Los Angeles company building national-security satellites that carry radio-frequency and optical sensors on the same spacecraft and compute in orbit: onboard models classify an RF signal, cue the camera and downlink a fused finding in minutes instead of raw data.
Founders
Kazi Farabi (CEO and co-founder; Northrop Grumman, Lockheed Martin and Booz Allen Hamilton; lead systems engineer on protected military communications and missile warning satellites, leading on-orbit testing for two of them), Tadeh Mirzakhanian (CTO and co-founder; led RF hardware on Starship Flights 2 through 9 at SpaceX, including the first Starlink phased array on Starship; electronic warfare at Raytheon on the MALD-N jammer), Jose Serrano (co-founder, listed as CSO and as CPO on the company site; software engineer for about a decade at Google, Amazon, Northrop and Lockheed Martin; leads government contracts and software architecture)
The deal
Speedrun's published standard deal: up to $1M, as $500K for 10% upfront on a SAFE plus up to $500K more in the next qualifying round ($2M+) within 18 months. Galileo Space's executed terms are not disclosed; its site lists a pre-seed backed by a16z Speedrun and Gutter Capital and a seed round as planned.
Most imaging and signals satellites still send raw data down and leave fusion to analysts on the ground, so a ship that turns off its transponder can be gone before anyone pairs the radio hit with a picture. Galileo's bet is that classification belongs on the satellite: hear the emitter, point the camera, send one answer. Running useful models within a small satellite's power budget is newer than the idea, and the team covers the hard parts unusually well, from missile warning payload testing to Starship RF hardware and electronic warfare, so it knows what the customer actually tasks. It was founded in 2025 and is in the newest Speedrun cohort. The case against is strong: it has no hardware in orbit, and its own roadmap puts the launch slot in Q4 2027 or Q1 2028; HawkEye 360 already sells space-based RF geolocation and BlackSky, Planet and Umbra sell fast imagery with onboard and ground analytics, and the Space Force can buy fused products from them; China's AdaSpace already runs models in orbit; defense procurement is slow, and the support letters, demos and a first $1.25 million contract still in development that the profile cites are not revenue; a three-person team building its own payload faces long, capital-heavy integration before it can prove anything.
A compiler extension that hardens existing C and C++ code at build time so memory-corruption bugs cannot be turned into remote code execution: it analyzes the source, inserts runtime guards into the binary, runs in monitor mode first and then blocks, with no code changes; aimed at IoT and OT firmware, open-source libraries and legacy codebases.
Founders
Noah Chander (CEO; built secure systems for Prime Video; advised Fortune 500 companies on cyber risk at McKinsey; led research on privileged enterprise data at a large financial market intelligence company), Finlay Moss (CTO; head of threat intelligence at Millennium; won the UK Government's national hacking competition two years running; maths and computer science at Imperial, where his thesis work was on call-graph analysis for forward-edge control-flow integrity in LLVM)
The deal
Entrepreneur First says it invests up to $250,000 in equity in teams that pass its Investment Committee and join the LAUNCH program in San Francisco, alongside an equity-free living-cost grant. Cofibrant's site also names Transpose Platform and angel investors as backers; no amount, instrument or executed terms are disclosed.
The pitch is a response to AI-found bugs. If models now find memory-safety flaws and write working exploits in days, patching cannot keep up, least of all in firmware that ships once and runs for a decade. Cofibrant gives up on finding the bugs and instead tries to make one class of them unexploitable: by its count from CISA's exploited-vulnerability catalog, 192 of 689 remote-code-execution entries go through memory corruption. The technique is not new; control-flow integrity has been studied for twenty years and Moss's own thesis was on it. What changed is the buyer's math, since a guarantee at build time beats a patch queue when time-to-exploit collapses, which is why it could not have sold two years ago even if it could have been built. A threat-intelligence lead and a national hacking champion know how exploits are actually chained. The case against: the site concedes this is weaker than full memory safety (Rust, Fil-C), so data-only attacks and information leaks remain; 'provable' has not been tested publicly beyond the company's own FFmpeg run and papers; Clang's CFI, Microsoft's CFG, Intel CET and Arm BTI already cover part of this for free; embedded firmware often builds on vendor compilers a compiler extension may not reach; and there is no named customer, only an early-access list.
An eye-movement test for early Alzheimer's risk: a headset with a patent-pending camera attachment records about fifteen minutes of reflexive and deliberate eye movements (look-toward and look-away tasks, inhibition, working memory, landing accuracy), and a model compares them with patterns linked to Alzheimer's; it sells a general-wellness version with a blood panel and runs an IRB-approved study toward a Class II device.
Founders
Shreyas Bhardwaj (Founder/CEO; previously built medical devices for hospitals; soft robotics researcher)
The deal
YC standard: $500K, as $125K post-money for 7% plus $375K uncapped MFN. No other round disclosed.
Alzheimer's drugs that slow decline now exist, but they work best early, and the early tests are a PET scan, a spinal tap or a blood draw sent to a lab, all gated by a doctor who already suspects something. Eye movements are an old research signal: people with Alzheimer's and amnestic MCI are slower and make more errors on the anti-saccade task, where you must look away from a flash. Saccade's bet is that a consumer headset can now capture that at clinical quality outside a lab. Two years ago the research rigs were desk-mounted trackers; a 2020 study of VR saccade measurement tested seven healthy adults. A UCSB engineering capstone describes the attachment as near-infrared cameras for an Apple Vision Pro. The non-obvious part is the two-track regulatory design: a wellness product at $897 a quarter, bundled with a blood panel and a doctor consult, pays for the IRB study that has to come before an FDA filing. The case against is strong. The company's own documents call the Alzheimer's test investigational and say it cannot diagnose, predict or rule out the disease; no validation study, sample size or accuracy figure is public; eye-movement changes are not specific to Alzheimer's; cheap blood tests for p-tau217 are moving into primary care and may make a headset unnecessary; and a wellness product that implies dementia risk invites FDA attention. It is also a team of two.
Montreal company whose software reads a circuit-board schematic and the datasheets of every part on it, then checks the design for electrical errors (voltage mismatches between parts, I2C/SPI/UART wiring, logic polarity, duplicate bus addresses, ratings and margins) at any stage, using AI to read and expert-system rules plus simulation to confirm each finding with its evidence.
Founders
Margot Blouin (CEO and co-founder; founding product manager for Microsoft's Azure Quantum; earlier Xbox; nearly a decade in hardware research), Scott Bright (COO and co-founder; co-founded Synapse Product Development, now part of Capgemini Invent; 35+ years in electrical engineering and product development)
The deal
$6.8M seed led by Bison Ventures with Innovation Endeavors and AI2 Incubator, announced December 19th 2024, after a $650K AI2 Incubator pre-seed. AI House's published incubator standard is up to $600K on a SAFE at a $10M cap plus 7% common stock; whether that applied to Cadstrom's pre-seed is not stated.
A board respin is the most expensive way to find a wrong pull-up resistor: weeks of schedule and a new fab run, often traced to something written on page 40 of a datasheet nobody reread. Senior engineers catch these in design review by memory; Cadstrom's pitch is that a model can now read the datasheets for every part, and rules and simulation can then decide whether the flagged interface is actually wrong, which keeps it from being an LLM guessing about electronics. Scott Bright co-founded Synapse, a product-development consultancy, and a consultancy is where you see the same classes of schematic error repeat across hundreds of client boards; that catalogue is the expert system. It is not strictly a could-not-exist-two-years-ago company, since it was founded in 2023 and funded in 2024, but its approach only became practical once models could reliably read long, table-heavy datasheet PDFs, and it is spotlighted now because hardware design is where agentic tools are moving next. It names Texas Instruments and StarFish Medical among users, says it is SOC 2 audited and does not train on customer designs. The case against: Cadence, Siemens and Altium own the tools engineers live in and can add the same checks natively; a validation tool earns trust only when it misses nothing important and flags little that is wrong, which is hard to prove; PCB teams are small buyers; and the 66% shorter design cycle is the company's own figure. No round since December 2024 is public.
Boulder company that turns cameras on municipal vehicles (existing fleet dashcams or its own windshield camera, about a 15-minute install) into a continuously updated map of a city's pavement defects, signs, markings, hydrants, manholes and drainage, with repair prioritization, dashboards and map-service APIs for public works departments.
Founders
Kyle Wandishin (CU Boulder computer science; runs an automation consulting company), Wade Turner (CU Boulder information and data science)
The deal
Techstars standard: $220K, as $20K for 5% through a post-money convertible agreement plus $200K on an uncapped MFN SAFE. Arterial's executed terms are not disclosed.
Most cities still know the state of their roads from a consultant's survey every few years and from residents reporting potholes. Arterial puts the inspection on vehicles that already drive every street, such as garbage trucks and parking enforcement, and says it refreshes about 80% of a jurisdiction every two weeks. What is new is cost: a vision model running on cheap cameras can now classify an asset inventory, not just potholes, and a third-party profile says its pipeline rebuilds corridors in near-LiDAR 3D from ordinary video, which would replace a survey van. In early 2025 its model was trained on about 55,000 labeled frames, and the City of Boulder signed a paid pilot. The case against is serious: this is a crowded category with older players (RoadBotics, Blyncsy, Cyvl, Good Roads), the founders are recent students without a public-works background, municipal procurement is slow and the budgets are small, and fleet telematics vendors that already own the dashcams can add the same detection. Included for the distribution idea, not the technology.
Edits P&IDs (the piping and instrumentation diagrams behind refineries, fabs and power plants) in their native DWG/DXF files from a plain-language request, carries the change into line lists, valve lists, datasheets and cable schedules, and checks the whole document set against engineering rules before an engineer approves it.
Founders
Oliver Wesche (CEO; previously machine learning engineer at Blue Origin, earlier at Asurion; Harvard Business School MBA; aerospace engineering at UAH), Ujjawal Madan (CTO; previously AI engineer at Promise (YC 2018); principal data scientist and senior manager at Asurion; M.S., Georgia Tech)
The deal
YC standard: $500K, as $125K post-money for 7% plus $375K uncapped MFN. No other round disclosed.
A process plant is thousands of documents that are supposed to agree, and engineers keep them in sync by hand; the profile's example is one tank carrying three different design pressures across three drawings already issued for construction. Two years ago a model could not reliably read a P&ID's symbols and line connections, let alone write a valid edit back into the CAD file and trace it through every downstream list. It keeps the engineer as approver and never becomes the system of record. The case against: owners on data-centric tools such as Hexagon Smart P&ID, AVEVA or AutoCAD Plant 3D already keep tags in a database, so Verno's market is the long tail of projects still on plain drawings; an error that reaches an issued drawing at a refinery is a safety matter, so engineering contractors will buy slowly; neither founder lists time inside one; no customer is named, and the website demo is an animation.
Alibaba's wholly owned data center chip designer, which confidentially submitted a Hong Kong listing application on September 11th 2026; the exchange has cleared Alibaba to proceed with the spin-off.
Alibaba's November 10th announcement says T-Head filed a Form A1 confidentially on September 11th and that the exchange confirmed under Practice Note 15 that the spin-off may proceed. The plan is a global offering of H shares on the Main Board; Alibaba expects T-Head to remain a subsidiary, and the size of the stake sold is not set. It still needs Listing Committee approval and a CSRC filing. The disclosure is the parent's own exchange announcement, so this is filed, not rumored; no size, price or date. It would join Denglin, Moore Threads and MetaX among Chinese AI chipmakers raising public money while export rules keep Nvidia's best parts out of China.
Xiamen designer of vision AI chips for security cameras, IoT and car devices, already listed on Shenzhen's ChiNext, seeking a second listing in Hong Kong.
SigmaStar first applied to the Hong Kong exchange on September 26th 2025 with CICC and China Securities International as joint sponsors, and an application proof was published on HKEXnews on March 27th 2026. On November 13th Bloomberg reported it may list as early as this year, raising up to $300 million, and has begun talking to investors; that report reached us through a relay, not the original. Its filings describe more than 600 SoC designs and 454 million chips shipped by mid-2025, with a push into edge AI inference chips and AI glasses. A dual listing of an established edge chip supplier: filed, not priced.
Quality scoring for robot training data: it scores each human demonstration and each operator across teleoperation, UMI handheld grippers and egocentric video, adjusts for task difficulty, separates execution mistakes from capture problems, and flags quality drops before bad episodes pile up.
Founders
Ramin Keene (co-founder; CTO of StockX; built Rover's first ML platform; streaming data infrastructure at Nordstrom; led infrastructure for training and serving generative video models at Hedra; O'Reilly author), Alan Guo (co-founder; head of business and chief of staff at Hedra; Harvard Business School MBA)
The deal
Speedrun's published standard deal: up to $1M, as $500K for 10% upfront on a SAFE plus up to $500K more in the next qualifying round ($2M+) within 18 months. Munari's executed terms are not disclosed.
Robot learning is now bought by the hour: labs and their contractors pay people to teleoperate arms, carry handheld grippers or wear head cameras, and those hours vary widely in what they teach a model. Munari scores the person behind each demonstration, takes MCAP, ROS bags, LeRobot or raw video with a control log without an integration, and promises a first report within 24 hours. Two years ago there was no such market, because demonstration collection at the scale generalist robot policies now need had not started. Keene ran training and serving infrastructure for video models at Hedra, close kin to deciding which footage is worth training on. The case against: whether a demonstration was good is settled by whether the policy trained on it improves, which only the lab sees, so a score without that feedback loop is a proxy; the largest robotics labs build this in-house; scoring operators invites them to game the score; neither founder lists robotics research; no customer is named.
An agent-driven sales CRM: agents research accounts, rank which to work and draft outreach so reps spend less time on admin.
$100 million Series B at $1.3 billion, co-led by General Catalyst and March Capital with Sequoia and GV returning and Prosperity7 new. CEO Ishan Mukherjee sold Pixie Labs to New Relic for about $150 million in 2020; co-founders include Stanford's Christopher Ré and AWS Aurora founding member Shriram Sridharan. Customers include Ramp, MongoDB and Groq; ARR is up more than tenfold since its first year, with no total disclosed.
AI investigators for consumer protection agencies: agents browse live websites the way a consumer would to find junk fees, subscription traps and deceptive ads, triage incoming complaints within a minute, and map each finding to the elements of the law with timestamped, hashed evidence.
Founders
Kavin Kumaravel (CEO; consumer protection and capital markets at the U.S. Senate Banking Committee; a causal inference startup, Haus; B.S. EECS, UC Berkeley), Christopher Bannon (CTO; software development engineer at AWS building security and compliance infrastructure, including services for the EU Sovereign Cloud; BA Computer Science, UC Berkeley)
The deal
YC standard: $500K, as $125K post-money for 7% plus $375K uncapped MFN. No other round disclosed.
Consumer protection enforcement mostly starts with a complaint, and most harmed people never file one, so agencies find a junk fee or a cancellation maze years after it started. Particularity flips the order: agents walk checkout and cancel flows across many sites, save dated, hashed copies of pricing and terms, and hand staff a case already sorted by legal element. Two years ago a browsing agent could not reliably get through a real checkout, and a founder who worked consumer protection at the Senate Banking Committee knows the bottleneck is investigator hours. The case against: the buyers are state AG offices and city consumer affairs departments with slow procurement and small budgets, and federal enforcement at the CFPB has been cut back since 2025; evidence gathered by an agent will be challenged in court on how it was collected; sites can spot and serve automated visitors a cleaner page; the launch post is asking to meet agency staff, so no customer is named yet; and the team is two people.
Suzhou GPGPU designer serving Chinese enterprise AI compute, listing in Hong Kong under Chapter 18C with stock code 9917.
Its counsel Sidley said on November 9th the offering raised about HK$1.41 billion (US$180 million) before the over-allotment option, and TechNode reports the shares were priced at HK$231, the bottom of the HK$231 to HK$288 range, and began trading in Hong Kong on November 10th under stock code 9917. It listed under Chapter 18C, the regime for specialist tech companies that are not yet profitable. Reuters: opened at HK$229 and closed at HK$213 on November 10th, down 7.8% (SCMP: 7.7%). Retail was 2.11 times subscribed. Pricing at the floor, with Moore Threads and MetaX already trading in Shanghai, is a read on how much appetite is left for China's second tier of GPU designers.
A public question-and-answer network for coding agents: installed into Claude Code, Codex or Cursor, an agent searches fixes other agents have already found, posts what it learned after redacting keys, names, hosts and paths and getting the user's approval, and votes on fixes it has used.
Founders
Ansh Tulsyan (co-founder; previously built AI tools for chip design used by thousands of engineers at AMD), Ishaan Chamoli (co-founder; previously AI at Boardy and AMD; CS at UIUC)
The deal
Speedrun's published standard deal: up to $1M, as $500K for 10% upfront on a SAFE plus up to $500K more in the next qualifying round ($2M+) within 18 months. ChatOverflow's executed terms are not disclosed.
Stack Overflow's traffic went to chat assistants, and the knowledge those assistants produce while debugging now disappears at the end of each session. ChatOverflow tries to rebuild the commons with agents as both askers and answerers: a tool error one agent spent twenty minutes on becomes a ranked fix the next agent finds in seconds. Two years ago no coding agent ran long enough unattended to need this, and there was no standard way to install a shared memory into Claude Code, Codex and Cursor at once. It differs from Agent Relay (November 8th), which passes messages between one team's agents; this is a public pool across strangers. The case against: a public feed of fixes that agents read and act on is a prompt-injection and supply-chain target, and votes cast by agents are easy to farm; automated redaction will eventually let a secret through; the 44% faster figure is the company's own controlled SWE-bench run, and the site's headline counters still read zero; and model vendors can add shared memory across their own users without a third party.
Munich robotics company founded in 2018 by German Aerospace Center researchers, now building humanoids in its own factory near Munich.
Raised $500 million led by SoftBank, with Nvidia, Hitachi and T.Capital, at an $8.5 billion valuation, about $2 billion raised in total. CEO Zhaopeng Chen expects about 2,000 humanoids built in 2026 and two to three times that in 2027, which makes it one of the few humanoid companies quoting volume rather than demos.
A payments account an AI agent can use to send and receive money, connected to ChatGPT, Claude, Muse or a custom agent over MCP; recipients need no account and cash out to a bank through a link, with banking handled by licensed partners.
Founders
Harshith Latchupatula (Stripe Connect; AWS Fargate autoscaling at Amazon), Vishnudev Poil (Ads at Amazon; observability at OpenText), Param Thakkar (vector search at Databricks; Deep Research at Glean; ads ML at Quora)
The deal
YC standard: $500K, as $125K post-money for 7% plus $375K uncapped MFN. No other round disclosed.
The bet is that the first money agents move will be small, social and boring: splitting a dinner bill, paying a friend when you miss a running goal, settling rent with roommates, paying an invoice out of the inbox. Two years ago no assistant had a standard way to reach a wallet; MCP gives it one, and a founder who worked on Stripe Connect knows where payouts break. The case against: OpenAI is building agent payments with Visa and Stripe already sells agent toolkits, so this is a feature the platforms can ship; consumer money transmission through partners is a compliance business before it is a software one; and the first agent that pays the wrong person will set the regulator's view of the whole category.
Predicts leaks and cooling failures in liquid-cooled infrastructure, starting with AI data centers, by learning how each cooling loop normally behaves from non-invasive sensors and flagging the early signs of failure before liquid escapes.
Founders
John Khazraee (CEO and founder; 15+ years building data center infrastructure at Microsoft, Oracle and IBM; engineer, MBA, US veteran)
The deal
Techstars standard: $220K, as $20K for 5% through a post-money convertible agreement plus $200K on an uncapped MFN SAFE. The class announcement does not give MayimFlow's executed terms or any outside round.
The founder's line is the thesis: as AI racks go liquid-cooled, every rack becomes a plumbing problem. Two years ago most data center halls were air-cooled; Blackwell-class racks made direct liquid cooling the default, and the company cites insurers putting about 24% of data center loss costs on liquid. Its scientific advisor, Stanford's Butrus Khuri-Yakub, co-invented the CMUT ultrasonic transducer and a clamp-on flow meter, which is the sensing know-how a software-only rival lacks. The case against: cooling vendors such as Vertiv and Schneider can build leak prediction into the units they already sell, predictive maintenance is crowded, and a single-founder team selling to hyperscalers faces year-long pilots; the website's statistics are its own selection.
AI-native primary care: a chat-based medical AI that hands users to licensed physicians, running inside the Oura app since June 2026; being acquired by Oura for $625 million.
An exit, not a round: Oura agreed on November 10th to buy Counsel for $625 million, closing expected before the end of 2027, with CEO Muthu Alagappan becoming Oura's chief clinical AI officer. Counsel had about 70 staff and about $36 million raised, including a $25 million Series A led by a16z and GV, so the price is roughly 17 times capital in. It is the clearest sign yet that wearable makers want to own the clinical step after the reading.
Sells reinforcement-learning environments to frontier labs built from de-identified time-series records (health data, industrial sensors, compute-cluster logs) that arrive on a clock while the agent works, with every answer graded against what actually happened.
Founders
Antanas Žilinskas (solo founder; first author of two ICLR 2026 papers on rare-event forecasting; previously co-founded Entangl, YC S24, as CTO, taking AI for engineering into Fortune 500 deployments; MEng, Imperial College London)
The deal
YC standard: $500K, as $125K post-money for 7% plus $375K uncapped MFN.
A bet that the gap left in agents is time: a model can pass the bar exam and still fail to watch a grid or a GPU cluster for weeks and say what happens next. Reific's own leaderboard (version 17, November 4th) has no model above 43.1% in any of its three families, and the founder's ICLR work on rare-event forecasting is the method underneath. Two years ago there was no market of labs buying training environments for agents. The case against: the buyers are a handful of labs that build environments in-house and squeeze vendors on price, the leaderboard is self-run with no named lab customer, de-identified health records carry licensing and re-identification exposure, and it is one founder. Adjacent to Tokenless (November 4th), which audits RL environments; Reific sells them.
$300 million Series E led by Georgian at $2.45 billion post-money, about four times the roughly $570 million of its November 2025 round; D. E. Shaw Ventures and Naver Ventures joined existing backers. The company says ARR rose 17-fold in 18 months and it has raised about $610 million since 2018. Welding is a skilled-labor shortage with a measurable output, which is why it is one of the first physical-AI jobs to price like software; the plan to move into defense and energy infrastructure is where that thesis gets tested.
$240 million Series C at $3 billion, led by Lightspeed with Insight co-leading, eight months after an $80 million round at $1 billion. Hardware test data is the unglamorous part of building drones, aircraft and spacecraft, and defense spending is pulling it forward; it holds an Air Force Test Center contract with a $53 million ceiling.
Turns a property address into an explorable, AI-generated simulation of the surrounding neighborhood that buyers can walk before a showing, shared as a link, listing embed or QR code; sold to agents, brokerages and developers.
Founders
Caleb Lo (CEO; says his mother's work as a real estate agent showed him buyers researching a neighborhood across disconnected tools), Andrew Shum (CTO), Joshua Zhang (COO)
The deal
$500K pre-seed from a16z Speedrun, per RuntimeWire on October 1st 2026 citing the CEO's announcement. Speedrun's published standard is $500K for 10% via SAFE plus up to $500K in a qualifying $2M+ round within 18 months; whether Mirai's check took those terms is not disclosed.
Mirai calls itself the application layer for world models. Two years ago a walkable, generated version of a real street was a research demo; now it can be a listing add-on, and real estate already pays for 3D tours and drone footage. The test is whether a generated neighborhood helps sell a house. The case against: a scene that gets a street wrong misrepresents the biggest purchase most buyers make, a risk RuntimeWire raised itself; Street View is free and captured tours are already bought; and the site does not say whose world model it runs, so a model vendor or a listing portal could ship the same feature. The profile lists the team as ex-TikTok, Shopify, Waabi and Walmart Blue Labs without saying who came from where.
Tel Aviv agent-security company that checks each action an AI agent proposes against the whole session and allows, blocks or escalates it within milliseconds.
Out of stealth with $50 million: a $37 million Series A led by Battery Ventures with Micron Ventures, M12 and Wiz's Assaf Rappaport, plus a $13 million seed from YL Ventures and Striker. Founded by Yuval Zacharia, Aner Morag and Boris Danilovich. It arrives the same week as research showing agents overstep permissions in roughly half of runs with no attacker present; the pitch is that per-action approval scales where hand-written rules per tool do not.
Learns each enterprise codebase's conventions, history and the mistakes its coding agents make, then adds merge gates that check agent code with deterministicverifiers rather than another model, so the same mistake cannot ship twice.
Founders
Edward Ahn (CEO; former founder and CEO of a 3D AI startup backed by Nat Friedman and Daniel Gross; AI agent researcher at Apple; RL researcher at CMU), Zhengming Wang (CTO; built 0-to-1 products at Retool and helped scale Character.AI; behind a GitHub repo with 40K+ stars that was acquired in 2023)
The deal
YC standard: $500K, as $125K post-money for 7% plus $375K uncapped MFN. No other round disclosed.
The claim is that the limit on coding agents inside large companies is no longer what the model can write but what a reviewer will approve, and that the answer is not a second model grading the first. Plover turns each failure into a deterministic check at the merge gate, so trust accumulates per codebase and agents can be handed more of the work as the gates pile up. Two years ago there was no volume of agent-written pull requests to gate, and an Apple agent researcher paired with someone who scaled Character.AI's engineering has seen both the agent side and the review backlog. The case against: GitHub, Cursor, Anthropic and OpenAI all ship review and policy features and can add learned rules where the code already lives; deterministic gates catch the mistakes you have already seen, not the novel ones, which is where shipping without review is dangerous; the company's own site does not yet explain the gates; and the tripled PRs per engineer at Fieldguide is the company's own figure, and more PRs is not the same as fewer defects.
A market-discussion network whose participants are AI trading agents: an agent registers through an API or MCP server, posts signals, debates positions and builds a reputation scored on how its signals perform, while humans sponsor agents or read the feed.
Founders
Not named on the Techstars page or the company site. Product Hunt lists three makers: Berkay Baygut ('Building cool stuff'), Ceyhun Derinbogaz ('Bringing AI Creation to Every Text Box') and Kemal Apaydin ('AI Safety Expert'); their roles at the company are not stated.
The deal
Techstars standard: $220K, as $20K through a post-money convertible equity agreement for 5% plus $200K on an uncapped MFN SAFE. The class announcement says only that founders 'receive capital investment'; Marx Finance's executed terms, and whether the ABN AMRO partner program differs, are not disclosed. A Stork AI listing of a $5 million Series A could not be verified.
Most finance-agent companies sell an agent; Marx bets the agents need a place to meet. It ships a skill file an agent can read to register itself, an MCP server and a ticker server, and it ranks agents on the track record of their signals rather than their followers, which makes it closer to a public leaderboard for trading models than a chat room. That only makes sense now that people run their own agents that can sign up for a service unaided, which was not true two years ago. The case against: a reputation score built on signal performance invites the old newsletter trick of many accounts and survivorship, and signals aimed at retail readers sit close to investment advice that regulators already police; the network showed about 1,300 agents and 149 discussions when checked, small for a product whose value is the crowd; and neither the Techstars page nor the site names the founders, so the team's edge cannot yet be judged.
A market for reserved AI inference: buyers lock a price today for a fixed volume of tokens from a chosen model and provider over a set window, providers bid against each other, and unused capacity can be resold to another buyer.
Founders
Ilia Bolgov (founder and CEO; product for wealth and trading at Revolut; mathematics and finance at Imperial College London), Roman Yanushevskyi (founder and CTO; two quantitative research internships at Citadel Securities; AI engineer at Lovable; math and CS at the National University of Singapore; IOI gold medalist)
The deal
YC standard: $500K, as $125K post-money for 7% plus $375K uncapped MFN. Touchmark's site also lists Inception, Liquid 2, Kima Ventures and Pareto as backers; amounts and company-specific terms not disclosed.
Two years ago inference spend was too small, and too concentrated in a couple of closed labs, for a forward market to make sense. Now open-weight models such as Kimi K3 are served by many hosts at competing prices, and token bills are large enough that finance teams want them fixed. Touchmark's contract is a fixed token volume over a window, with output tokens weighted at 1.0, input at 0.2 and cached input at 0.02, and service levels attached: 99% availability, under one second to first token at p95, 250 tokens a second. Its pitch to sellers is that prepaid future capacity finances GPUs without dilution or interest. The founders bring a trading desk's habits, from Revolut's trading product and Citadel Securities, to a market that has never had a forward curve. The case against: per-token prices for a given capability keep falling, so a buyer who locks a price for months can end up above spot, and the resale market meant to fix that is unproven; labs and large hosts already sell committed-use discounts directly; the site names no customer, labels its provider bids illustrative, and its '4T+ tokens' figure is requested volume rather than volume traded; and the launch post limits trading to commercial producers and end users, wording that suggests the company is already managing how regulators would classify these contracts.
Business Insider's report, relayed by secondary sites, has the round closing within days, below the roughly $50 billion discussed late last year. It lands in the week OpenAI took $1.25 trillion below its ask, so it is a third data point for a reset in frontier private prices. The lab released Inkling 2 Small the same day and promised Apache weights, which makes it the rare frontier-scale lab raising on an open-weight strategy. Unconfirmed by the company.
AI trade-compliance platform paired with licensed US customs brokers, lowering import duties and preventing border holds for enterprise importers of fashion, beauty, food and consumer goods.
Founders
Cameron Bell (CEO; grew up in his family's clothing and electronics import business; founded a customs brokerage SaaS company and took it to $100K ARR; BlackRock and Teya, then built a sales channel at a $9B London fintech to a third of UK revenue in six months), Ufuk Dede (CTO; founding engineer at Cenoa, a payments company, where he scaled international payments infrastructure to more than 500,000 users)
The deal
Entrepreneur First says it invests up to $250K after a team forms. Greenwich Mercantile's specific terms are not disclosed on the EF profile; its site quotes a fixed fee for the initial review and does not publish ongoing pricing.
Since 2025, stacked US tariffs have turned customs classification from clerical work into a margin decision: the company's own lead magnet is a checklist for Section 301, Section 232, IEEPA and antidumping duties applied on top of each other. Greenwich Mercantile reads supplier documents, keeps one record per product, flags mismatches before filing, and has a licensed US customs broker review and sign every classification, origin and value call, so the liability sits with a license holder while the software does the reading. Reading messy invoices and spec sheets across hundreds of SKUs is what current document models made cheap. Bell grew up in an import business and built customs software before, so he has seen the paperwork from the importer's side. The case against: Tarifflo, in YC's Summer 2026 batch and led by a licensed broker, gives its filing software away and earns on the overpayments it recovers, which undercuts a fee-for-scope service; Flexport and the large brokers already pair software with licensed staff; the savings depend on tariff policy that can change by executive order or court ruling; and the logo strip on its site mixes one consumer-goods customer with a university, the UK Civil Service and the CTO's own employer, Cenoa.
Open-source messaging layer that lets a team's coding agents (Claude Code, Codex, Cursor, Devin and others) message each other directly across tools and machines, read each other's sessions with permission, and share a searchable history.
Founders
Will Washburn (founder and CEO; spent 14 years building Tailwind, a marketing SaaS for small businesses, to eight-figure ARR and more than 50,000 paying customers), Khaliq Gant (founder and CTO; founding engineer at Nango (YC W23) and lead engineer at Overview (YC W19); Cornell and Columbia)
The deal
YC standard: $500K, as $125K post-money for 7% plus $375K uncapped MFN. Agent Relay's site also lists Hustle Fund, Active Capital, Cortical Ventures and YONDER as backers; amounts and company-specific terms not disclosed.
The problem only exists once each engineer runs several coding agents and a teammate runs several more. Then the people become the message bus, pasting one agent's summary into Slack and from Slack into another agent. Agent Relay lets those agents ask each other directly, sort out a merge conflict between branches, and get pinged when a teammate's work lands, across vendors and machines. The founders say they run more than 50 agents in parallel themselves, which is where their read on the bottleneck comes from, and the relay is open source under Apache 2.0 with about 1,100 GitHub stars. The case against: Anthropic already ships experimental agent teams and cross-session messaging inside Claude Code, so vendors are building the single-user version and the cross-person version is a short step from there; a layer that sits between competing vendors' agents is the kind of thing a standard or the vendors themselves tend to absorb; letting a colleague's agent read your session and make requests of your agent widens the prompt-injection surface, which the site answers only with a rule that messages can ask but never grant permission; and the one named customer quote comes from Pomerado, a company in the same YC batch.
Parent filed in August to raise 33 billion yuan ($4.9 billion); the listing is on hold.
Reuters reported on November 4th 2026 that the flash-memory maker paused the listing, citing a company memo that gives only "various factors" and people familiar with the matter. Sources say the delay could last months, with a listing possible in the first half of 2027. It stays filed until withdrawn or priced. [B] November 11th: Bloomberg reports YMTC now targets a listing as soon as the first quarter of 2027, later than the year-end timing observers expected in August; still filed, not approved.
Mines venom biology with AI for peptide drug candidates and tests the survivors in its own lab; first target is antimicrobial resistance.
Founders
Kartik Sunagar (co-founder and director; evolutionary biologist and professor at the Indian Institute of Science, about two decades on venoms and antivenoms), Vinay Prabhu (co-founder and CTO; Carnegie Mellon PhD, IIT Madras MSEE; repeat founder who has built AI at scale)
The deal
$2.7M pre-seed announced April 2026, per the company's site, with no investors named there. SPC says it invested in a post dated November 4th 2026 but gives no amount, instrument or co-investors, and neither source says whether SPC is in that round.
The computational half of this idea was proved in public only last year: a Penn team mined more than 16,000 venom proteins, found 386 candidates, and 53 of the 58 peptides it synthesized worked against drug-resistant bacteria. VenomicsAI bets the scarce part is the rest of the loop. Sunagar has spent two decades on snakes, scorpions, spiders and wasps, and the company pairs that sourcing with patch-clamp screening under one roof, so a prediction reaches the bench quickly. The case against: the Penn work, which shares the phrase 'Venomics AI' but lists neither founder as an author, shows the mining step is open to any good lab; SPC's own post says the hard part is turning discoveries into drugs; the site's 0.8nM potency figure and 'IND-enabling ready' label for VNM-402 carry no citation; and antimicrobials are a weak first market, because new antibiotics are held in reserve and sell little even when they work.
Baidu's AI chip unit, reported to have filed confidentially in Hong Kong in January for a roughly $2 billion IPO that may now slip into 2027.
Bloomberg, citing people familiar with the matter, reported on November 3rd 2026 that the listing may slip to next year while Kunlunxin waits for approval from China's securities regulator. A confidential filing reported through sources stays rumored here until a public document appears. Bloomberg's original was not opened; this rests on a relay. Reuters reported in December 2025 that investment materials showed a goal of completing an IPO by early 2027, so a slip into next year would land near the original plan. [B]
Self-healing operations for GPU fleets: it watches servers, diagnoses faults against the fleet's own history, and applies or suggests a fix an engineer can approve, for smaller GPU clouds and AI teams running their own machines.
Founders
Param Shah (co-founder and CEO; led product at Microsoft AI Supercomputing, managing GPU deployments for frontier AI labs; Berkeley), Sahaj Putcha (co-founder; Amazon Robotics; computer science at UIUC)
The deal
Speedrun's published standard deal: up to $1M, as $500K for 10% upfront on a SAFE plus up to $500K more in the next qualifying round ($2M+) within 18 months. Harbor AI's executed terms are not disclosed.
Fleet monitoring is old; the buyer is new. Two years ago few small operators ran bare-metal 8xH100 servers for AI tenants without a hyperscaler's reliability team behind them, and those are the buyers Harbor targets. Shah managed GPU deployments for frontier labs inside Microsoft, so his claim about what breaks comes from the largest fleets. The product works on specifics such as Xid errors, InfiniBand link flaps and NCCL checks, then stores each resolved fix for reuse across sites, which makes the compounding asset a fault history. The case against: Nvidia's Mission Control, announced in March 2025, already excludes faulty hardware, runs recovery playbooks and flags parts for RMA, though Nvidia describes it around its own reference architectures; larger clouds can build this in-house; the site names no customer and its goodput figures sit in illustrative mockups; and reseating a card still takes a person at the rack.
Suzhou maker of printed circuit boards and optical transceivers, a supplier of AI server components, listed in Hong Kong under stock code 2384.
Pipeline catch-up: raised HK$9.88 billion at HK$86.80 and listed October 28th 2026, opening at HK$82.40 and closing 0.75% lower at HK$86.15, per The Standard. AI-adjacent, not AI: the demand it sells into is AI servers, and a large supplier closing below its offer price in the first session is a read on appetite for that trade. [B]
Chengdu AI-satellite maker running Alibaba's Qwen model in orbit; listed in Hong Kong on November 13th under code 6935 at HK$24.90.
ChinaTechNews reports 15 completed missions, 35 satellites launched and 3.475 billion yuan ($485 million) of pending orders for 114 AI satellites, with proceeds earmarked for a 543 million yuan Shenzhen base. The Bamboo Works reports the post-hearing information pack was published October 29th; that page was not opened. A cleared hearing is a gate passed: no offer size, price range or listing date was found, so this is filed, not priced. [B] November 8th 2026: IPOX reports 30,121,000 H shares at an indicative HK$24.90 to HK$33.20, stock code 6935, trading expected November 13th; Finimize reports up to about HK$1 billion. Both are relays and the prospectus was not opened, so it stays filed until it prices. [C] November 13th: priced at HK$24.90, the bottom of the HK$24.90 to HK$33.20 range, selling 30.12 million shares for about HK$750 million (Reuters). It opened at HK$24.44, down 1.8%. Retail was 3.43 times subscribed and the international book 1.79 times.
A three-person applied research lab building tools to detect flaws in reinforcement-learning environments and evals, with the stated aim of ending reward hacking.
Founders
Rohit Agarwal (CS PhD student at Princeton PLI, Hazan Lab; Google DeepMind; previously Duolingo; top EECS graduate at Berkeley in 2024), Andrew Liu (UC Berkeley researcher on language documentation tools; CS, mathematics and linguistics), Kev Wang (SkyPilot, multi-cloud GPU infrastructure for neolabs; founding engineer at Eventual; Berkeley EECS)
The deal
YC standard: $500K, as $125K post-money for 7% plus $375K uncapped MFN. Company-specific terms not disclosed.
Labs now spend heavily on RL environments, and a flaw in one teaches a model to cheat at scale, so the customer base for this barely existed two years ago. The non-obvious claim is that environment quality, more than compute, limits post-training. The case against: the best labs treat eval integrity as core and will build it themselves, which leaves neolabs and environment vendors as a thin market, and a research lab that sells tools can end up as neither.
Turns clinical trial protocol PDFs into a structured, system-agnostic model that generates study documents and system configurations, with humans reviewing the AI extraction.
Founders
Zach Sawaged (founder and CEO; worked across research sites, sponsors and vendors, including NYU Langone, Amgen, Johnson & Johnson, Takeda and Veeva)
The deal
Round led by Surface Ventures with Techstars, Innovation Works and Gaingels, announced August 25th 2026; amount undisclosed. Techstars standard: $220K, as $20K post-money convertible equity for 5% plus a $200K uncapped MFN SAFE. Company-specific terms not disclosed.
Reading a dense protocol reliably enough to configure downstream systems is what recent models made affordable, and the founder has worked on every side of the table that handles those documents. The company claims six to eight weeks off study activation and estimates configuration waste at about $25 million a study, both its own figures. The case against: Veeva and Medidata already own the systems being configured and can add extraction themselves, and sponsors are slow to trust a new source of truth.
Security verification for AI agents in regulated industries: it simulates attacks, proposes a tested fix in about four hours, and checks each agent action in production in milliseconds without a GPU.
Founders
Daniel Begimher (CEO; led the global application security technical field community at AWS and built the AI investigative agent for AWS Security Incident Response; co-author of SIR-Bench), Cristian Leo (CTO; led AI security science at AWS; created ThreatForest, an open-source threat modeling agent; co-author of SIR-Bench)
The deal
YC standard: $500K, as $125K post-money for 7% plus $375K uncapped MFN. Company-specific terms not disclosed.
The founders shipped a customer-facing security agent inside AWS and wrote a benchmark for evaluating one, so they know where these systems fail in production. The ARTEX intrusions make the timing plain. The case against: its claim to beat every detector on six public benchmarks is its own, and incumbents are buying this category outright, with CrowdStrike paying about $495 million this week for Fiddler and Dreadnode, so distribution may decide it before detection quality does.
Runs reinforcement-learning training and inference for companies, with its own researchers working on customer projects.
The deal
$350 million led by Elad Gil at a $3.25 billion valuation, with Nvidia, Shopify Ventures and HubSpot Ventures.
Founded by Rhythm Garg, Yash Patil and Linden Li, it launched less than 18 months ago and says annual recurring revenue topped $100 million, with Microsoft and DoorDash as customers. A services-heavy model valued like software is the thing to test.
Verifies the evidence behind health and wellness products (lab certificates, doses, claims, market rules) and serves it as source-linked feeds that AI shopping assistants can trust.
Founders
Karina Valencia (co-founder; 18 years of cross-border operating; three-time scale-up operator who directed market entries at Palantir in the run-up to its IPO and managed an $11M CPG portfolio at P&G), Dr. Leonie Bode (co-founder; medical doctor and founder of Switzerland's first venture-backed longevity clinic; leads Kensai's science and regulatory intelligence layer)
The deal
Techstars standard: $220K, as $20K through a post-money convertible equity agreement for 5% plus $200K on an uncapped MFN SAFE. Kensai's site also lists Innosuisse as a backer. Company-specific executed terms not disclosed.
Shoppers now ask AI assistants which supplement to buy, and the answer depends on facts that product feeds were never built to carry: whether a heavy-metal test was a full four-metal panel or a single mercury check, whether the dose matches the clinical study, whether a claim is allowed in that market. Kensai's point is that this proof sits in supplier PDFs and certificates of analysis that no assistant reads, so the products with the best evidence get skipped. That buyer, a brand or retailer worried about how an AI agent ranks it, did not exist two years ago, and pulling structured facts out of lab certificates at catalog scale needed current document models. A physician who ran a longevity clinic knows which claims regulators and informed customers actually check. The case against: Techstars lists it as a tool for checking formulations and claims against global rules for international launches, while the site now leads with AI shopping feeds, which reads as a recent repositioning; a verifier paid by the brand whose claims it verifies has a conflict that assistants and retailers may discount; marketplaces and product-data syndication incumbents can add the same layer; and its headline figure, that one in three SKUs contradict across channels, comes from its own check with no sample stated.
Hosts and runs AI video, image and audio models for developers.
The deal
$1.3 billion Series F led by Bessemer at a $16 billion valuation, up from $8 billion in April.
Bloomberg reports annualized revenue passed $800 million earlier this year. Demand has run strongest among customers using open models, and gross margins stay thin because compute is expensive to buy or lease.
A 'product brain' that turns code, tickets, docs, meetings and customer feedback into one knowledge graph, then writes agent-ready specs and tracks the work, with plugins for Slack and Claude Code.
Founders
Chris Bee (co-founder and CEO; more than two decades of product and engineering leadership at Amazon, Uber, Zillow and Lessen), Anton Safonov (co-founder and CTO; built global-scale infrastructure at Snap, Meta and LinkedIn, focused on reliable systems and trustworthy data)
The deal
$2.5 million seed led by AI2 Incubator (now AI House), with Acequia Capital, Mighty Capital, Grand Ventures and eLab Ventures, announced June 18th 2026. AI House publishes standard incubator terms of up to $600K on a SAFE at a $10M cap plus 7% common stock; whether those applied inside this round is not disclosed. Valuation not disclosed.
The thesis is that coding agents moved the bottleneck: once code is cheap, the scarce work is deciding what to build and keeping everyone, people and agents, on the same context, which Bee says eats half the week of good people without shipping anything. Devplan's answer is a single graph of why things were built, from sales calls to commits, that answers questions with sources and hands a coding agent a spec in one command. That product only makes sense in a world where agents like Claude Code take specs as input and MCP lets tools share context, neither of which existed two years ago. The case against: Atlassian's Rovo, GitHub and the coding agents themselves are all pulling the same context through MCP, and a team already on Jira or Linear may never need a separate layer; the speed and cost win over 'a standard Claude configuration' is the company's own internal benchmark, and the eight hours a week saved comes from a survey of its early users; and a knowledge graph built from Slack and meeting notes is only as good as what teams actually write down, which is the coordination problem it claims to solve.
Real-time sensors for the coolant in liquid-cooled data centers, flagging failing valves and bacterial growth before they block flow.
The deal
Buying Aquatic Labs, a water-chemistry sensor maker, for $16 million in cash and stock, after a $31 million Series A in June 2026.
A small company on an unglamorous problem that grows with every rack of liquid-cooled GPUs: operators still mail coolant to labs and wait days. Omen says it will monitor 250 to 300 megawatts of chips by year-end and wants a gigawatt by the end of 2027. Aquatic's founder Allan Adams, a physicist and former MIT professor, becomes chief science officer.
Bengaluru lab building real-time text-to-speech models for Hindi and 10+ other languages, meant to carry a native speaker's accent, rhythm and emotion; two models released as open weights.
Founders
BS Dheemanth Reddy (co-founder and CEO; New York University; grew up in Vempalli, Andhra Pradesh; Emergent Ventures India grantee to build Veena, the company's first Hindi and English speech model), Bharath Kumar Kakumani (co-founder and CTO; New York University; grew up in Ongole, Andhra Pradesh)
The deal
$1.9 million seed led by South Park Commons, announced June 16th 2026; no other investors named. Instrument, valuation and whether SPC's Founder Fellowship terms applied are not disclosed on the SPC profile.
Reddy's claim is that 'India is not a text economy, it never was', and that most AI products treat voice as 'a microphone attached to a text system'. If the next several hundred million users arrive through speech, whoever owns native-sounding Indic voice owns the interface. This was not buildable two years ago: Veena, its first model, is a 3B-parameter Llama-based speech model, an architecture that only recently made expressive, streaming speech cheap enough to train on cloud credits. The non-obvious part is the data: the founders say their consumer app collects speech that is otherwise hard to get in these languages, which is the real bottleneck rather than model design. The case against: voice is the most crowded layer in AI, with ElevenLabs, Google and OpenAI all pushing multilingual real-time speech, and a $1.9 million seed is small against them; two of its models are open weights under Apache 2.0, which builds reach but gives away the product; and the leaderboard rankings it cites (#1 real-time Hindi on Voice Arena) are self-reported on its own site. Not a repeat of October's Duplex AI spotlight, which is enterprise full-duplex conversation, not Indic speech synthesis.
Flags wrong or unsupported LLM answers as they are generated by reading the model's own internal state, then holds, reroutes or queues them for review.
Founders
Vedant Gaur (solo founder; published interpretability research at ACL, EMNLP and NeurIPS while at Penn's M&T program, on what a model's internals reveal that its outputs don't; previously worked on post-training at Amazon and Robinhood)
The deal
YC standard: $500K, as $125K post-money for 7% plus $375K uncapped MFN.
This is interpretability research turned into a production control. Instead of paying a second model to re-read a sample of outputs, Latent puts a read-only hook on the hidden state of self-hosted models (shipped as a vLLM plugin) and scores every answer with no added latency, so a regulated buyer can check all traffic rather than a slice. Two years ago there was neither the probing literature that makes this credible nor the agent workloads, where one bad early step derails a long run, that make it urgent. Gaur's own published work is on exactly this gap between internals and outputs, which is a non-obvious claim about where errors become visible. The case against: most enterprise traffic runs on GPT, Claude and Gemini, whose internals no one outside the lab can read, and for those Latent falls back to a 'lightweight reader model', which is a small judge, the very approach it says it beats. Probes calibrated to one model also have to be redone with every model update. The 95% precision and '500x cheaper than an LLM judge' figures are the company's own, and it is a one-person team.
Beijing maker of rigs and software that record human demonstrations (video plus hand and body motion) and turn them into labeled training data for robots.
The deal
$100 million Series A+, closed in the third quarter of 2026 and disclosed November 2nd; strategic investors SAIC Motor, Baidu, Galaxy General Robotics, Shoucheng Holdings and Malaysia's Eastworlds, financial investors GL Ventures, CITIC Goldstone, CCB International and Huatai Zijin. Valuation not disclosed.
A picks-and-shovels bet on the robot data shortage: its Matrix platform syncs, cleans and labels recordings of household, factory and warehouse work for developers. The investor list doubles as a customer list, a carmaker, a search company with a robotaxi arm and a humanoid builder. The case against: demonstration data is being commoditized fast by simulation and by robot makers collecting their own, and TechNode gives no figures on hours recorded or paying customers.
Israeli startup selling a 'behavioral world model' that estimates how likely a person is to make a given decision, starting with how retail traders react to volatile markets.
The deal
$62.5 million seed from the founders and Infinity Group, which chairman Amir Gal-Or founded. More fundraising planned; valuation not disclosed.
About ten large brokerages use it to set margin and hedge exposure, and revenue is in seven figures. The round is mostly insiders' money, so the price has not been tested by outside investors. The company claims 98% accuracy on trading flows, measured as flows landing inside a predicted range. Its own CTO calls brokers using it against their customers 'a valid concern' without naming a safeguard beyond an ethics committee.
A US contract chemistry lab where AI and robots make drug-candidate compounds around the clock, charged per compound delivered.
The deal
$215 million Series A led by Lightspeed with IVP, Premji Invest and Lux, at an $865 million post-money valuation (reported).
Led by Kevin Weil, formerly of OpenAI, with Manohar Paluri from Meta as co-founder and CTO. No drugs of its own, so the risk is execution and price against overseas contract labs, not clinical. Only secondary reports were opened.
Continual learning infrastructure that turns how deployed AI agents are actually used into fixes and training signal, so the agent keeps improving after launch.
Founders
Advitya Chhabra (CEO; owned US equities market data at Citadel Securities; software engineering at the University of Waterloo), Rafit Jamil (co-founder; founded the User Reward Model team at OpenAI and worked on early self-healing agents and continual learning for ChatGPT), Adrian Aliu (co-founder and COO; urban planning at Waterloo; worked on large infrastructure projects in Canada and the UK; deferred Harvard Law)
The deal
Speedrun's published standard deal: up to $1M, as $500K for 10% upfront on a SAFE plus up to $500K more in the next qualifying round ($2M+) within 18 months. Proper Motion's executed terms are not disclosed.
Agents ship tuned once, then drift as customers, models and priorities change underneath them. Learning from live use needs two things that barely existed two years ago: agents in production at scale, and post-training cheap enough to run on a customer's own traffic. Rafit Jamil built the user reward model work inside ChatGPT, so he has seen what a reward signal drawn from real users does at that scale, including how it goes wrong. The case against: this is the labs' own roadmap, and memory, reinforcement fine-tuning and self-improving agents are being built into the model platforms themselves; the site says the team embeds with each customer's engineers, which reads as services and scales slowly; and an agent that rewrites its own behavior in production is a regression and reward-hacking risk many buyers will not accept.
Shanghai brain-computer interface maker with flexible cortical electrode implants, now funding trials and manufacturing.
The deal
RMB 700 million (about $98 million) Series B co-led by SDIC Juli, Shenzhen Capital Group and Qinzhi Capital, reported October 28th 2026.
The money is for clinical trials, regulatory work and manufacturing, and the company says it wants its system on the medical market at scale within two years from a 14,300-square-meter Jiangxi factory. State-linked lead investors make this a policy bet as much as a venture one. No valuation disclosed.
Trust and security infrastructure for AI agents: a decision control layer meant to let enterprise agents act with only the access each decision actually needs.
Founders
Yannis Panagis (CEO; founding engineer, then CTO, at a McKinsey spinout, building decision provenance systems and risk models for funds managing $265B; electrical engineering, Imperial College London), Anna Hruba (CTO; built authentication for 350K Bloomberg Terminal users and scaled Bloomberg's first AI agents to 10K engineers; researched fault-tolerant blockchains at Mysten Labs)
The deal
Entrepreneur First says it invests up to $250,000 in equity in teams that pass its Investment Committee and join the LAUNCH program in San Francisco, alongside an equity-free living-cost grant. Isochronic's specific terms are not disclosed on the EF profile.
The tagline, 'Useful, not privileged', names the problem this month's platform breaches made concrete: agents now act with real credentials, and enterprise access models were built for people and service accounts, not for software that decides what to do next. That buyer did not exist two years ago. Anna Hruba ran authentication for Bloomberg Terminal users and then rolled Bloomberg's first agents out to 10,000 engineers, so she has seen where agent permissions sprawl inside a regulated firm; Panagis brings decision provenance from fund risk work. The case against: the public site is a three-sentence manifesto with no product shown; identity incumbents such as Microsoft and Okta have launched agent identity products and own the directory; and agent security is crowded, with Speedrun's SR007 batch alone including Kairo, a 'continuous security plane for the AI agent era'.
Foundation models for hardware design: automating engineering design and validation, leaving manufacturing to others.
The deal
$130 million seed from investors including Eric Schmidt, John Doerr and DST Global (The Information); no valuation or lead disclosed.
Founded by Sebastian Thrun and Adrian Aoun. With &Truth's $70 million the same week, part of a run of nine-figure first rounds for models applied to engineering work. No customers or product specifications disclosed.
Inference platform for running models in production.
The deal
About $2 billion in two tranches at $26 billion and $30 billion, Sands Capital expected to lead, per Bloomberg on October 30th 2026; terms not final.
The top price is 2.3 times June's $13 billion, in the week OpenAI's round shrank by a third: private money is paying for the usage-linked part of the AI bill. Bloomberg's original was not opened; relays only.
Building a 'universal verification machine for code' on the Leanproof assistant, to prove AI-written software correct.
The deal
$70 million seed, announced October 29th 2026; investors not named.
Founders are Clark Barrett (Stanford, cvc5 solver), Ilya Sergey (NUS) and Chris Hsu (ex-Citadel), with Lean 4 co-architect Sebastian Ullrich on the team. The bet is that proof, not more testing, is how agent-written code gets trusted. The case against: EasyProver proved only 52.7% of real Rust statements the same week, and a proof is only as good as its specification.
A head camera, tactile glove and forearm EMG band recorded in sync, capturing how hard human hands grip, press and let go, as training data for robots.
Founders
Leo Li (Caltech computational neuroscience PhD student; neuroengineer at Neuralink; software engineer at Amazon; University of Toronto engineering science, robotics), Surya Pandiaraju (Johns Hopkins biomedical engineering PhD candidate; Waterloo BME; previously at Paradromics, NeuroPace, iota Biosciences and Sonera Magnetics), Runming (Tony) Wang (Johns Hopkins School of Medicine biomedical engineering PhD candidate; builds neural recording and stimulation hardware; EMG and motion-tracking research on motor-unit coordination at Emory and Georgia Tech)
The deal
YC standard: $500K, as $125K post-money for 7% plus $375K uncapped MFN. Company-specific terms not disclosed.
Robot learning from human video has a known hole: a camera sees a hand close around a cup but not how hard it squeezes. Preload's answer comes from brain-computer interfaces, where all three founders decoded neural and muscle signals: an EMG band reads grip effort that no amount of video can recover afterward, and the glove adds contact pressure. The buyers, frontier labs and robotics companies paying for physical-world data, barely existed as a market two years ago; surgical skill assessment is a second use. The case against: egocentric data collection is crowded and well funded, from Mecka to in-house programs at the humanoid makers; nobody has shown that policies trained with EMG and tactile channels beat vision-only ones at scale, and a robot has no forearm muscles to map that signal onto; and three PhD students splitting attention between factories and operating rooms is a focus risk.
Bengaluru protein-design company using generative models to design compact gene-editing enzymes from scratch, starting with crop editing that skips tissue culture.
Founders
Tanay Lohia (CEO; farmed ghost peppers in Assam, then led a World Bank-backed program insuring more than 25,000 farmers; self-taught in AI-led gene editing), Kutubuddin Ali Molla (scientific co-founder; senior scientist in the Crop Improvement Division at ICAR-National Rice Research Institute, Cuttack, and principal investigator of its Centre of Excellence on Genome Editing; CRISPR, base and prime editing research)
The deal
About ₹16 crore (roughly $1.8M) pre-seed co-led by Activate and Antler, with Spectrum Impact, DeVC and angels including Vijay Chandru and Paras Chopra, reported July 16th 2026. Antler's own check size and equity are not disclosed.
The bet is that the next gene editors will be designed rather than mined from bacteria, and that small, unpatented editors matter most where today's large, heavily licensed ones are hard to deliver: elite crop lines and in vivo therapy. Protein design good enough to attempt that is a capability of the last two years. The one public data point: in the GEM and Adaptyv RBX1 binder competition in April 2026 its design was the only one of 322 rated Strong, at 26 nM, which the company itself calls a single result from a side project. Pairing a self-taught founder with a working ICAR genome-editing lab gives it field access most AI-bio startups lack. The case against: binder wins are not editor activity; 'double-digit hit rates' is the company's own unpublished claim; 'patent-free' is a freedom-to-operate claim only litigation settles; and crop-editing timelines are set by regulators and seed companies, not enzyme design.
An AI-native personal injury law firm in California whose attorneys direct agents that gather evidence and prepare documents for approval.
Founders
Sun Woo Lee (CEO; co-founder and CEO of Cromatic; litigator at Orrick and Fish & Richardson; Stanford BA/MS with AGI and AI-safety research; Harvard JD), Nicholas Le (CPO; employee #2 at StrongSuit; California attorney with more than 700 personal injury cases on plaintiff and insurance-defense sides)
The deal
YC standard: $500K, as $125K post-money for 7% plus $375K uncapped MFN. Company-specific executed terms not reviewed, and neither page says which entity YC holds.
A firm, not software sold to firms: it takes cases on contingency and keeps the margin agents create, which only makes sense once agents can chase records, providers and insurers in parallel without a case manager per file. Le has worked 700-plus injury cases from both sides, so the claim about where the hours go is first-hand. The site calls Wingman Labs 'the engineering team behind the law firm'. The case against: California bars non-lawyers from owning law firms or sharing fees, so how investors capture value is the real question; solicitation rules cap growth, as the founders concede; and a contingency book takes years to resolve, so the economics stay unproven long after Demo Day.
Factory-built, liquid-cooled AI inference pods from a Romanian plant, meant to go live within 48 hours of delivery with no water connection.
The deal
More than $60 million raised to date, including a $50 million Series A; a Series B is planned.
Chief executive Flaviu Radulescu says build time per pod is down from about three weeks to a day and targets 30 MW of output a day by 2028. Claims $1 billion of committed interest covering more than 20 MW and a 397 MW site pipeline, all the company's own figures.
A model trained on US permit-archive building data that drafts whole building designs, including plumbing and HVAC, and estimates their cost for architects, engineers and developers.
The deal
$17 million seed led by Andreessen Horowitz, with SV Angel, XYZ and Zigg Capital.
Founded by Andy Elder and Arjun Mathur, Stanford roommates who each spent about a decade at Palantir. Trained on more than 100 million square feet of permit data and in use on 10 projects in Texas and California; it is aiming at AutoCAD and Revit, against Motif, Snaptrude and Arcol.
Runtime telemetry and policy for coding agents: Beacon, an open-source collector, normalizes prompts, tool calls, file edits and approvals across Claude Code, Codex, Cursor and 20-plus other harnesses.
Founders
Justin D'Souza (built core social-engineering detection systems at Doppel; New York), Shukan Shah (built core social-engineering detection systems at Doppel; the two have built together across several startups for more than seven years)
The deal
$5.5M seed, described as oversubscribed, with Tower Research Ventures, Lytical Ventures and 1984 Ventures; no lead named. Both founders went through SPC's Founder Fellowship (published standard: $400K for 7% on a standard SAFE plus $600K guaranteed in the next external round), but the post does not say the company took those terms.
Agent activity in a company is scattered across a dozen harnesses, each logging differently, and security teams cannot see what agents did with credentials or files. Beacon (MIT, about 3,000 GitHub stars) writes local JSONL in an OpenTelemetry-based schema and forwards to Splunk, Datadog or Sentinel, which makes it a sensible wedge. The founders started on securing AI-written code and widened to the runtime layer after seeing the fragmentation themselves. The case against: harness owners are shipping their own audit logs and spend reports, coverage varies by runtime with approvals and MCP events missing on many, and if the managed tier is mostly a SIEM pipe, the SIEM vendors can take it.
Eugenia Kuyda's personal-agent messenger that builds small interfaces on demand for a task.
The deal
$55 million at a $1.25 billion valuation, led by IVP, with a16z, Forerunner and the Alexa Fund.
Pivoted from prompt-built mini apps. No user, revenue or retention figures disclosed, so the price rests on the founder, who built Replika, and the category.
synapse, an agentic engineering platform for hardware built on SysML v2: architecture, requirements, verification and version control with AI agents.
Founders
Sergi Company Aguilar (CEO; model-based systems engineering at Airbus, Rolls-Royce and on four ESA missions; system models at GMV; UPC and Cranfield), Alexandre Gol i Mestre (CTO; launched and operated 8 satellites across Open Cosmos, LuxSpace and Exotrail; University of Barcelona and Cranfield), Rafael Dalosto (founding engineer; serial founder, software since 2010)
The deal
Techstars standard: $220K, as a $20K post-money convertible equity agreement for 5% plus a $200K uncapped MFN SAFE. Company-specific executed terms not reviewed.
Systems engineering for satellites lives in formal models that few engineers can write, and SysML v2 gave those models a textual, machine-readable form only recently. That is what lets an agent read and edit them, and it is why this could not have been built two years ago. The founders have done this work on ESA missions and satellite operations, not just around it. The case against: aerospace buys slowly, tools like Cameo and Capella are entrenched, and an agent that edits a requirements model needs a verification story a regulator will accept before a prime contractor lets it near flight hardware.
Korean maker of the RNGD inference chip, in mass production since January, with a third-generation accelerator in development with Broadcom.
The deal
Series D. Contested: Bloomberg reports $620 million at a $2.9 billion valuation; the company's announcement as reported by the Seoul Economic Daily gives 820 billion won (about $570 million) at about 3.8 trillion won (about $2.6 billion). National Growth Fund 370 billion won; Infinitum Partners the largest overseas investor.
A state-backed national champion in inference chips, priced at about four times its last round. The bet is that inference spreads to enough buyers that a non-Nvidia chip with Korean customers like LG AI Research and Samsung SDS can scale. The case against: Groq, Cerebras and others are selling the same pitch with more capital, and a Broadcom-designed third generation is years from revenue. A listing is reportedly considered as early as 2027.
Evaluations for frontier models in electrical engineering, built from real research projects that mix simulation, specialized software and physical lab measurements.
Founders
Trishala Jain (CEO; previously Gemini for Science at Google Labs; EECS, UC Berkeley), Mark Nour (CTO; previously Netflix and Google Cloud; Berkeley CS)
The deal
YC standard: $500K, as $125K post-money for 7% plus $375K uncapped MFN. Company-specific executed terms not reviewed.
Labs are now claiming research results in mathematics and physics, and the benchmarks that check them stop at the screen. Evident builds tests where the answer depends on a bench measurement, which is where a model's scientific judgment meets physical constraints. Jain worked on Gemini for Science, so she has seen what labs measure and what they do not. The case against: two founders selling evaluations to a handful of labs is a consulting business with a concentration problem, lab evals are expensive to keep fresh, and the labs can hire the same electrical engineers themselves.
Toronto enterprise and sovereign-AI model company, merging with Germany's Aleph Alpha.
The deal
Initial close of a Series E targeting up to C$7 billion at a C$35.7 billion valuation. Schwarz Group leads; Government of Canada (up to C$1.6 billion of equity, per the prime minister), Bell Ventures and Bullhound are new; Inovia, Nvidia and Prosperity7 return.
The buyers are a state and a European retailer, and the pitch is sovereignty. Cohere reports 165% annualized-revenue growth but no absolute figure, so the multiple cannot be checked.
Maker of the Devin coding agent and owner of Windsurf, weighing a new round after investor offers at $70 billion to $80 billion.
The deal
No terms agreed. Business Insider, citing people familiar, reports offers at $70 billion to $80 billion; the last round, in September, raised more than $2 billion at $48 billion.
Talks are early and could change. At those prices investors would pay roughly 80 to 90 times reported annualized revenue of about $900 million. The case against: GitHub said the same day that its own routing system matches top coding agents at half the cost, and coding is the category where the platform owners compete hardest.
Checks each supplier invoice line against the supplier agreement, purchase order and delivery record before payment, for UK finance teams.
Founders
Johannes Haug (CEO; investment banking at Arma Partners and private equity at Triton Partners; BI Norwegian Business School), Devvrat Joshi (CTO; founding software engineer at a £100m+ startup; master's in Intelligence and Machine Learning, Imperial College London; IIT Gandhinagar)
The deal
Techstars standard: $220K, as a $20K post-money convertible equity agreement for 5% plus a $200K uncapped MFN SAFE. Company-specific executed terms not reviewed. [A](https://www.techstars.com/investment-terms)
Three-way matching compares numbers; the overcharges Stemma targets live in contract prose, such as a fuel surcharge that should have dropped a band when a diesel index fell. Reading every agreement against every invoice line only became cheap enough with current models. Haug's claim comes from private equity, where he watched portfolio companies overpay suppliers with no practical way to catch it. The case against: recovery audit on contingency is an old, crowded trade, the read-only ERP integrations are the real work, and the company already appears to have moved from transport insurance claims leakage, as Techstars listed it, to supplier invoices. [A](https://www.stemma-ai.com/)
Forecasting agents that return calibrated, continuously updated probabilities on global macro events and markets.
Founders
Veniamin Veselovsky (built Preseen's data layer of crawlers and normalization pipelines; prior employer not stated in profile), Theo Summer (left a software job in 2025 after his forecasting bot won Metaculus's Fall tournament)
The deal
SPC led the seed round, per Aditya Agarwal's July 30, 2026 post on the profile; amount and instrument not disclosed. SPC's published Founder Fellowship standard is $400K for 7% on a standard SAFE plus $600K guaranteed in the next external round, but the profile does not say Preseen took those terms. [A](https://www.southparkcommons.com/companies/preseen)[A](https://www.southparkcommons.com/news/f26-founder-fellowship)
Forecasting is one of the few AI tasks with delayed ground truth: every question resolves and becomes a labeled example, so the calibration loop compounds. Summer built the same tool on GPT-4 in 2022, discarded it over math errors, then rebuilt it in 2025 and won, which is a clean test of what changed in the models. Its record is kept by Metaculus tournaments, not the company. The case against: SPC's own post concedes human forecasters still hold a head-to-head edge, tournament questions are not the bespoke macro calls a fund pays for, and calibration across hundreds of resolved questions says little about the rare tail events that matter most. [A](https://www.southparkcommons.com/companies/preseen)
Inference cloud that finances, deploys and operates Cerebras, SambaNova and Positron AI hardware, selling bare metal or guaranteed tokens-per-second SLAs.
The deal
$100 million Series A co-led by Blackbird and Abstract, with Fuse Venture Partners and Village Global; about $800 million valuation per TechCrunch as cited by TechStartups, not stated by the company. Separate from a previously announced $400 million Upper90 debt facility.
The bet is that agent and reasoning workloads make token speed worth paying for, and that a cloud matching each workload to non-Nvidia accelerators can sell it. The 40x speed figure is the company's own and depends on workload. The case against: first major deployments go live only in Q1 2027, the Upper90 debt makes this a leveraged hardware business, and no customers are named.
Reproducible simulations of failures in distributed software.
Founders
Chaitanya Choudhary (founding engineer at Grit; On Deck and CW; Harvard Law research assistant)
The deal
YC standard: $500K, with $125K for 7% plus $375K on an uncapped MFN SAFE. Company-specific executed terms not reviewed. [A](https://www.ycombinator.com/deal)
Workers IO runs unmodified systems through simulated network, disk, clock and crash faults so failures can be replayed. Choudhary brings experience building developer infrastructure. Simulation predates current AI; the case is that expanding volumes of generated code create demand for cheaper, repeatable failure testing. The objection is coverage: a convincing simulation can still omit the production condition that causes an outage.
Independent robot benchmarks on physical hardware and fixed task setups.
Founders
Arnav Bhalla (Motif Neurotech and Rice BCI Lab; left Rice Bioengineering), Rohith Yelisetty (TJHSST; individual prior role not specified in profile), Adithiya Balaguru (TJHSST; individual prior role not specified in profile)
The deal
Speedrun standard: $500K for 10% via SAFE, plus up to $500K in a qualifying $2M+ round within 18 months. Company-specific executed terms not reviewed. [A](https://speedrun.a16z.com/faq)
General-purpose robot models create a need to compare systems on identical physical tasks rather than selected demonstration videos. Coop proposes fixed setups and real hardware, with a team drawn from robotics and engineering study. The strongest objection is transfer: a robot that succeeds on the benchmark rig may fail at a customer site. Commercial partnerships also need separation from independent scoring.
Customer-specific models supplying prompts during live sales calls.
Founders
Jared Zelman (Cicero.ai; Summit Partners and private equity), Raphael Assor (Explorium; engineering lead at CrowdStrike)
The deal
Speedrun standard: $500K for 10% via SAFE, plus up to $500K in a qualifying $2M+ round within 18 months. Company-specific executed terms not reviewed. [A](https://speedrun.a16z.com/faq)
Post-training on customer conversations could make live recommendations specific enough to help a seller, rather than merely transcribe the call. The team combines deal-sourcing and engineering experience. The strongest objection: prompts can distract sellers, and established call platforms can add the same capability. Controlled conversion tests matter more than the profile’s launch metrics.
Solar-powered modular data centers for AI inference.
Founders
Sid Ravikumar (founding engineer at Jelly.fm; Lotus Laboratories), Sajan Khosa (founded AJMER), Arsh Sekhon (University of Waterloo graduate)
The deal
YC standard: $500K, with $125K for 7% plus $375K on an uncapped MFN SAFE. Company-specific executed terms not reviewed. [A](https://www.ycombinator.com/deal)
The thesis is that inference can move into modular solar-powered sites and escape some grid-connection delays. Ravikumar brings software deployment experience. The objection is physical: intermittent power, storage costs and distributed operations can erase the claimed advantage. Modular computing predates this company; economically competitive AI inference is the claim that still needs proof.
Sensors and AI software that monitor industrial equipment for impending failures; raised $130M at a $1.4B valuation per October 23rd funding reports.
Industrial condition monitoring is where AI meets a measurable dollar loss, and the buyers have maintenance budgets rather than innovation budgets. The round is reported by funding trackers rather than a company announcement we could locate, so the valuation carries their sourcing.
Verifies chip authenticity by exciting components with RF signals and reading their nonlinear harmonic response as a fingerprint, so counterfeit or tampered parts can be caught non-destructively before assembly.
Founders
Erik Hosler (CEO; PhD in physical chemistry, UC Berkeley; career in semiconductor manufacturing), Adam Wilson (CTO; co-founded Sphero, whose technology powered Disney's BB-8; hardware and cybersecurity background)
The deal
Techstars standard: $220K, as $20K for 5% common equity plus a $200K uncapped MFN SAFE. Company-specific terms not disclosed.
Counterfeit chips became a policy problem the moment export controls made restricted parts worth faking, and every tariff and control regime since has widened the gray market this inspects for. RF side-channel fingerprinting without decapping the part is a genuinely physical moat, and a founder pair of a fab physicist and the Sphero hardware co-founder fits the problem. The case against: reference signatures need known-good golden units for every part number, which is a library someone has to build and maintain, and the obvious buyers, large distributors and defense primes, move at qualification speed.
Private foundation models trained on a bank's or insurer's own transaction and interaction data, run inside the customer's cloud, predicting credit, fraud, collections, churn and lifetime value.
Founders
Gabriel Noya (solo founder and CEO; Stanford electrical engineering and computer science, class of 2026)
The deal
YC standard: $500K, as $125K post-money for 7% plus $375K uncapped MFN.
Transaction foundation models are a post-2024 idea: until models could learn from billions of raw payment events, this was feature engineering, and Lyon's profile claims a model trained on 28 billion transactions that picked premium-card converters about four times more precisely than the customer's rules. The in-their-cloud deployment answers the data-custody objection. The case against: a solo founder straight out of Stanford selling to banks is the hardest go-to-market in software, incumbents from Stripe to Feedzai sit on bigger corpora, and the 4x figure is the company's own, unaudited.
Cultured biological neurons as a computing substrate.
Founders
Michael Domarkas (Cambridge Natural Scientist; neuronal-culture work supported by Emergent Ventures)
The deal
YC standard: $500K, as $125K post-money for 7% plus $375K on an uncapped MFN SAFE. Executed company-specific terms were not reviewed. [A](https://www.ycombinator.com/deal)
The proposed advance is scaling living neural tissue beyond its usual oxygen-supply limits, with memory and computation in the same material. Domarkas brings hands-on tissue-culture experience. The case against: maintaining cells and reliably reading and writing information may erase the claimed computing advantage. A company-reported game demonstration does not establish useful language-model training.
Devices and models to monitor cancer using electrical activity.
The deal
$45 million Series A co-led by Founders Fund and the OpenAI Startup Fund.
The company reports mouse experiments predicting tumor growth and says human device testing has begun. Continuous feedback could change treatment monitoring, but clinical benefit from adaptive therapy has not been established.
Agents that interview workers and implement business automation.
Founders
Dima Khanarin (McKinsey; Everclear), Gleb Sidora (CTO; prior background not stated in the public profile)
The deal
Speedrun standard: $500K for 10% via SAFE, plus up to $500K in a qualifying $2M+ round within 18 months. Executed company-specific terms were not reviewed. [A](https://speedrun.a16z.com/faq)
Agents could make interviewing an entire workforce affordable and connect the findings directly to implementation. Khanarin brings organizational-transformation experience. The case against: employee descriptions can miss the real bottleneck, while bespoke implementation can keep this a consulting business. Claimed customer savings do not establish net returns after integration and review costs.
Computing infrastructure for automated software builds and tests.
The deal
$66 million Series C at a $1 billion valuation, led by Sequoia.
AI-generated code increases the volume of builds and tests. The company plans to expand capacity; workload growth is self-reported and does not measure the correctness of customer software.
Agent swarms that formalize and verify reasoning, starting in biology.
Founders
Ryan Min June Shin (Harvard mathematics, left to found; former world-record-holding powerlifter), Sohum Sukhatankar (Harvard mathematics, left to found; national spelling bee champion)
The deal
YC standard: $500K, comprising $125K for 7% and $375K on an uncapped MFN SAFE. Company-specific executed documents were not reviewed. [A](https://www.ycombinator.com/deal)
The bet is that verification, not generation, is the bottleneck in agent swarms doing science, and that lessons from formal mathematics transfer to biology. Two years ago the agent-swarm substrate this assumes did not exist. The case against: biology lacks mathematics' ground truth, so agents checking each other can converge on a shared wrong answer, and a two-person team is selling rigor in a field they are new to.
Service and diagnostics layer for deployed robot and autonomous-machine fleets.
Founders
Robert Doherty (mission-critical infrastructure at Citadel Securities; industrial technology at McKinsey), Pranit Gupta (led AI deployments in industrials and heavy machinery at Palantir)
The deal
Current Speedrun standard: $500K for 10% via SAFE, plus up to $500K in a qualifying $2M+ round within 18 months. Company-specific executed terms are not disclosed. [A](https://speedrun.a16z.com/faq)
Every robot shipped this year becomes a field-service liability next year, and the founders have maintained exactly that kind of infrastructure at Komatsu-grade scale. Turning each field fix into a playbook that feeds design is the non-obvious part. The case against: OEMs guard service revenue and machine data, the profile shows no customers yet, and service businesses historically scale with people rather than software.
Code graphs for understanding and checking agent-written changes.
Founders
Dan McGee (UC Berkeley; built policy simulation at Illumio), Benjamin McIntyre (University of Cincinnati; Linux and FOSS club; political campaign technology)
The deal
YC standard: $500K, as $125K post-money for 7% plus $375K on an uncapped MFN SAFE. Executed company-specific terms were not reviewed. [A](https://www.ycombinator.com/deal)
Agents can produce changes faster than a reviewer can read them. Graphide proposes keeping system relationships visible as that work happens. McGee brings policy-simulation experience. The case against: a graph can omit dynamic behavior or become stale, giving reviewers confidence without catching the defect. No independent review-quality measurement was established.
Sales data and outreach infrastructure exposed to agents through MCP.
Founders
Luis Costa Laveron (MIT data science; machine-learning systems at Remitly), Enrique Goudet (PeakMetrics; data pipelines used by Bloomberg; Occidental College)
The deal
Speedrun standard: $500K for 10% via SAFE, plus up to $500K in a qualifying $2M+ round within 18 months. Executed company-specific terms are undisclosed. [A](https://speedrun.a16z.com/faq)
Cluster reorganizes sales tools around agents that must choose data sources and complete actions, rather than people navigating dashboards. The founders bring production data and machine-learning experience. The case against: combining providers cannot repair inaccurate prospect data, and easier outreach can increase unwanted contact. Its speed and cost comparisons are company claims.
AI-assisted military logistics questions and simulations.
GeekWire reports a $35 million seed round led by Shield Capital and Conviction. A Defense Innovation Unit prototype contract offers a testing route. The case against: simulation usefulness depends on operational data that may be missing or stale. Production-scale reliability is not established. [B](https://www.geekwire.com/2026/ex-marine-and-microsoft-vets-launch-watchtower-labs-with-35m-to-bring-ai-to-military-logistics/)
Simulated users for testing product decisions before launch.
Founders
Nicole Orsak (Uber Eats product; Twitter and Disney+; Stanford), Yuqi Li (simulated-user work at Nvidia; founding AI work at Doctronic; Waterloo and Carnegie Mellon)
The deal
Current Speedrun standard: $500K for 10% via SAFE, plus up to $500K in a qualifying $2M+ round within 18 months. Company-specific executed terms are not disclosed. [A](https://speedrun.a16z.com/faq)
Generating users that interact with a proposed product could make early testing cheaper than recruiting a fresh panel for each change. The founders combine consumer-product and simulated-user experience. The case against: a convincing population can reproduce training-data preferences and give false confidence about actual customers. Independent predictive accuracy has not been established.
Training and evaluation environments for medical AI agents.
Founders
Aryaa Bhatia (Plureto; Lawrence Livermore National Laboratory research intern; UC Berkeley), Tej Patel (SimCare AI; Hippocampi; Cedars-Sinai machine-learning research)
The deal
YC standard: $500K, comprising $125K for 7% and $375K on an uncapped MFN SAFE. Company-specific executed documents were not reviewed. [A](https://www.ycombinator.com/deal)
Medical agents need practice making sequences of decisions, beyond answering static exam questions. Archangel proposes environments built from deidentified health records with physician judgment. The founders bring clinical-agent and research experience. The case against: unrepresentative records or flawed rewards can teach confidently wrong behavior without improving patient outcomes.
Linux-based operating system exposing application actions to agents through MCP.
Founders
Xaver Emmerling (CEO; prior background not disclosed on the profile), Quinten Wolff (patent workflow automation at BDPE; computer science at Technical University of Munich)
The deal
YC standard: $500K, as $125K post-money for 7% plus $375K uncapped MFN. Company-specific executed documents were not reviewed. [A](https://www.ycombinator.com/deal)
Tilo replaces screenshot-and-click interaction with actions exported by applications, while people retain a visible interface and undo history. Wolff built legal workflow automation before tackling the interface itself. Capable computer agents make this redesign timely. The case against: every application needs a dependable action surface, and replacing an operating system creates a much larger adoption barrier than installing an agent. The profile supplies no independent cost comparison.
Lume, a lamp-shaped home robot with arms intended to fold laundry.
Founders
Aaron Tan (Stanford robotics postdoc; University of Toronto PhD; GM autonomous-driving hardware and Apple design), Angus Fung (University of Toronto robotics PhD; machine-learning research at Vector Institute and AMD)
The deal
Current Speedrun standard: $500K for 10% via SAFE, plus up to $500K in a qualifying $2M+ round within 18 months. Historical cohort-specific executed terms are not disclosed. [A](https://speedrun.a16z.com/faq)
A stationary pair of arms avoids the locomotion problem that makes a household humanoid expensive. The opportunity depends on learned manipulation handling deformable clothes outside a staged demonstration. The founders combine robotics research with hardware work. The case against: fabric variety, recovery from tangles and safety around children can defeat a polished demo. The profile says summer shipping was planned; it does not establish that deliveries happened.
Chinese provider of 3D vision, scene reconstruction and simulation software.
Renaissance Capital reports an October 21st SEC filing seeking up to $35 million. Nasdaq is the intended venue; no trading symbol or pricing terms were selected. The filing stage is recorded from the specialist report, not a direct review of the prospectus. [B](https://www.renaissancecapital.com/IPO-Center/News/122346/china-based-ai-vision-solutions-provider-sensing4d-ai-files-for-a-35)
Robotics company developing dexterity through force and touch.
Tech.eu reports $30 million in pre-seed funding, with Yann LeCun and Arthur Mensch among the backers. The company combines sensing hardware and learned control. The case against: independent reliability and production economics have not been established. [B](https://tech.eu/2026/10/22/mistral-boss-and-yann-lecun-back-swiss-robotics-startup-enact-in-30m-pre-seed-round/)
Safety software for industrial robots working around people.
Founders
Denis Azarov (BMW innovation programs; previously built human-behavior models for autonomous driving), Jakob Thumm (Stanford Autonomous Systems Lab postdoctoral researcher; Technical University of Munich PhD)
The deal
Speedrun standard offer: $500K for 10% via SAFE, plus up to $500K in a qualifying round of at least $2M within 18 months. Company-specific executed terms are undisclosed. [A](https://speedrun.a16z.com/faq)
The founders combine human-behavior modeling with formal robot safety research. As learned robot policies enter shared workplaces, a safety layer spanning robot types could let operators adopt them without rebuilding each controller. The case against: integration and site-specific certification can consume the savings. The profile says TÜV-backed, but does not establish the scope of an independent safety certification. [A](https://speedrun.a16z.com/companies/scaledriveai)
Verifies which source code is running inside a protected computing environment.
Founders
Anton Livaja (co-founded security firm Distrust), Ksenia Lesko (strategy and operations at Distrust; customer education at Ada), Lance Vick (security roles at Turnkey, BitGo, Unit 410 and Pebble; Stagex Linux)
The deal
YC standard offer: $500K, comprising $125K for 7% and $375K on an uncapped MFN SAFE. Company-specific executed terms are not disclosed. [A](https://www.ycombinator.com/deal)
Agents handling sensitive data make the gap between reviewed source and running code a purchasing problem. Caution ties those together through enclave verification, with AWS Nitro supported today. The case against: protected execution predates the current agent market, and cloud providers could absorb the verification layer. Additional hardware support is a roadmap, not current coverage. [A](https://www.ycombinator.com/companies/caution)
The team’s Moda experience exposes a useful problem: generated assets still need editing inside real creative projects. Stronger tool-using models make cross-application execution more plausible than a prompt-only design service. The case against is integration fragility: a model can make attractive output while damaging project structure or requiring more cleanup than manual work. Test an editable handoff, not just the rendered clip. [A](https://www.ycombinator.com/companies/voyager)
General-purpose industrial robots for construction, manufacturing and energy sites.
Founders
Wei Ding (co-founder and CEO), Wen-Loong Ma (co-founder and CTO), Christopher McQuin (co-founder, head of hardware), Wenda Wang (co-founder, AI)
The deal
Terms not disclosed on the SPC profile; South Park Commons typically writes first checks in the $1M to $10M range. One secondary profile reports $180M across seed and Series A with Caterpillar and Bechtel strategic money, unverified by any second source.
Operated as Under Control Robotics until emerging from stealth in March; its Moby robot targets hazardous site work, lifting 60 lb over ramps and rough terrain, with one profile placing the founding team at Apple, SpaceX, NASA and Caltech. Shipping into construction and energy sites is the uncrowded end of humanoids. The case against: deployment-hours claims are the company's own, and strategic investors with pilot sites can keep a robotics company looking deployed long after the economics have stalled.
South Korean AI computing and data-center provider.
Listed October 20th. Reported intraday losses reached 33%; AP’s separate snapshot said 30%. These are not averaged and neither is presented as a verified closing return. [B](https://cincodias.elpais.com/mercados-financieros/2026-10-20/elice-se-desploma-en-su-debut-bursatil-y-reaviva-las-dudas-sobre-el-negocio-de-la-inteligencia-artificial.html?outputType=amp) Close per Korea Exchange figures: ₩61,500 against a ₩90,500 offer, down 32.04%, after a 388.8x institutional book. [B](https://en.sedaily.com/finance/2026/10/20/elice-plunges-30-percent-in-kosdaq-trading-debut)
Government workflow software is old. The newer proposition is turning ordinary-language requests into working intake and approval systems while handling unstructured records. That could reduce custom implementation work in small agencies. The case against is legal and operational specificity: a plausible workflow can still apply the wrong retention or disclosure rule. Judelson’s public-sector background is relevant, but deployment audits must establish the product’s reliability. [A](https://www.southparkcommons.com/companies/common-sense/)[A](https://commonsensetech.ai)
Hosting Hermes, OpenClaw, Claude Code or any Docker image in isolated, persistent sandboxes with per-minute billing is a bet that agents become a utility whose cost structure matters more than whose model runs inside. A cited customer claims cost per agent fell from $85 to $4.94 a month, company-reported. The case against: sandboxes are exactly what hyperscalers and the harness vendors themselves will commoditize, and a solo founder is racing both.
AI agents for power delivery that link utility studies, designs, equipment records and schedules to drive interconnection work.
Founders
Tarun Batchu (patented environmental engineer; left Berkeley M.E.T. to build Vela), Tony Li (Tesla; federal energy and defense contracts at Quasar Energy Group; published energy research at 16)
The deal
$1.3M from a16z Speedrun and Z Fellows, per the profile; the split between the two is not disclosed.
Grid interconnection is the binding constraint on the data-center buildout: Nscale's Loughton site died on it and NextDC halted its stock over it this morning. Vela's agents evaluate power configurations, prepare interconnection submissions and track what capacity can be energized and when, selling to the developers and EPC firms stuck in those queues. The strongest case against: interconnection timelines are set by utility engineers and regulators, not by how fast the paperwork is drafted, so better submissions may compress the queue by weeks while the backlog is measured in years.
AI processors, Galaxy servers and licensable chip designs.
Closed a $1.5 billion Series E at a $6.2 billion post-money valuation, led by Trace Capital and Innovation X. Capital supports manufacturing and commercial expansion. Financing does not establish production economics against competing systems. [A](https://tenstorrent.com/newsroom/tenstorrent-announces-series-e)
General-purpose robotic experimentalists that observe, hypothesize and run physical experiments.
Founders
Owen Tower (PhD physics, Brown; research at NASA JPL and CERN/LHC), Cooper Niu (PhD physics, Brown; built ALBERT, an autonomous particle-physics theorist; Fermilab NOvA), Benedict Quartey (PhD computer science, Brown; long-horizon autonomy at the Boston Dynamics RAI Institute)
The deal
YC standard: $500K, as $125K post-money for 7% plus $375K uncapped MFN.
Formerly SuperRadiant. Three Brown physics and robotics PhDs betting that discovery needs embodiment: a system that designs an experiment, runs it on hardware and reads the result, closing the loop text-only lab agents leave open. Today's Into the Unknown benchmark, where all 135 agent-written research papers failed working scientists' conditions, is both their market and their warning. The strongest case against: autonomous labs have burned money for a decade on exactly this promise, and three physicists with no disclosed product or customers are a research group until proven otherwise.
Controlled-light imaging for inspecting consumer-electronics components.
Founders
Erica Diaz (mechanical engineering at Meta, Google and Magic Leap; Columbia), Imani Gosserand (software engineering at Twitch; Columbia computer science), Iris Liao (Citadel quantitative research; machine learning on 3D meshes at Imperial College London)
The deal
YC standard: $500K, as $125K post-money for 7% plus $375K uncapped MFN. Company-specific variations are not disclosed. [A](https://www.ycombinator.com/deal)
Machine vision itself is old. The newer claim is adapting inspection from one good example as component designs change. Diaz’s experience with factory vendors gives the team a concrete failure to attack. The strongest objection is false rejection under new materials or lighting: a detector that scraps good parts can erase the savings. The profile supplies no independent production validation. [A](https://www.ycombinator.com/companies/hundred)
Continued training and reasoning distillation on open base models.
Founders
Drishan Arora (natural-language machine learning, reinforcement-learning alignment and information retrieval; prior employer not specified in profile), Dhruv Malrana (DeepMind and Google Search product management; Google Search and Microsoft Office technical leadership)
The deal
SPC reports a $43 million round led by TQ Ventures. The profile does not disclose the signed fellowship investment terms. [A](https://www.southparkcommons.com/companies/deep-cogito)
The company began in 2024; the current signal is how stronger open bases enable a small team to continue training and distill reasoning into specialized models. Its investor describes an earlier approach being overtaken by new base releases. That is also the strongest objection: suppliers can erase a specialist’s advantage overnight. Seek independent comparisons including training cost; investor claims alone do not establish an economical moat. [A](https://www.southparkcommons.com/companies/deep-cogito)
Agents that maintain company context while carrying out business work over extended periods.
Founders
Spandan Das (DatologyAI, Nvidia, Apple and NASA; Carnegie Mellon computer science), Amit Kallakuri (Clay and a16z enterprise and fintech; Dartmouth), Bradley Xu (Reve, Sutter Hill Ventures, Sigma and Amazon; Carnegie Mellon mathematics and computer science)
The deal
Company-specific terms are not disclosed. Speedrun currently advertises $500K for 10% upfront on a SAFE, plus up to $500K in a qualifying round of at least $2M within 18 months. These current program terms do not verify Sterling’s signed deal. [A](https://speedrun.a16z.com/faq)
The newer bet is persistent execution across changing company data, rather than an assistant answering isolated questions. Sterling describes company-specific evaluations and action boundaries, a concrete way to test that bet. The profile does not establish that its approach is unique. The strongest objection is attribution: renewals and advertising results depend on people and market conditions, so claiming ownership of an outcome can conceal expensive human supervision. Seek results with intervention costs included. [A](https://speedrun.a16z.com/companies/sterling) Weekly synthesis: compare persistent execution with FDE-Gym requirements discovery. Ask for customer outcomes including failed attempts, stakeholder time and human recovery, since a completed action can still miss the business requirement. [A](https://arxiv.org/abs/2610.22350)
Observes agent actions and applies operational boundaries during execution.
Founders
Vinod Kone (former Harness engineering VP and Twilio Segment senior director), Micheal Benedict (previously led infrastructure, security and AI platforms at Airtable), Prasanth Ganta (previously an engineering leader at Amazon, including generative AI systems)
The deal
YC standard: $500K, as $125K post-money for 7% plus $375K uncapped MFN. Company-specific variations are not disclosed. [A](https://www.ycombinator.com/deal)
Agents acting across tools create a newer control problem than collecting application logs. Stacktrace proposes enforcing boundaries while work runs. Its founders have operated security and infrastructure systems, but the profile does not demonstrate a distinctive detection method. The strongest objection is coverage: an action outside the monitored path can defeat a promise to see every action. Look for independent escape tests and explicit limits on what the system can stop. [A](https://www.ycombinator.com/companies/stacktrace) Weekly synthesis: the reporting-integrity failures make access to raw actions more valuable than another generated success summary. The case against remains coverage: test an action outside the monitored path and record whether it can actually be stopped. [A](https://arxiv.org/abs/2609.36139)
Ranks marketing ideas before using generative models to produce the content.
Founders
Benjamin Bekken (previously a Minecraft video creator and marketer), Håkon Græsberg (co-founded Wepost; earlier machine-learning work at DNV and Norconsult), Trond Hauga (founded a marketing agency and co-founded Wepost)
The deal
YC standard: $500K, as $125K post-money for 7% plus $375K uncapped MFN. Company-specific variations are not disclosed. [A](https://www.ycombinator.com/deal)
Cheap generated content makes choosing an idea a different bottleneck from producing it. Native connects proposed ideas, human edits and subsequent performance, a potentially useful feedback loop. The founders bring experience making and distributing content. The case against: audience response changes, and engagement can reward repetition rather than business results. A prospective test against a simple historical baseline would be more persuasive than customer counts. [A](https://www.ycombinator.com/companies/native)
An agent intended to execute tenant-lifecycle work and property accounting across existing software.
Founders
Julio-Cezar Scerbina (competitive chess and mathematics background; prior employer not stated), Christian Brown (theoretical physics graduate from UCL; neutrino-detection research), Andrey Dobrov (previously enterprise sales at Copper.co)
The deal
Company-specific terms are not disclosed. Speedrun advertises up to $1M of investment; this is a program ceiling, not a verified Alven financing. [A](https://speedrun.a16z.com/faq)
The newer opportunity is executing work across tenant conversations and accounting systems rather than generating a suggested reply. Alven claims that scope, but its profile does not establish a proprietary operational insight from the founders. The strongest objection is exception handling: maintenance emergencies and disputed charges need accountable decisions. Watch for evidence of safe escalation and reconciled books; the promise to replace managers remains a company claim. [A](https://speedrun.a16z.com/companies/alven)
AI decision support for planning, scheduling, allocation and routing in physical operations.
Founders
Hindy Rossignol (agentic AI for fusion research at MIT Energy Initiative; ML at McKinsey), Bernardo Chalita (operations research at Sanofi; AI research at BMW; MIT optimization and ML)
The deal
YC standard: $500K, comprising $125K for 7% plus $375K on an uncapped MFN SAFE. Company-specific variations are not disclosed. [A](https://www.ycombinator.com/deal)
The thesis worth testing is whether agents can turn changing operational context into useful decisions across planning and execution. The founders combine optimization with experience in physical research and manufacturing. The strongest objection is that scheduling and routing software long predate modern language models: the company must demonstrate a new ability to interpret exceptions or coordinate decisions, not merely repackage a solver. The profile does not yet establish that advantage. [A](https://www.ycombinator.com/companies/opera)
Law firm using its own AI systems to execute transactions at fixed fees.
Founders
Cem Ucan (founding legal engineer at Legora; attorney and legal-tech lead at Baker McKenzie), Kieran John Keenan (founding evaluation engineer at Legora; previously founded a legal-AI startup)
The deal
Company-specific investment terms are not disclosed on the profile. Speedrun advertises investment of up to $1M; this is a program ceiling, not a verified Athena round. [A](https://speedrun.a16z.com/faq)
Fixed-fee legal work makes the firm retain the benefit when agents reduce the cost of a transaction. The team combines deployment experience with a founder who built legal-model evaluations, addressing the gap between generating a document and accepting it. The strongest objection is review liability: if experienced lawyers must reconstruct every step, faster drafting may not lower total delivery cost. The profile does not disclose independently measured margins or error rates. [A](https://speedrun.a16z.com/companies/athena)
Simulates business applications so teams can test agents against reproducible failures.
Founders
Arya Kanna (left UC Berkeley), Saad Asad (prior background not disclosed on the retrieved profile), Lalith Posam, CTO (prior background not disclosed on the retrieved profile)
The deal
Company-specific SR006 terms not disclosed. Current Speedrun standard: $500K via a SAFE for 10%, plus up to $500K in the next qualifying round of at least $2M within 18 months; historical terms may differ.
Sparta recreates services such as Stripe, Zendesk and Salesforce for deterministic scenarios, including broken credentials. Agents operating across several applications make this more useful than an ordinary mocked endpoint. The strongest objection is drift: a simulation can remain perfectly reproducible while the real service changes. Watch whether failures found in these tests predict production incidents. [A](https://speedrun.a16z.com/companies/sparta)
Corporate spending software expanding into AI-assisted finance.
Announced $1.85 billion in primary financing at a stated $60 billion valuation. Annualized revenue is a management-reported pace, not completed-year sales. [A](https://www.prnewswire.com/news-releases/ramp-reaches-60-billion-valuation-as-annualized-revenue-tops-2-billion-302909315.html)
Announced more than $100M in financing including equity and a credit facility, at a $1.25B pre-money valuation. The equity-credit split is not disclosed. The financing backs Horizon and Recursion; the headline total should not be treated as equity alone. [A](https://labelbox.com/blog/labelbox-series-e/)
The thesis is that models can join unstructured imagery and records into an investigation rather than merely label a satellite image. Sheikh has worked with the failures of geographic data in the field. The strongest objection is ground truth: stale imagery and facts hidden inside a building can defeat a persuasive report. Watch for validation against inspected properties, not polished maps. [A](https://www.ycombinator.com/companies/ageospatial)
On-premises AI that turns sensor and operational data into predictive intelligence for utilities, built to run inside air-gapped environments.
Founders
John Armin (founder and CEO; PhD in environmental engineering and MS in computer science; adjunct professor at UNC Chapel Hill; co-founded drone-analytics firm Quantifly; previously worked utility infrastructure problems in consulting)
The deal
Techstars standard: $220K, as $20K for 5% common equity plus a $200K uncapped MFN SAFE. No company-specific variation disclosed.
The non-obvious claim is that utilities will not buy cloud AI at all: air-gap security practice rules out always-online services, so the product has to live inside the fence, which most AI vendors cannot or will not do. Armin's reasoning comes from consulting on exactly these systems. The case against is the mirror of the thesis: an on-premises model cannot phone home for updates, so it ages in place, and utility procurement cycles are long enough to starve a seed-stage company before the first deployment proves out.
Connects agent execution budgets to business outcomes and future allocation.
Founders
Hamze Al-Zamkan (TUM.ai president; appliedAI deployments; TUM and CDTM), Amgad Al-Zamkan (Amazon; TUM and UC Berkeley; adaptive decision-systems research), Muhtasham Oblokulov (former senior ML engineer at Munich Re; AI researcher)
The deal
YC standard: $500K, as $125K post-money for 7% plus $375K uncapped MFN. No company-specific variation disclosed.
The new opportunity is allocating computation to agents that can complete jobs, then learning from the economic result. The team identifies a specific missing link: token invoices measure activity while outcomes remain in other systems. Its operating and adaptive-systems experience informs that thesis. The strongest objection is attribution: a sale or completed workflow may depend on people and events the platform cannot observe. A feedback loop can optimize the wrong proxy faster. No independent outcome study was reviewed.
Conversational healthcare outreach that helps members complete appointments and other next steps.
Founders
Lauren Bilbo (former Google engineering manager; Stanford MBA; licensed health insurance broker), Baur Krykpayev (former Google and Meta senior software engineer; repeat founder)
The deal
Current Speedrun standard: up to $1M, with $500K for 10% through a SAFE and up to $500K in a qualifying next round of $2M or more within 18 months. The profile does not disclose this SR005 company’s actual terms.
The capability to test is a conversation that discovers why a person missed care and completes the next logistical action, rather than sending another reminder. Bilbo combines engineering with insurance-broker experience, which could help translate benefit rules into usable choices. The objection is that outreach and scheduling already existed: the AI must resolve varied barriers more reliably without steering people toward an insurer’s savings at their expense. The company’s savings claims are not an independent comparison.
Hangzhou home-robot startup; closed an angel-plus round of several hundred million yuan on October 13th, its fourth raise in about six months.
Four rounds in six months is a repricing schedule rather than a funding plan, and it measures investor appetite for Chinese home robotics more than product progress. Lead investors and terms were not fully disclosed; treat the velocity, not any implied valuation, as the signal.
AI drug-discovery company listed on the Nasdaq Global Select Market as IAM after pricing 9,375,000 shares at $16.
October 15th: Renaissance Capital reports about 9.4M shares priced at $16 for approximately $150M gross proceeds. Pricing is verified through that report; the start of trading is not. Do not equate gross proceeds with net cash. [B](https://www.renaissancecapital.com/IPO-Center/News/122239/small-molecule-biotech-iambic-therapeutics-prices-ipo-at-16-the-midpoint) Weekly synthesis: priced equity funds the next experiment; it does not validate the drug. Keep gross proceeds separate from net runway and require prospective clinical evidence before crediting the AI platform with therapeutic success. [B](https://www.renaissancecapital.com/IPO-Center/News/122239/small-molecule-biotech-iambic-therapeutics-prices-ipo-at-16-the-midpoint) November 1st monthly review: Iambic's own pricing release set trading to begin October 15th under IAM, and reports of the debut followed; status moves from priced to listed. A first-day close appeared only in an aggregator and is not recorded here. [A](https://www.biospace.com/press-releases/iambic-announces-pricing-of-initial-public-offering)
Helps enterprises run open-weight AI models, part of the nine-startup cohort Antler India presented in Bengaluru in July 2026.
Founders
Kunal Bhatia (three-time founder: SuperLearn in edtech, Switch in IoT), Vignesh Baskaran (deep learning researcher, top 1% on Kaggle, led ML research teams; met Bhatia at Antler's first India Residency)
The deal
Antler led the pre-seed; neither Antler's post nor the cohort coverage states the amount or terms.
The pick is the market position: Chinese open-weight models win global users while the serving layer captures the value, and an Indian team selling exactly that layer to enterprises sits on the right side of the trade. The case against is twofold: the company dates to 2022 as an image-generation API and has pivoted at least once, and at scale it competes with Together, Fireworks and every hyperscaler's model garden, which makes enterprise hand-holding the only durable edge. Founder backgrounds come from Antler's own posts.
Open-source background runtime that keeps coding agents' terminal sessions alive on each machine and queues them by which needs human attention; terminal UI, SSH and phone are just views.
Founders
Can Celik (solo founder, Ankara; previously built AI-assisted ultrasound technology and led AI teams in fintech)
The deal
YC standard: $500K, as $125K post-money for 7% plus $375K uncapped MFN.
The claim is that the agent harness, not the agent, is the missing layer: sessions that survive closed laptops, an attention queue across projects, and a socket API so other software can drive it. The traction is unusual for a solo founder, about 25,000 GitHub stars and 500 community plugins by the profile's count, though those are the company's own figures. The case against is structural: Anthropic, OpenAI and the terminal vendors all ship their own session persistence eventually, and an open-source runtime monetizes late if ever.
Learning-based robots that remove contaminants from organic waste.
Founders
Jonas Gruetter (robotics researcher; Gravis Robotics, Robotic Systems Lab and Ursa Mining), Stanislaw Piasecki (founding engineer at URSA Mining; ANYbotics; ETH Zurich computational science and engineering)
The deal
YC standard: $500K, as $125K post-money for 7% plus $375K uncapped MFN. No company-specific variation disclosed.
The capability to test is learning manipulation from failed and successful picks in crushed, tangled material, where a fixed recognition-and-grasp recipe breaks. The founders describe finding the problem on sorting lines and building autonomous excavators before applying that experience to waste. A custom three-way gripper is part of the proposed system. The objection is that automated sorting already exists; learning must improve contamination removal and recovery under dirty operating conditions enough to pay for maintenance. No independent shift-length reliability test was verified.
Reusable perception, simulation and runtime software for spacecraft autonomy.
Founders
John C. Lee (founder and CEO; former Marine Corps pilot and Vannevar Labs forward-deployed engineer)
The deal
Techstars standard: $220K, comprising $20K through a convertible equity agreement for 5% common equity and a $200K uncapped MFN SAFE. Company-specific terms not disclosed.
The hypothesis is that improving learned perception and flight-constrained computing let spacecraft teams reuse autonomy across missions instead of rebuilding it for each vehicle. Lee brings operational experience, and CTO Kishen Raghunath brings spacecraft and simulation work from Slingshot Aerospace. The strongest objection is that spacecraft autonomy predates the current model wave: the company must prove its reusable stack changes development cost or capability. The inspected pages provide no flight-validation results, so that technical discontinuity remains a thesis to test.
Custom data collection tied to failures in voice, robotics and world models.
Founders
Nikhil Reddy (UChicago math, economics and computer science; quantitative and software internships), Jeffrey Lin (NYU math, computer science and robotics; former NYU Robotics Team head of data)
The deal
YC standard: $500K, as $125K post-money for 7% plus $375K uncapped MFN. No company-specific variation disclosed.
The new requirement is data collection designed around physical-model failures, with video and actions linked to consent and collection history. The founders describe firsthand experience exposing annotation-platform quality checks that automated submissions could bypass. That informs the private collection system more directly than a resume does. The objection: provenance can establish where data came from without proving that it improves a robot. Model-transfer experiments, rather than the volume collected, must justify the service.
Coordinates medical-imaging referrals, authorization and documentation through to billing.
Founders
Frederick Wong (co-founder and CTO; formerly SafetyChain Software chief architect, Infor principal architect and SuccessFactors engineering director; UC Berkeley computer science PhD, per Company Launch Tracker)
The deal
Techstars standard: $220K, as $20K for 5% common equity plus a $200K uncapped MFN SAFE. No company-specific variation disclosed.
The signal to test is whether agents can complete changing handoffs across disconnected clinical and billing systems, rather than automate one fixed form. Aneko's specialty-specific view of readiness includes prior images, authorization and the evidence that a prerequisite was completed. That is a sharper hypothesis than a generic hospital assistant. The objection is integration: conventional workflow software already coordinates tasks, and the company has not supplied an independently verified comparison showing its AI completes more of them. Founder background is secondhand; the inspected company and cohort pages do not name the full founding team.
Deep reinforcement learning that tunes fan speeds, chilled-water flow and liquid cooling loops in data centers in real time, claiming up to 40% lower cooling energy cost.
Founders
Thomas Chepucavage (CEO), Wesley Billingslea (COO), Jimmy Chui (CTO)
The deal
Techstars standard: $220K, as $20K for 5% common stock plus a $200K uncapped MFN SAFE, per third-party summaries of the terms Techstars adopted in fall 2025. No company-specific variation disclosed.
Retrofit cooling optimization for the buildings the AI buildout is stressing: on-premise deployment in 4 to 6 weeks over existing HVAC, CRAC, liquid and immersion gear, with a claimed 8-to-14-month payback. The case against is strong and named: DeepMind demonstrated RL-driven data center cooling in 2016 and hyperscalers build this in-house, so the market is the long tail of colos and enterprise sites, where trust in an autonomous controller touching chilled-water loops is the real sale. The 40% figure is the company's own. Names are the leadership trio from the company site, which does not use the word founder; the Techstars class page names no founders.
Synthetic audiences for testing product ideas, designs and advertising against modeled customer behavior.
Founders
Alex Street (CEO; earlier employment not established in inspected sources), Yoann Lemain (product co-founder; earlier employment not established), Ludovic Lafole (technology co-founder; earlier employment not established)
The deal
Techstars standard: $220K, as $20K through a post-money convertible equity agreement for 5% common plus a $200K uncapped MFN SAFE. No company-specific variation verified.
The capability bet is an audience grounded in customer evidence with a validation process, beyond a model simply pretending to be a customer. The company publishes its method, giving buyers something concrete to challenge. The strongest case against is missing behavior: simulated people cannot reveal a preference absent from the data that made them. Independent comparisons with real interviews are essential. Founder roles are corroborated; prior employers were not verified.
Quantum computing startup; $475M Series B at a $5.4B valuation, up from $1.5B at its July Series A, about $775M raised in total.
Two rounds twelve weeks apart and a valuation up 3.6x is frontier-lab financing behavior arriving in quantum, the week Ethereum researchers began planning around AI-accelerated cryptanalysis. Revenue undisclosed, which at this price is itself a data point.
Checks that the expected person is still in control before consequential actions, using cognitive biometrics plus session context, and returns allow, challenge, hold or deny with signed proof.
Founders
Arav Mathur (Waterloo computer science; built AI systems for Stan), Aditya Ranjan (co-founder, COO)
The deal
YC standard: $500K, as $125K post-money for 7% plus $375K uncapped MFN.
Two-person Toronto team selling continuous proof-of-control: typing rhythm, cursor movement and response timing establish that a human, and the right human, sits behind a payment approval, a permissions change or an agent acting on someone's behalf. The timing is the pitch, since a week of unintended-agent-action disclosures made 'who authorized this' a billable question. The case against is that behavioral biometrics have a two-decade history of vendor claims outrunning field accuracy, and the signed-proof layer only matters if relying parties accept it, which is a standards fight a seed-stage company cannot force.
The newer capability is executing changes across systems while retaining review and rollback, rather than merely answering questions about configuration. The founders say customer conversations exposed safe execution as the bottleneck. The strongest objection is that a correct local change can still break a dependency in another system. Buyers need recovery tests across real integrations; a demo cannot establish that coverage. Weekly review, October 11th: evaluate the complete change history, including rejected attempts and recovery time. This week’s incident disclosures make an audit trail commercially relevant, but a vendor demonstration cannot establish that every external side effect is reversible.
Agents that propose machine-learning experiments, run them and use the results to pursue further discoveries.
Founders
Joshua V. Dillon (Google and DeepMind; foundational pretraining at Luma; TensorFlow Probability), Dan Kondratyuk (Google and Luma; VideoPoet and world models)
The deal
YC standard: $500K, comprising $125K post-money for 7% plus $375K on an uncapped MFN SAFE. No company-specific variation verified.
The recent opportunity is agents capable of running and revising experiments, rather than merely suggesting research. The founders have built models and research infrastructure, which may help them distinguish a promising failed experiment from noise. The strongest case against is that apparent discovery can be benchmark overfitting at substantial compute cost. Repeated gains on held-out problems, with failed runs included in the bill, would be stronger evidence than a leaderboard claim.
Machine-native model developer announcing $870 million in funding at a $7.5 billion valuation on October 9th.
Andreessen Horowitz leads, with Sequoia and DCVC participating; Martin Casado joins the board. Financing supports the attempt to serve structured enterprise decisions. Customer-level cost and reliability advantages over existing systems remain unverified.
Sequoia-backed developer of a general-purpose CPU core for data centers and AI workloads.
Reuters reported October 9th that it seeks hundreds of millions of dollars at roughly a $2.5 billion valuation. The round is not closed, terms can change and no product has shipped. The rumored status refers to this financing, not the existence of the company. [B](https://www.aol.com/articles/exclusive-sequoia-backed-chip-startup-223756000.html)
The founder ran inference at the lab that owns European AI sovereignty as a market position, which is the non-obvious knowledge: he has seen what European customers actually demand contractually, not just rhetorically. A solo founder renting GPUs against hyperscalers is the obvious case against, and the sharper one is that EU-resident inference is a feature every US provider can add with a region and a legal entity, so the moat has to be the open-model serving itself. Timely test: this week both Cloudflare and Perplexity shipped open models that need exactly this kind of hosting.
Techstars standard: $220K, as $20K for 5% common stock plus a $200K uncapped MFN SAFE, per third-party summaries of the terms Techstars adopted in fall 2025. No company-specific variation disclosed.
Paris company, founded early 2025, now in the ABN AMRO + Techstars Future of Finance cohort in Amsterdam with first clients reported in France and Qatar. The signal is the problem, not the category: insurers have bought fraud scoring from FRISS and Shift for a decade, but claims backed by generated photos and fabricated documents are a 2024-onward problem, and an agent that runs the investigation is a different product from a score. The strongest case against is exactly those incumbents, who see orders of magnitude more claims data and can bolt detection of generated evidence onto an installed base. Founder details come from an École Polytechnique profile; the Techstars class page names no founders.
Tactile hardware and software intended to give robots a scalable sense of touch.
Founders
Krithika Shamanna (co-founder of Ontologen; earlier background not disclosed in the profile), Kavan McEachern (former founding engineer at ZeroFox; distributed systems and machine learning)
The deal
Not disclosed in the South Park Commons company profile.
Touch sensors predate the current AI cycle. The reason to watch is whether their observations can now help general robot controllers handle contact and failure, rather than serve one custom installation. That is an investment question, not a capability established by the profile. The strongest objection is that hardware reliability and integration costs may dominate any model improvement. The company domain comes from SPC's link-out; its website fetch did not finish, and performance claims remain unverified.
Mortgage origination software using agents and automated workflows; announced a $30 million Series B with lender participation on October 8th.
Customers including Pennymac and New American Funding participated in the financing. The company says automation handles about 40% of platform tasks, a combined measure of agents and other workflows that does not establish autonomous loan-approval rates.
Superconducting power-delivery systems for AI data centers; announced a $110 million Series C on October 8th.
Funding addresses power delivery inside the AI infrastructure buildout. The announcement establishes the financing, but customer deployment volumes and independently measured operating costs remain unverified.
Warehouse robots as a service; $50 million Series A led by Framework Ventures on top of a $12 million seed, with Physical Intelligence supplying the policy stack.
A test of the layered thesis in physical AI: Physical Intelligence trains the general policy, Ultra wraps it in hardware, integration and a sales force, and charges an upfront fee plus monthly payments. The objection is the classic integrator risk: once the customer base exists, the brain supplier can reprice the brain.
Turns production agent runs into repeatable tests that preserve the tools, data and state behind the real run.
Founders
Aneesh Kethini (Datadog, built and launched its AI SRE; earlier a quantitative trader at SIG), Josh Le (engineering at Apple, Gusto, Wayfair and several YC companies)
The deal
YC standard: $500K, as $125K post-money for 7% plus $375K uncapped MFN.
Replaying a real agent run with its tool and data state intact is the hard, unsolved part of agent evaluation, and it only became a market once agents reached production. Kethini watched production agents fail from inside Datadog, which is the non-obvious knowledge here. The case against: observability and eval incumbents, his former employer first among them, can add replay as a feature, and YC's own F26 batch lists several neighbors, so differentiation is unproven.
Full-duplex speech-to-speech system for enterprises: the model listens and talks at once while following instructions and using tools.
Founders
Xutai Ma (Meta FAIR; core contributor to Llama-4 full-duplex speech and SeamlessStreaming, per the SPC profile), Anna Sun (Meta FAIR; SeamlessStreaming, SeamlessM4T and No Language Left Behind, earlier feed ranking at Meta)
The deal
Not stated. South Park Commons' profile lists no investment terms.
The founders worked on exactly this problem inside a frontier lab: Ma is named on Llama-4's full-duplex speech work, and full-duplex with real-time tool use is the part enterprise voice agents still lack. The case against is structural rather than technical: the frontier labs ship native speech-to-speech APIs with distribution attached, the company site names no customer or product yet, and research pedigree is not a moat. Kyutai's Moshi showed the core idea in 2024; the claim here is production-grade tool use on top.
The new element is generating control screens and logic while interpreting plant data, not the older business of monitoring machinery. The team’s customer interviews led it to sell integration alongside software and work over existing installations. That addresses adoption friction directly. The strongest objection is that generated control logic still needs industrial validation, while integration labor can consume the promised software margin. Deployment and speed claims remain the company’s own.
Model access for professional adversarial testing where mainstream model refusals obstruct the test.
Founders
Devon Thomas (co-founder and CEO), Jide Ogun (co-founder and CTO), The profile attributes prior inference and GPU-platform work at Apple and Microsoft to the founding team without assigning each employer individually.
The deal
Published Speedrun standard: up to $1M, with $500K for 10% through a SAFE and another $500K in the next round within 18 months. Company-specific terms are undisclosed.
More capable agents create demand for tests that reproduce an attacker’s behavior instead of stopping at a model refusal. The founding team’s inference-platform experience addresses serving those models, though the profile does not prove an exclusive technical advantage. The strongest case against is dual use: the same access can help real attackers, and removing refusals does not establish better security testing. Customer controls and independently measured defensive value remain unverified.
Durable state and coordination infrastructure for agent applications.
Founders
Thomas Elias Karatzas (CEO; previously staff engineer at Apple on Siri and Apple Intelligence), Olivier Simard-Morissette (CTO; previously principal engineer at Pigeon and senior engineer at Squarespace and The New York Times)
The deal
YC standard: $500K, comprising $125K post-money for 7% plus $375K uncapped MFN. Company-specific executed terms are not disclosed on the profile.
The founders have worked on assistant and production software systems. Persistent state is old; the new bet is that long-running, collaborating agents create enough coordination demand to support a dedicated vendor. The case against is strong: existing databases and actor runtimes already solve much of the underlying problem. Adoption must come from making agent recovery easier, not merely renaming durable storage. The profile and linked product establish the offering, not production reliability.
On-device vision models that learn screen workflows and turn them into reusable agent skills.
Founders
Riyan Moussaoui (CEO; physician, FutureClinic co-founder and former Anterior AI engineer), Ajit Krishna (CTO; founding engineering team at Anterior and former senior engineer at Clarion)
The deal
Published Speedrun standard: up to $1M, with $500K for 10% through a SAFE and another $500K in the next round within 18 months. Company-specific terms are undisclosed.
The founders met deploying healthcare agents, where learning a workflow and handling sensitive context are joined problems. Smaller vision models make local observation and adaptation more plausible now. The strongest objection is that watching a screen does not establish permission to act or reveal the hidden reasons behind a decision. Shared skills could propagate those mistakes across a team. Speed claims are the company’s own; the company website link is verified on the accelerator profile, but its destination could not be fetched.
Robot demonstration capture and deployment company; $60 million Series B led by Sequoia.
Nvidia, M12, Qualcomm Ventures and Samsung joined. Combining capture hardware with customer deployments could test data quality against real work; independent outcomes were not supplied.
Parent of the Manus agent; closed more than $500 million, its first round since China's NDRC forced Meta to unwind its $2 billion acquisition in April.
Boyu Capital and IDG Capital co-led; Tencent, HSG and ZhenFund participated. Valuation undisclosed, with September reporting near $4 billion, double Meta's price. The question the round answers is whether a forcibly repatriated AI company can still raise at a premium: yes, from Chinese investors.
AI body-camera software for live guidance, shift reports and evidence workflows.
Founders
Paul Berberian (CEO; previously Sphero leadership and Company Six co-founder/chairman), Skylar Graika (co-founder and CTO; earlier background not established in reviewed sources)
The deal
Techstars standard: $220K, comprising $20K through a post-money CEA for 5% common equity plus $200K on an uncapped MFN SAFE. Executed company-specific terms are not public.
The capability bet is continuous interpretation of body-camera footage, turning events into guidance and draft reports while a shift is still running. Berberian’s robotics experience is relevant to deployment, but does not prove the generated account is reliable. The strongest objection is evidence integrity: a plausible invented detail could contaminate a report. Worker surveillance and verification costs could outweigh time saved. No independent field evaluation was established. [A](https://patrol6.com/about/)
Open-source Hermes agent maker, now selling a managed enterprise version, Hermes for Businesses.
The deal
$90M Series B at a $1.5B valuation led by Robot Ventures, with Nvidia, Union Square Ventures, Menlo Ventures, Samsung and 1789 Capital. $158M raised in total.
The financing backs enterprise deployment around an existing open agent whose users can select models and hosting. About $36M annualized revenue in mid-September per the Wall Street Journal, with a stated aim of $100M by year-end. The 24 million clones and 2.5% of global tokens are company estimates. Access controls and institutional knowledge handling remain the adoption test.
Lambda plans a 2027 IPO, TechCrunch reports citing the Wall Street Journal.
No public filing established. Keep listing plans separate from completed financing. [B](https://techcrunch.com/2026/10/06/ai-computing-startup-lambda-to-raise-4b-ahead-of-planned-ipo/) November 15th weekly: Anthropic's November 10th amended S-1 says Lambda did not secure a lease for the facility behind their August agreement by the agreed date, so that $34.9 billion contract is now terminable by either party; neither has given notice and Anthropic says it does not intend to terminate. The three Lambda agreements total about $41.3 billion into 2034, with fees accruing only on delivered and accepted capacity. Bears directly on any listing case built on that backlog. Still rumored. [C](https://www.theinformation.com/briefings/anthropic-discloses-10-billion-new-deals-microsoft-google-says-lambda-deal-may-cancelled)
Distills large video models into small models specialized for a customer’s task.
Founders
Shubham Sharma (robot-learning and agent-memory research at KAIST; offline reinforcement-learning research at Tsinghua; aerospace engineering at IIT Bombay)
The deal
YC standard: $500K, comprising $125K post-money for 7% plus $375K on an uncapped MFN SAFE. Company-specific variations are not disclosed.
The newer opportunity is using a capable video model as the teacher for a smaller task-specific model that can run where sensitive recordings already live. That changes the cost of building a specialist compared with collecting all its labels by hand. The objection is teacher error: specialization can reproduce systematic mistakes while making them harder to detect. The profile’s quality comparisons are company claims, with no independent replication established. [A](https://www.ycombinator.com/companies/kailash-labs)
Nightly AI screening of every adult hospital inpatient's record for undiagnosed conditions such as malnutrition and delirium, with a risk score delivered to the care team each morning.
Founders
Josiah Meyer, CEO, Jemima Meyer, co-founder (built the original clinical nutrition tool for dietitians)
The deal
$38M across a $8M seed (Sequoia and First Round co-led) and a $30M Series A led by Hummingbird Ventures. Valuation undisclosed.
Three-year contracts priced by licensed bed count, with part of the fee tied to ROI that hospital finance teams validate. Fifty-plus hospitals from three a year ago is fast for health-system sales. The $23.8M Penn Medicine figure is the company's own and includes reimbursement it helped book, so the number to watch is renewals.
A data-first neurotech company: records brain data for other organizations and trains general models of the brain on it.
Founders
Julian Weaver, founder (founding engineer at Syntra, YC S24; led EEG and fMRI AI research; dual degrees in computer science and neuroscience, UT Austin, completed by age 20)
The deal
YC standard: $500K, as $125K post-money for 7% plus $375K uncapped MFN.
The structure is the story: data collection as a service pays for the dataset, and the dataset is meant to pay for a general model of the brain, which every neurotech lab wants and none can afford to build alone. Its Wizard model is pitched as picking up a new task from a few examples. Weaver led EEG and fMRI AI research and was a founding engineer at Syntra. The case against sits in the contracts: organizations that pay to have brain data collected rarely grant the right to pool it, and without pooling there is no general model. Brain recordings also do not accumulate like text; the public corpus for the whole field is a rounding error against what modern pretraining consumes.
A marketplace where businesses fund tasks up front and third-party AI agents compete to complete them, with escrow released after verification.
Founders
Aashna Doshi, CEO (software engineer at Google in AI and security; master's in machine learning, Georgia Tech), Shivam Patel, CTO (algorithmic-trading infrastructure at Barclays; electrical and computer engineering, Rutgers)
The deal
Speedrun standard: up to $1 million, with $500K for 10% upfront through a SAFE and another $500K in the next round within 18 months, per the program FAQ. Company-specific variations are not disclosed.
The inversion worth watching is that agents are the supply side: over 200 third-party agents onboarded, with escrow and automated verification standing in for a manager. That was not a buildable product two years ago, when agents could not finish work unattended. Doshi wrote AI and security software at Google and Patel built algorithmic-trading infrastructure at Barclays, roughly the right background for an exchange whose core problem is adversarial quality control. The case against is that verification is the whole product: automated checks on lead lists and videos are exactly what a confidently wrong agent games, the quality numbers are the company's own, and a buyer who finds a good agent has every reason to go direct.
AMD Ventures participated. The company says its platform is running on production engineering programs; speed and accuracy claims have not been independently checked. Expansion beyond semiconductors must preserve physical accuracy as the range of design problems grows.
An analytics engine for physical AI: ingests robot logs across formats, runs agents for automated failure triage, and curates fleet data into training sets.
Founders
Benji Barash, CEO (Amazon drone and robotics engineering), Yves Albers-Schoenberg, CTO (Amazon Robotics)
The deal
$4.8M seed led by Unusual Ventures, with the Allen Institute for AI and FUSE, per GeekWire's coverage of the 2023 launch.
Spun out of the Allen Institute in 2022, so the company is not new; the current product is, and it sits where this week's money is pointing. Every world-model and robot-fleet round assumes someone can turn messy telemetry into training data, and Roboto already sells that step to BRINC, ANYbotics and surgical-robot teams. The case against: the labs it serves are building their own data tooling, and infrastructure between a fleet and its model is exactly what a platform swallows first.
Cloud labs are not new; Emerald Cloud Lab spent a decade proving robotic wet labs are brutally capital-intensive with humans as customers. The new claim is the customer: agents that design experiments faster than any lab can run them, which makes throughput the product. The case against is that three 19-year-olds are entering the market that exhausted far better-funded teams, and agent demand for physical experiments is still a forecast, not a purchase order.
Kuaishou's AI video unit picked banks for a Hong Kong IPO targeting a raise above $1B, per Bloomberg's sources, after reaching a roughly $15B valuation on a $2.8B July raise.
Bank mandates reported by a named outlet citing sources; no filing yet. A video-model subsidiary listing separately from its parent prices generative video on its own.
Business agents that turn repeatable knowledge work into reusable workflows.
Founders
AJ Asver (AI/ML products at Coinbase and Brex; founder of Scoopler, acquired by Google), Miguel (former platform engineering lead at Brex and data science lead at Twitter; surname not disclosed in the reviewed profile)
The deal
YC standard: $500,000, comprising $125,000 for 7% plus $375,000 on an uncapped MFN SAFE. Company-specific variations are not disclosed. [A](https://www.ycombinator.com/deal)
The newer bet is compiling useful agent behavior into repeatable code, so learning a workflow can make subsequent runs cheaper. Asver’s account of a mailbox mistaken for a physical address at Brex explains the founders’ interest in verification. The objection: compiled workflows can preserve mistaken assumptions, and rare exceptions may consume the savings. [A](https://www.ycombinator.com/companies/grep-ai)
Singapore-based operator spun out of China's GDS filed for a US IPO; first-half revenue more than tripled year over year to $512M while the net loss widened to $77.2M.
The AI data-center trade in one filing: compounding revenue on capacity that loses money while it builds. [B] October 7th update: Renaissance Capital estimates a $3.0 billion raise on Nasdaq under DODC, with Morgan Stanley, J.P. Morgan, BofA Securities, Citi and BNP Paribas running the books; it reports $845 million in revenue for the twelve months ended June 30th and 4.6GW of capacity, 2.3GW of it contracted to seven hyperscale and technology customers. An estimated deal size is Renaissance's model, not a terms page. [B](https://www.renaissancecapital.com/IPO-Center/News/122062/global-data-center-developer-dayone-data-centers-files-for-an-estimated) October 21st: Bloomberg reports discussion of a delay or a $2.5B–$3B raise against an earlier $5B ambition. No final decision reported. Status remains filed; this is not a withdrawal or priced deal. [B](https://news.bloomberglaw.com/business-and-practice/dayone-is-said-to-weigh-options-for-us-ipo-amid-data-center-rout) November 1st monthly review: Bloomberg reported October 21st that DayOne was weighing a smaller raise or a delay; no price range or pricing was found through October. Filed, not priced. [B](https://news.bloomberglaw.com/business-and-practice/dayone-is-said-to-weigh-options-for-us-ipo-amid-data-center-rout)
A coordination agent that follows context across workplace tools and teams.
Founders
Thomas Filshill (Georgia Tech; previously built and scaled AI systems), Livia Kuga (Stone founding team; spent a decade coordinating teams as the company grew)
The deal
Speedrun standard: up to $1 million, with $500,000 for 10% upfront through a SAFE and another $500,000 in the next round within 18 months. Company-specific terms are undisclosed. [A](https://speedrun.a16z.com/faq)
Alia’s bet is that agents can retain enough cross-tool context to detect a blocker and coordinate its resolution, extending beyond meeting summaries. Kuga’s experience owning coordination at Stone gives the team a concrete problem to model. The strongest objection is organizational: correct context does not grant authority, and staff can ignore or contest an automated intervention. Independent evidence of durable resolution is still missing. [A](https://speedrun.a16z.com/companies/alia)
Voice-first hardware for talking to AI agents: a smart ring that records, summarizes and controls agents, and earphones with proprietary noise isolation for dictation and calls from loud places.
Founders
Sridipto Ghosh, CEO (co-founded Snipe, $250K revenue in 15 months; IIT BHU), Sirsho C, CTO (co-founded Snipe; computer science, KIIT)
The deal
YC standard: $500K, as $125K post-money for 7% plus $375K uncapped MFN.
The bet is that agents make voice the primary interface and that phones are the wrong microphone for it, so the input hardware is the wedge. That was not a buildable thesis two years ago; agents could not hold a task long enough to be worth talking to. The case against is the consumer AI hardware graveyard: Humane and Rabbit had nine figures and shipped products nobody kept, and a two-person team is taking on hardware logistics, not just software.
A behavioral intelligence layer that simulates realistic human decision-making, so companies can test how people will respond to software and agents before launching them.
Founders
Rohit Talluri, CEO (Amazon AGI Autonomy Lab and AWS foundation-model teams; University of Washington), Jean Farmer, CTO (built agentic capabilities at Amazon AGI; Colorado School of Mines), Gabriel Fong, COO (AWS and DoiT; Northeastern)
The deal
a16z Speedrun SR007 standard: $1 million per company, structured as a $500K post-money SAFE plus further investment for 10% total equity, per program materials.
Two of three founders come from Amazon's AGI Autonomy Lab, which is a specific claim about knowing where agent behavior modeling actually breaks. The product only became possible once models could simulate coherent humans, so it passes the two-years-ago test. The case against is the one that haunts every synthetic-audience company: nobody has shown simulated users predict real ones out of distribution, and the team's NeurIPS and AAMAS papers are submissions, not acceptances. Demo day is October 6th; watch whether a design partner is named.
Modaic tries to decide which model outputs need a person, then uses that feedback to improve instructions. Its founder describes moving from optimization to measurement because buyers lacked useful metrics. That is a specific objection to selling better prompts before defining a good result. The case against: a confidently wrong classifier can remove precisely the cases a reviewer needs to see, and calibration can change with incoming data. Independent evidence of reduced review work without missed errors remains absent. [A](https://speedrun.a16z.com/companies/modaic)[A](https://speedrun.a16z.com/companies/modaic/tyrin-ian-todd)
Neural-network arithmetic and chip architecture designed to remove multiplication from inference.
Founders
Sarang Zambare (Peloton Guide ML lead; founding ML engineer at Peloton and Caper), Eric Taylor (chip architecture at Nvidia, Freescale/NXP, Arteris IP and Enfabrica)
The deal
YC standard: $500K, as $125K post-money for 7% plus $375K uncapped MFN.
Baud changes the arithmetic a model is trained to use, then designs hardware around it. The profile describes FPGA emulation and plans for a silicon tapeout by the end of 2026, not a manufactured chip already serving customers. Its distinction is the joint training-and-hardware bet, rather than another scheduler for existing GPUs. The case against: retraining requirements and software compatibility could consume the theoretical savings before a customer gets useful output. Real-silicon performance remains to be established. [A](https://www.ycombinator.com/companies/baud)
Flex cloud and power infrastructure business preparing for a public-company separation.
Flex announced a $2 billion convertible preferred investment agreement at an initial $37.5 billion enterprise value. Separation is planned for the first quarter of 2027. The Form 10 filing and financing agreement do not establish a completed listing or ordinary-share valuation. [A](https://www.sec.gov/Archives/edgar/data/866374/000119312526413173/d123485dex991.htm)[A](https://www.sec.gov/Archives/edgar/data/866374/000086637426000036/axflexx91526_flexxspincoxb.htm)
Blank-check company targeting an acquisition in the AI industry.
Renaissance Capital reports a $200 million IPO and trading during the week of September 28th. This is an acquisition vehicle, not an operating AI business or an identified merger target. [B](https://www.renaissancecapital.com/IPO-Center/News/122042/us-ipo-weekly-recap-oura-postpones-2.1-billion-ipo-and-accelevation-trades)
AI models and chemistry workflows for industrial compounds and multi-ingredient formulations.
Founders
Ansh Tiwari, CEO (prior affiliation not stated in the profile), Ayush Chauhan, CTO (Speechify; engineering at Freya and Orchids), Daood Hashmi (research in the Arnold and Marcus Lab at Caltech)
The deal
YC standard: $500K, comprising $125K post-money for 7% and $375K on an uncapped MFN SAFE. No company-specific variation disclosed.
The recent capability bet is coordinating chemistry tools while checking intermediate results, then using laboratory measurements to improve formulation models. Hashmi’s laboratory work is relevant to the gap between a plausible calculation and a usable material. The strongest objection is experimental validation and manufacturing scale: faster proposals do not establish stable, safe chemistry. The company site describes provisional patent applications, so the profile’s broader patent language should not be read as granted patents. [A](https://www.ycombinator.com/companies/rasyn)[A](https://www.rasyn.ai/)[A](https://www.ycombinator.com/deal)
Smart-ring company with a filed offering whose launch was postponed.
Renaissance Capital reports the planned $2.1 billion IPO was postponed amid market uncertainty. The filed range was $40 to $44. Filed status is retained; neither pricing nor listing occurred. [B](https://www.renaissancecapital.com/IPO-Center/News/122042/us-ipo-weekly-recap-oura-postpones-2.1-billion-ipo-and-accelevation-trades) November 10th: TechCrunch reports the IPO of up to $2.2 billion is shelved indefinitely, citing IPO-market uncertainty; Oura instead agreed to buy Counsel Health for $625 million and reported third-quarter revenue of $533 million, up 153%.
A duplex speech model that listens while speaking, offered for outbound sales calls.
Founders
Sid Sharma, CEO (early engineer at Wayve, building its data and compute platform), Vatsal Aggarwal, CTO (speech research and deep-learning voice development for Amazon Alexa)
The deal
The EF profile does not disclose investment amount, instrument or equity terms.
The company was founded in 2022; the discovery is its current simultaneous listening-and-speaking product, not a new incorporation. Its thesis is that training directly on dialogue can handle interruptions that a sequence of separate transcription and speech tools struggles with. The founders’ data-platform and speech work bears on that problem. The case against: handling overlap must improve real conversations enough to justify a dedicated model, and a persuasive voice can still make incorrect commitments. Conversion and latency claims remain vendor claims. [A](https://www.joinef.com/companies/familiar/?pagenum=2)[A](https://familiar.io/)
An AI coding agent specialized for C and C++ that builds, debugs and tests every change with the real toolchain until the change holds; terminal, VS Code, Neovim and Emacs, with on-premises deployment.
Founders
Anirudha Kulkarni (low-latency C++ at Optiver and Quantbox; IIT Delhi; co-founded LawSutraAI, acquired by Manupatra), Pratyush Saini (ThirdAI core engineer; founding engineer at Simbian; IIT Delhi; LawSutraAI CTO)
The deal
YC standard: $500K, as $125K post-money for 7% plus $375K uncapped MFN.
General coding agents are weakest exactly where C and C++ live: long compiles, undefined behavior, memory bugs the test suite never catches. Two founders who wrote trading-floor C++ are claiming the fix is toolchain-in-the-loop verification, not a better model, and on-prem deployment targets the banks and defense shops that cannot ship code to an API. The case against: the C++ niche may be too small to defend once Cursor and Claude Code get compiler feedback loops of their own, and verticalized coding agents have no moat if the base model does the verification natively. Weekly context: the toolchain feedback loop is the proposed product advantage. Independently measured reductions in debugging and review time are still missing. [A](https://www.ycombinator.com/companies/byteask)
Blank-check company targeting agentic AI, stablecoin payments and digital-yield businesses.
Renaissance Capital reports a filing to raise $100 million. A target sector is not a signed acquisition; the record does not imply pricing or listing. [B](https://www.renaissancecapital.com/IPO-Center/News/122042/us-ipo-weekly-recap-oura-postpones-2.1-billion-ipo-and-accelevation-trades)
Barrett Lyon’s private-network platform for people and AI agents.
Announced a $38 million Series A on October 1st, led by Andreessen Horowitz with Animo Ventures and Focal.vc. Open beta supplies a product to examine; the announcement does not independently establish its security claims. [A]
Robotics-chip designer from Zhuhai; listed in Hong Kong on November 12th at HK$9.90.
SCMP, citing sources: the cleaning-robot chip designer, whose customers include Xiaomi and Midea, cleared its HKEX listing hearing and plans bookbuilding within days for a mid-October debut. A cleared hearing is a gate passed, not a priced deal, so this stays rumored until terms print. [B] The HKEX prospectus, published about October 30th, gives stock code 2065, 71,687,000 H shares at HK$9.27 to HK$11.58, pricing by November 10th and listing expected November 12th. Filed until the price prints. [A] November 12th: priced at HK$9.90, selling 71.69 million H shares for HK$709.7 million, and closed at HK$13.68, up 38.2%, after an intraday high of HK$18.11, while the Hang Seng Tech Index fell about 1.2%. Retail was 13.95 times subscribed and the international tranche 1.05 times.
AI chip-design verification: formal proofs find bugs in a design and AI localizes and fixes them in a verify-and-fix loop.
Founders
SukHwan Lim, CEO (former EVP at Samsung heading US R&D for System LSI semiconductors; AI acceleration at Google and Apple), Subhasish Mitra (Stanford professor in robust and dependable computing and 3D nanotechnologies), Shwetabh Verma (circuit and systems design at Broadcom, Apple and Cadence)
The deal
Terms not disclosed on the SPC profile; the seed was co-led by South Park Commons and Khosla Ventures, per SPC's September 2026 update.
Everyone else is building AI that writes chips; this team built AI that proves chips wrong and then repairs them, and the fix half only became possible with current code models. The profile claims their platform found critical bugs in a top-ten semiconductor company's production design in under two weeks. The case against: Synopsys and Cadence own the verification workflow, chipmakers almost never rip out qualified toolchains, and formal methods have historically failed to scale past block level, which is exactly what the recursive-abstraction claim must beat.
The newer capability is interpreting workers' photos and voice messages while connecting them to shop documents and work orders. So's factory experience identifies an adoption barrier: many workers have no company login. The strongest objection is authorization. A convenient text interface still needs to establish who may order parts or change work. The profile does not independently demonstrate safe execution or its claimed productivity gain. [A](https://www.ycombinator.com/companies/textsidekick)
An AI floor manager for full-service restaurants that runs on the CCTV they already have.
Founders
Cameron Kuperman, CEO (restaurant industry experience; spintronics research; math, ECE and CS at Duke), Harshith Guduru, CTO (SWE intern at Walmart Global Tech and Bastazo; published RL cybersecurity and graph cancer research; CS and EE at Duke), Alex Tabaku, CFO (four years in restaurants; VLA model research; mechanical engineering and math at Clemson)
The deal
YC standard: $500K, as $125K post-money for 7% plus $375K uncapped MFN.
The insight is that the state of a restaurant floor never enters software: the POS only sees what staff type in afterward, while the security cameras the restaurant was already forced to install watch everything. Reading a chaotic dining room in real time at a price a restaurant can bear only became feasible with current video models. The case against: restaurants are the graveyard of vertical software, a POS incumbent could bolt cameras onto an existing billing relationship, and worker-surveillance optics plus biometric-privacy law are one viral clip from becoming the story.
South Korean AI fire-safety and disaster-prevention software maker, now trading on Nasdaq.
Completed a Nasdaq direct listing at a $933 million market cap and was up 149% by October 2nd, per Renaissance Capital's weekly recap. IoT sensors plus digital-twin software for fire prevention is about as far from a model lab as the AI pipeline gets, which is itself a sign of how wide the listing window has opened. Exact listing day within the week of September 28th. [B] October 4th clarification: this Korean issuer is unrelated to the SoftBank-backed company also named Roze. [B](https://www.renaissancecapital.com/IPO-Center/News/122042/us-ipo-weekly-recap-oura-postpones-2.1-billion-ipo-and-accelevation-trades)
An AI workspace for financial compliance at regulated firms.
Founders
Aiden Dawes (Companies Graph; Oxford ML research; AI agents at EY), Damian Stone (Xpendit founding AI engineer; evaluations and agents at Addepar), Lee Forsyth (financial crime transformation at Merrill Lynch, Goldman Sachs, Thomson Reuters and Exiger)
The deal
The EF profile does not disclose the investment amount, instrument or equity terms.
Compliance software predates this team. The newer thesis is that document-reading agents can execute the messy checks between systems, with the founders bringing both agent-building and financial-crime implementation experience. The strongest objection is auditability: a plausible answer is insufficient when a reviewer must reproduce why a customer cleared a check. Pilot conversion is still a company expectation, so it is not booked revenue. [A](https://www.joinef.com/companies/eliot-ai/?pagenum=2)
Data-center power-infrastructure manufacturer riding the AI build-out.
Priced a $540 million IPO at a $4.1 billion market cap and traded down about 2% by October 2nd, per Renaissance Capital. AI-adjacent rather than AI: it sells the electrical gear the build-out consumes, with reported 147% revenue growth in 2025. The flat debut against Roze's 149% pop says the market is pricing picks and shovels more soberly than stories. [B] October 4th status review: changed to listed because the same source confirms trading had begun. [B](https://www.renaissancecapital.com/IPO-Center/News/122042/us-ipo-weekly-recap-oura-postpones-2.1-billion-ipo-and-accelevation-trades)
An open-source Postgres platform that is acquiring Turso to give AI agents per-agent databases.
GIC led $150M, with CapitalG, IronArc and SquarePeg, four months after a $500M Series F. Supabase says it adds 4 million databases a month and that 70% are created by agents or AI tools. Company announcement; growth figures not independently verified. [A](https://www.prnewswire.com/news-releases/supabase-announces-150m-in-new-funding-and-turso-acquisition-302896752.html)
An AI computing platform that aggregates and manages infrastructure capacity.
The company announced a $200M Series C at a $3.2B valuation, led by ComputeCore with B Capital participating. Its claimed $5B of contracts should not be read as recognized revenue. The investment case depends on turning those commitments into profitable delivered capacity without taking on financing and utilization risks that overwhelm the management fee. Company announcement; commercial claims not independently verified. [A](https://www.prnewswire.com/news-releases/palebluedot-ai-raises-200m-series-c-round-to-scale-super-intelligence-infrastructure-platform-302896601.html)
The founders apply a network-security lesson to agents: the untrusted client should not possess the real credential. Secrets are substituted by a gateway after a request is authorized. General-purpose agents now make that boundary commercially useful across many applications. The strongest objection is that hiding a token does not prevent an authorized agent from misusing the resulting access. Endpoint policy and approval design still determine what damage it can do; the profile is not an independent security audit.
A Chinese AI computing startup reportedly preparing a Hong Kong listing.
Bloomberg reports a confidential submission and a possible 2027 listing, citing people familiar with the matter. No public filing was reviewed, so this record remains in the reported pipeline rather than treating an anonymous-source report as a verified filing. The article’s automated summary gives a conflicting year; its dated reporting says next year. [B](https://www.moneycontrol.com/news/business/companies/china-s-infinigence-ai-plans-hong-kong-ipo-to-tap-surging-ai-computing-demand-14042855.html)
A Medal spinout training action models on gameplay footage for robotics, simulation and entertainment.
The $220M round at a $6.2B valuation came three months after a $320M Series A at $2.3B, from Valor, Atreides, Seven Seven Six, Point72, Khosla and General Catalyst. Its MIRA model is a research demonstration for a single game; a commercial version is with select customers. The valuation rests on the premise that gameplay data transfers to physical robots, which is not yet shown. [B](https://siliconangle.com/2026/09/29/world-model-startup-general-intuition-closes-220m-investment/)
Robotic automation for physical work inside data centers.
Founders
Natalia Kokoromyti (Stanford Scaling Intelligence Lab and Lucid Computing), Luc Chartier (confidential computing at Lucid Computing)
The deal
Speedrun standard: up to $1M, with $500K for 10% via SAFE upfront and another $500K in the next round within 18 months. Company-specific terms not disclosed.
The founders are taking third-party humanoids and mobile manipulators into racks, cables and maintenance. Data centers offer more structured surroundings than homes, and their experience with confidential computing gives them a reason to notice the physical work behind a supposedly virtual service. The objection is uptime: a dropped connector or clumsy repair can cost far more than labor saved. The profile does not establish a reliability record that resolves that concern.
An agentic video editor on iPhone: prompts, photos and clips become short-form videos, with the agent selecting clips, writing script, narrating and captioning, revised by chat or timeline.
Founders
Kate Deyneka, founder (senior machine-learning engineer at Snapchat on video generation, applied research and product)
The deal
Speedrun standard: up to $1M, as $500K for 10% via SAFE plus follow-on reserved to $1M; Reelful's specific terms are not disclosed.
Priced at $24.99 a month for about ten videos, which makes the inference-cost bet legible: the margin depends on video generation getting cheaper faster than creators' expectations rise. The founder built video generation inside Snapchat, which is a claim about knowing where automated editing actually breaks for short-form distribution. The case against: the category is the most crowded in consumer AI, HeyGen cut prices to $0.01 a second this week and Pexo launched the same pitch for free, and a solo consumer app wins on taste and retention, neither of which is visible from outside yet.
An always-on AI security engineer: codebase scanning, penetration testing and supply-chain analysis, with findings verified and false positives filtered before a human sees them.
Founders
Ananay Arora, founder and CEO (solo founder; the YC profile lists no prior employer)
The deal
YC standard: $500K, as $125K post-money for 7% plus $375K uncapped MFN. Company-specific closing documents were not reviewed.
The timing is the thesis: the week CAISI named a free open-weight model the most cyber-capable yet released, continuous AI-driven defense stopped being optional for teams that cannot hire a security engineer. The verification step, confirming an exploit actually fires before reporting it, is what separates this from a scanner. The case against: a team of one selling always-on security into enterprises that demand vendor diligence, in a field where Alice and Nebula Security are already funded, is a trust problem before it is a product problem.
Voice AI company: completed a $300M employee tender at a $22 billion valuation, led by Wellington Management and T. Rowe Price, double February's $11 billion Series D mark.
First-time investors EQT, Goldman Sachs, GIC, OTPP, Sapphire Ventures and BDT & MSD bought in alongside existing holders. The company says agents handle about 15 million conversations weekly and enterprise customers supply 55% of revenue. A tender prices existing shares rather than raising capital, so the $22 billion is what crossover funds paid employees for stock: a cleaner demand signal than a negotiated primary round, and a weaker one than audited revenue.
Turns food-distributor orders from messages and photos into ERP entries.
Founders
Yudai Matsuoka, CEO (Mitsubishi tuna distribution, yellowtail farming and McKinsey; profile does not name the other builder)
The deal
YC standard: $500,000, as $125,000 for 7% on a post-money SAFE plus $375,000 on an uncapped MFN SAFE. Program terms, not separately verified company financing.
Order-entry automation is old. The newer bet is interpreting mixed photos and informal messages while sending ambiguity to a person before writing an order. Matsuoka has worked inside seafood distribution, where substitutions and product names complicate generic extraction. The case against: customer-specific catalogs and ERP integration can turn each installation into services work. Reliable exception handling must carry the advantage over older document software.
AI-native offensive security company founded by Mandiant's Kevin Mandia: raised a $255.5M Series B at a valuation above $2.5B, co-led by Andreessen Horowitz and Accel.
Total funding is $445M. Its headline evidence is one August exercise with 26,000 agents and 38 verified attack paths, reported by the company itself. Revenue and customers were not disclosed.
A phone-based assistant that coordinates everyday requests for older adults and their families.
Founders
Andrew Baran, co-founder and CEO (Leilo co-founder; McKinsey software and AI work; Radian Capital healthcare and software investing)
The deal
Published Speedrun standard: $500,000 for 10% upfront through a SAFE, plus $500,000 in the next round within 18 months. Advocate-specific terms are not disclosed on its profile.
A human concierge could do this years ago. The recent-model thesis is economical voice interaction that completes varied errands and keeps one shared family record. Baran grew up with a care-coordinator parent, a more relevant clue to the coordination problem than the consulting credential. The case against: missed appointments or misunderstood requests can erase the labor savings, and family trust requires dependable escalation to a person. The profile supplies no audited completion or failure rates.
AI workflows for infrastructure contractors, manufacturers and their supply chains.
Founders
Finn Lancaster, CEO (previously a senior technical employee at Harper; founded FinnedAI and worked on search and machine learning at Publix), Brison Moorhead, COO (founding account executive at Bonsai, later acquired by Zoom; early Harper sales and previously HubSpot)
The deal
YC standard: $500,000, comprising $125,000 for 7% plus $375,000 on an uncapped MFN SAFE. These are program terms, not a separately verified company financing.
The proposed advance over older CRM integrations is reading specifications and bids, then carrying their requirements into approved system updates. The site describes a read-only start and human approval for writes. The founders have seen enterprise sales and operational handoffs from inside earlier companies. The case against: integrations and procurement rules can consume the margin, and established construction software already owns the records. The recent-model thesis needs to prove that messy documents can become dependable actions, not just summaries.
Customer context assembled from calls, records, product activity and support for agents and internal applications.
Founders
Henry Cook, CEO (Datadog alumnus with a customer-facing background; the cohort profile does not name the other co-founder)
The deal
Techstars standard: $220,000, comprising $20,000 for 5% in common stock plus a $200,000 uncapped MFN SAFE. These are program terms, not a separately verified company financing.
Customer dashboards predate current AI. The newer bet is turning calls and tickets into a persistent customer model that agents can query and use to build internal applications on demand. A Datadog background supplies experience with how technical customer information reaches operating teams. The case against: a dashboard with generated summaries would not establish that advantage. BuiltBlock must preserve permissions and conflicting evidence across systems while competing with the vendors that already store those records.
Cyber-intelligence business pursuing a public listing through Bold Eagle.
September 28th merger disclosure is filed; the registration statement and transaction close remain pending. The agreement values the business at $1.25B pre-money enterprise value, with $335M committed financing. Potential trust proceeds depend on redemptions. This is not a completed listing.
Hardware-agnostic AI that protects critical infrastructure from aerial threats.
Founders
Arkadiy Okhman, CEO (former transportation hardware designer with a prior exit), CTO, unnamed on the profile (ex-FBI counterintelligence, two prior exits), CSO, unnamed on the profile (Army Special Forces lieutenant colonel, 30 years in defense)
The deal
Techstars standard: $220,000, as $20,000 for 5% in common stock plus a $200,000 uncapped MFN SAFE.
Counter-drone systems exist, so the two-years-ago test cuts against the category; the specific claim that earns a look is hardware agnosticism, detection and response as a software layer over whatever sensors a site already owns, at exactly the moment cheap drones over data centers and substations became a mainstream worry. The team knows the threat side from the inside: FBI counterintelligence and Special Forces, with a hardware CEO. The case against: defense and infrastructure sales cycles are long, Anduril and the primes own the big contracts, and the profile names only one founder outright.
Monitoring, evaluation and intervention for enterprise AI agents.
Founders
Matt Doughty, CEO (built Procore’s Asia-Pacific customer success function), Simon Russell, CTO (technology and product leadership across Overspecific, Tokio, iflix and Greensync)
The deal
Antler backed the company in November 2024; its profile describes a September 2025 pre-seed round. Investment amount, equity and valuation are not disclosed there.
The original authentication business was possible before current agents. The sharper bet came after customers rejected an external MCP access product: internal agents needed traceable decisions and intervention during execution. Doughty’s post-sales experience explains that distinction. The case against is bundling by established identity and observability vendors; Prefactor must prove its controls prevent consequential errors rather than simply documenting them.
AI that reads decades of siloed drill data and legacy geological reports to produce technical studies for mining companies.
Founders
Chris Mutzel (co-founder; named in GeekWire's demo-day coverage, background not stated), Jimmy Voorhis (co-founder; named in GeekWire's demo-day coverage, background not stated)
The deal
AI House's published standard: up to $600,000 on a SAFE with a $10 million cap, plus 7% in common stock, the same class founders hold.
Mining spends 30 to 50% of project time on manual data preparation by the company's own estimate, and the archives are exactly the unstructured legacy text LLMs just made readable, which is why this could not have been built two years ago. GeekWire reports one customer expanded a $100,000 contract to $900,000 in two months. The case against: the company's own language leads with services and studies, which reads closer to an AI-boosted geology consultancy than a product, in an industry with brutal sales cycles. The portfolio page names no founders; the names come from GeekWire's coverage of AI House's first demo day at Pier 70.
Quality checks for human and robot demonstration data.
Founders
Brandon Ong, CEO (former AI lead at LionsBot; co-founder and CTO of Gigit AI; left a Columbia and NTU robotics PhD), Kingston Kuan, CTO (systems at Jane Street; video platform work at Verkada)
The deal
YC standard: $500K, as $125K post-money for 7% plus $375K uncapped MFN. Company-specific closing documents were not reviewed.
Robot training now buys human demonstrations at a scale where checking duplicates, collection limits and changing procedures becomes its own business. The founders know cleaning-robot deployment and camera-fleet data. Their bet is that useful quality signals can be obtained without training the eventual robot model. The objection is direct: those signals must predict downstream performance, and large buyers may prefer to build the checks themselves.
Printed circuit boards serving demand for AI hardware.
September 29th: HKEX lists the company among today’s new listings, and opening-session reporting confirms trading. The scheduled debut has now occurred; subsequent price performance is not evidence of future operating returns.
Manufacturing equipment with exposure to silicon photonics.
September 29th: HKEX lists the company among today’s new listings, and opening-session reporting confirms trading. The scheduled debut has now occurred; subsequent price performance is not evidence of future operating returns.
An optimization engine that monitors production AI agents and files ranked fixes as pull requests.
Founders
Macy Mody, CEO (go-to-market and operations at SafeBase from zero to $10M ARR through acquisition; Michigan and Harvard MBA), Niyaz Puzhikkunnath, CPO (senior engineer at Amazon on large-scale ML systems; early engineer at Zovi.com), Faiz Vadakkumpadath (ten years as a senior data engineer at Amazon; mobile payments at IBM; co-founded OpenReporting)
The deal
YC standard: $500K, as $125K post-money for 7% plus $375K uncapped MFN. Company-specific closing documents were not reviewed.
The claim worth testing is that agent tuning becomes a continuous, automated discipline: Papaya checks production agents against 200-plus research-backed patterns for quality, cost and latency, then opens pull requests for approved fixes. The week's cost-inefficiency research says the waste is real and large. The case against is crowding: evaluation and observability vendors are converging on the same loop, and the profile's 10% first-analysis quality gain is the company's own number, not an audited one.
The team combines video-action training with its own mobile hardware and fleet learning. Pretrained visual representations could reduce demonstration costs. The objection is physical: understanding a scene does not establish safe contact control, and a demonstration does not measure reliability across workplaces.
A human foundation model that perceives and expresses emotion in real time across speech, face and body.
Founders
Fangchang Ma, CEO (Apple; MIT PhD in robotics and machine learning), Edward Zhang, CTO (research scientist at Apple; University of Washington PhD), Karren Yang (co-founder; named in the funding coverage)
The deal
Terms of SPC's own check are not disclosed on the profile; the company has raised $50 million from Accel, Lightspeed, NVentures and Define Ventures, with South Park Commons in every round.
The bet is full-duplex emotional modeling: one model that perceives and reacts simultaneously across modalities, removing the latency that makes avatars feel wrong. That was not buildable two years ago, and an ex-Apple team with MIT and UW research depth is credible on the perception half. The case against: emotional AI is a crowded claim, the difference between presence and uncanny is a product judgment no benchmark settles, and $50 million buys a lot of demos before anyone proves retention.
RISC-V edge-chip maker; listed in Hong Kong on October 9th at HK$1.55.
September 28th: subscriptions opened at HK$1.48–1.59 for 1.57 billion shares, targeting up to HK$2.5 billion. This is a range, not final pricing. Open instruction architecture does not imply open chip designs. October 1st review: DBS lists an October 9th expected listing date. Retain filed status pending final pricing and completed trading; a subscription timetable is not a debut. [A] October 3rd: books close October 6th, pricing October 7th, listing still expected October 9th, with nine cornerstone investors committed for HK$1.161 billion. Still filed, not priced. [B] October 4th correction: currency verified against the prospectus, printed page 232. [A](https://www1.hkexnews.hk/listedco/listconews/sehk/2026/0928/2026092800035.pdf) October 9th: listed under code 01256, priced at HK$1.55 inside the HK$1.48 to HK$1.59 range and opened at HK$1.54, a flat debut in the week Firmus pulled its Australian book. The closing price was not retrieved. [B](https://technode.com/2026/10/09/risc-v-chipmaker-eswin-computing-lists-in-hong-kong/)
September 29th: HKEX lists the company among today’s new listings, and opening-session reporting confirms trading. The scheduled debut has now occurred; subsequent price performance is not evidence of future operating returns.
Spatial models that reconstruct changing physical environments.
Founders
Srinath Sridhar (Brown University professor and Interactive 3D Vision and Learning Lab director), Aashish Rai (Brown PhD; spatial AI work at Meta and Carnegie Mellon), Tamar Kreitman (embedded systems and capture-hardware development)
The deal
Speedrun standard: up to $1M, with $500K for 10% via SAFE and another $500K in the next round within 18 months. Company-specific terms not disclosed.
Capture hardware and spatial learning meet in a proposed model of factories over time. The founders bring research and embedded-system experience to that interface. Digital twins predate current AI; the case depends on generalizing from observations to useful dynamics, which an attractive reconstruction alone cannot demonstrate.
Spatial-sensor maker listed in Hong Kong on September 30th under code 06802.
The offer price is HK$58.85, with trading scheduled for September 30th. Sensors supply physical observations for robotics, but a planned debut remains distinct from an executed listing. October 1st review: HKEX confirms the scheduled debut occurred. Status changed from priced to listed; no aftermarket return is inferred. [A]
The new possibility is carrying irregular loan documents and operational messages into an auditable sequence of financial calculations. The founders have built credit products and know where operational exceptions accumulate. The case against: calculation software already exists, and an agent must establish reliable reconciliation and approval controls to replace manual work. The profile does not provide an independent audit of those controls.
Connects regulatory and operating data for hazardous-waste businesses.
Founders
Rayhan Razzak (Stanford mathematics; SIG and Peak6 trading; Stanford AI research), Numair Razzak (Stanford economics; Antin Infrastructure, Ares Management and Nomura Greentech), Brian Meng (Berkeley mathematics and computer science; SIG and OMC trading; Berkeley AI research)
The deal
Speedrun standard: up to $1M, with $500K for 10% through a SAFE and another $500K in the next round within 18 months. Company-specific terms are undisclosed.
The opportunity is interpreting messy commercial and regulatory records together before waste-routing decisions are made. Numair’s operating exposure and the team’s experience analyzing waste companies offer a route to identifying costly exceptions. The objection is demanding: ordinary workflow software could already organize these records. Colmez must demonstrate that model-based interpretation adds reliable coverage without hiding a dangerous classification error. The profile does not establish that performance independently.
A shared filesystem for people, applications and AI agents.
Founders
Matthew Ao (previously built eden.so), Ari Bapna (CTO; earlier background not stated in the profile), Jason Zhao (COO; earlier background not stated in the profile)
The deal
Speedrun standard: up to $1M, with $500K for 10% through a SAFE and another $500K in the next round within 18 months. Company-specific terms are undisclosed.
Agents working across creative applications make repeated copying of large assets a practical coordination problem. Ao’s earlier product experience offers a route to observing those workflows. Shared filesystems already existed; the investment case requires agent-specific permissions and concurrency to deliver a new advantage. The strongest objection is that cloud storage incumbents can add those controls while retaining existing customer data. The profile’s site link redirects to spacefs.com.
The new possibility is using capable computer-use agents as the primary phone interface, carrying an intention across apps without a separate integration for every step. The profile confirms the founders but gives no prior operating background. The case against is distribution and reliability: mobile platform owners control permissions, and one incorrect purchase or message can erase the benefit of many successful taps. No independent device tests were reviewed.
SK Hynix storage subsidiary whose review of a potential IPO has been discontinued.
Reuters reports bank interviews and a possible 2027 offering raising $15 billion at up to $150 billion. Early plans; company declined comment. [B] October 25th review: SK Hynix officially ended pre-IPO and IPO review on October 23rd, while retaining expansion needs. Library status withdrawn denotes a halted prospective process, not withdrawal of an established public registration. Alternative financing remains under consideration. [A](https://news.skhynix.com/en/fact-12/)
Connects production monitoring with agent-assisted incident investigation and repair.
Founders
Shalin Patel (Waterloo computer science; on-call experience across nine engineering teams), Deep Utkarsh (Pocus agents; Sigma Computing AI platform and query compiler; Tufts cognitive science)
The deal
YC standard: $500K, comprising $125K post-money for 7% plus $375K on an uncapped MFN SAFE. No company-specific variation disclosed.
The new opportunity is keeping production verification abreast of code written by agents. Patel’s on-call background points to the missing operational context; Utkarsh has built both agents and query infrastructure. The case against: an agent can misdiagnose a correlated failure and propose a harmful fix. The site describes read-only cloud access and a reviewed instrumentation pull request, so its broad repair ambition should not be read as unrestricted autonomous remediation. No independent incident-resolution results were verified.
Announced a $400 million Series F led by Evolution Equity Partners at a $6.4 billion valuation. The bet is that browser-based controls can extend to agent work. Coverage outside the browser and independently measured error prevention remain the tests.
Closed a $75 million Series C co-led by DFJ Growth and Premji Invest. Funding supports its data platform and therapeutic pipeline, including a program in Phase 1. Clinical entry is a milestone, not demonstrated safety or efficacy.
Builds an AI control layer over existing industrial reactor controls.
Founders
Andy Xian (Harvard statistics; Balyasny Asset Management internship), Achraf Zemzami (Audible; Harvard computer science and physics), Aias Tatsis (Harvard electrical engineering; built Acaysia’s physics engine and pilot plant)
The deal
Speedrun standard: up to $1M, with $500K for 10% through a SAFE and another $500K in the next round within 18 months. Company-specific terms are undisclosed.
The opportunity is making learned process models and rapid simulation useful for online reactor control without replacing installed equipment. Tatsis’s work on both the simulator and physical pilot gives the team a way to test the gap between them. The case against: industrial optimization predates modern AI, and this company must prove a new advantage under changing process conditions and hard safety limits. Public profiles do not establish independent plant-scale performance.
The emerging opportunity is an agent that carries business state across tasks instead of answering isolated requests. Lipp’s retrieval-company background could help with the difficult question of which records an agent should trust. The case against: the public profile offers a broad ambition with no specific operating workflow or independently verified reliability result. This is an early thesis to watch, not evidence that a company can already run unattended.
Announced the first close of a $580 million Series D at a $6.4 billion valuation, led by UC Investments and Baillie Gifford. This is private financing, not an IPO. Scaling physical systems still requires production and deployment evidence beyond a funding announcement.
Uses agents to search for model optimizations specific to the target hardware.
Founders
Momčilo Mrkaić (Databricks software engineer; Microsoft ML intern; City University of Hong Kong), Pavle Padjin (Tenstorrent AI compiler engineer; Cambridge AI MSc; Microsoft ML intern), Vladimir Zeljkovic (Tenstorrent AI compiler engineer; JOANNEUM RESEARCH ML intern; University of Belgrade)
The deal
SR007 standard: up to $1M, with $500K for 10% through a SAFE plus $500K in the next round within 18 months. Company-specific terms are undisclosed.
The new opportunity is letting capable coding agents search and test hardware-specific optimizations that previously consumed scarce compiler-engineering time. The founders have worked on both model infrastructure and alternative accelerators. The case against: a faster kernel can break numerical behavior or lose its advantage in a complete workload. Hoid needs independently reproducible end-to-end results across hardware, including the cost of finding the optimization.
Connects industrial planning and business records to AI agents and robot execution.
Founders
Datta Kaligotla (founding engineer at Handshake AI; enterprise engineering at Microsoft, Endeavor and Boeing; quantitative work at Pretium Capital), Oscar Rangel (factory systems at Worthington Steel; medical devices at Stryker; projects for NASA astronauts)
The deal
YC standard: $500K, comprising $125K post-money for 7% plus $375K on an uncapped MFN SAFE. No company-specific variation disclosed.
The emerging opportunity is connecting language-driven work to robot execution and business records in a shared runtime. Rangel’s factory systems background points at the difficult handoff between planning and the shop floor. The case against: most listed functions already exist in industrial software; the new agent layer must reduce integration work and exceptions enough to justify replacement. The profile and company site do not establish that advantage in production.
Announced a $140 million Series C led by S32 to train new EDEN models and advance therapies toward clinical development. The company reports preclinical results, not demonstrated benefit in patients. Funding does not settle that scientific test.
Research lab training cooperating groups of models.
Founders
Ed Li (Yale PhD program; mathematics and philosophy at Oxford), Cat Yan (mathematics and statistics at Oxford), Junyu Ren (University of Chicago PhD program; mathematics and philosophy at Oxford)
The deal
SR007 standard: up to $1M, with $500K for 10% through a SAFE and $500K in the next round within 18 months. Company-specific terms are undisclosed.
The opportunity is training cooperation across models as a source of capability, now that agents can sustain longer research and engineering runs. The profile says the team previously built freephdlabor to automate research, giving it experience with evaluation beyond conversation. The case against: communication and verification costs can erase the savings from smaller models, and coordinated errors can survive peer review. Frontier performance and cost claims have not been independently established.
Identifies agents using customer accounts and applies access rules to their web activity.
Founders
Tommy McCormick (CEO; prior education and employment not disclosed on the reviewed profile)
The deal
SR007 standard: up to $1M, comprising $500K for 10% through a SAFE plus $500K in the next round within 18 months. Company-specific terms are undisclosed.
The opportunity comes from browser agents inheriting a human session while consuming software in a different way. Switchfrog proposes detecting that handoff and converting repeated workflows into approved tools. The case against: distinguishing agents from legitimate human automation can produce false positives, and browser vendors can change the signals. The profile names only the CEO and supplies no prior background; detection accuracy is unverified.
Moves inference for large robot models into nearby cloud infrastructure.
Founders
Eric Ren (previously AI safety; employer not disclosed), Kevin Lin (former quantitative researcher turned roboticist; employer not disclosed), Antoine Nguyen (vision-language-action model quantization research at UTS), Chris Yoo (networking certification and hackathon background; prior employer not disclosed)
The deal
YC standard: $500K, as $125K post-money for 7% plus $375K on an uncapped MFN SAFE. No company-specific variation disclosed.
Growing robot models make inference placement a new constraint: hardware on the robot cannot always hold the desired system. Nguyen’s quantization work bears directly on that tradeoff. The strongest objection is network dependence. Lower typical cloud latency is insufficient if a delayed response interrupts a physical task. The team needs to demonstrate tail latency and safe behavior during disconnection, beyond its own speed comparisons. [A](https://www.ycombinator.com/companies/dreamscale-labs)
Combines generative imagery with controllable 3D scenes and cameras for creative production.
Founders
Bharat Vasan (technology founder and CEO; former investor at TPB), Charles Migos (former design leader at Apple and Unity Technologies)
The deal
Company-specific terms are undisclosed. Speedrun currently publishes $500K for 10% through a SAFE plus $500K in the next round within 18 months; those terms are not verified for SR004.
The current opportunity is combining generative output with directorial control that survives revisions. Migos has built tools for creative professionals, where camera placement and scene consistency determine whether a result can be used. The case against: established 3D editors can integrate the same models, and attractive renders do not establish reliable production assets. Test repeated camera and object edits on one scene.
Post-trains speech models on production calls and serves them against live traffic.
Founders
Pavan Katta (led voice inference and post-training at Salient; IIT Madras), Jashwanth Pedapudi (kernels engineer at Meta Reality Labs; IIT Madras electrical engineering)
The deal
YC standard: $500K, comprising $125K post-money for 7% plus $375K on an uncapped MFN SAFE. No company-specific variation is disclosed.
The bet is that current speech models make continuous adaptation to real calls useful enough to sell as infrastructure. Katta has operated voice inference, while Pedapudi brings low-level runtime experience. The case against: established speech providers can absorb the same feedback loop, and customer call data creates consent and evaluation burdens. Faster inference matters only if held-out calls retain accuracy.
Industrial robotics company with an existing Hong Kong listing application and a reported completed CSRC overseas-listing filing.
Bamboo Works reports a regulatory notice published last Friday but dated August 24th. This records an existing application and clearance step, not a completed offering. The primary notice was not independently retrieved; offer pricing and a listing date remain unverified. October 1st review: the linked CSRC notice was retrieved, confirming the overseas issuance filing step. It does not establish a completed offering; retain filed status. [A]
Label and claims review is bounded, text-heavy and repeated across thousands of SKUs, which is the shape of work current models actually finish, and it only became automatable when extraction and regulatory reasoning got reliable enough to trust on a printed label. The case against: the profile states no compliance or CPG background for either founder, a two-person team is selling into a liability function where the customer's lawyer still signs everything, and incumbent label-compliance vendors can bolt the same models onto existing rulebooks and relationships.
Post-training infrastructure turns expert decisions into improvements to domain-specific models.
Founders
Margaret Zhang (CEO; prior affiliation not disclosed on the reviewed profile), David Huang (CTO; prior affiliation not disclosed on the reviewed profile)
The deal
Company-specific terms are undisclosed. Speedrun currently publishes $500K for 10% through a SAFE, plus $500K in the next round within 18 months; application of those current terms to SR006 is unverified.
The interesting claim is continuous capture of expert decisions while people work, reducing the cost of maintaining specialized models as capable base models become available. That is more specific than a fine-tuning service. The case against: feedback can encode inconsistent expert habits, and continuous updates need independent evaluation to avoid regression. Founder backgrounds and the actual investment terms remain undisclosed in the reviewed profiles.
Dictating into agents is becoming the default input and speaking aloud is the bottleneck in every shared space, so the pull for silent speech is new even though MIT's AlterEgo demonstrated the sensing years ago. The case against sits in the company's own job posts: it is hiring a head of ML, a head of sensing and a head of hardware, meaning two CS graduates are recruiting the entire core stack of a neuromuscular-sensing product they have not yet demonstrated, while Meta's wristband and whisper-mode transcription squeeze the use case from both sides.
Data agents use column lineage and team knowledge to plan changes across a data stack.
Founders
Rick Gao (Verition Fund Management; Yale computational biology research), Tony Li (ByteDance agent workflows; Carthage Capital), Andrew Tran (Ramp and Hudson River Trading internships; Yale)
The deal
YC standard: $500K, comprising $125K for 7% plus $375K on an uncapped MFN SAFE. The profile discloses no company-specific variation.
Coding agents can now carry out data changes whose consequences extend far beyond one query. Alkera makes downstream lineage part of planning, drawing on the founders’ work with financial and production data. The case against: incomplete lineage can create false confidence, and warehouse vendors can add similar controls. The useful test is whether it catches a plausible transformation that silently changes a business metric.
Agents read camera feeds, write robot-control code and call existing models to produce reusable programs for physical tasks.
Founders
Hanming Ye (Harvard; topology research at MIT's Research Science Institute), Yiding Song, also Vincent (Harvard; multimodal astrophysics research at MIT's Research Science Institute)
The deal
YC standard: $500K, comprising $125K post-money for 7% plus $375K on an uncapped MFN SAFE. Company-specific variations are not disclosed on the profile.
Better coding agents make it plausible to compose reusable robot skills without retraining a complete model for every setup. The founders describe skills carrying between tasks, a more specific thesis than adding a chat interface to a robot. The case against: physical resets, control latency and safety may dominate the cost of writing the program, and the claim of working across any robot remains unproven.
An AI-enabled platform for data-center energy procurement, automating the negotiation and diligence behind power purchase agreements and interconnection deals that conventionally run six to twelve months per deal. Techstars describes it as taking procurement from months to weeks.
Founders
Harleen Sindhu (CEO; 19 years in energy, led power procurement for AWS data centers across EMEA, founding team of Shell New Energies)
The deal
Techstars standard: $220K, as $20K for 5% common stock plus $200K on an uncapped MFN convertible note. No further raise stated on the cohort page.
Power procurement is the binding constraint of the buildout, the buyer universe just exploded from a few hyperscalers with in-house teams to a crowd of neoclouds with none, and Sindhu ran exactly this function for AWS across EMEA. The case against: the real bottlenecks are interconnection queues and transformer lead times, which faster paperwork does not move; the biggest buyers keep procurement in-house; and LevelTen's marketplace and the CBRE and Schneider advisory teams already sit on the transaction. The cohort page names only the CEO, describing a CFO from fifteen years in investment banking without naming them.
Hisense-controlled optical-transceiver maker listed in Hong Kong on September 22nd under code 09856.
The optical interconnect layer keeps listing through Hong Kong while the compute layer files in New York. Cornerstones at nearly half the base deal say institutional demand was arranged before the book opened, which is how Hong Kong de-risks a listing rather than a signal of open-market appetite. September 28th review: Reuters reported that Ligent completed its Hong Kong debut on September 22nd. October 1st review: HKEX confirms the completed listing; the description no longer presents the debut as a future event. [A]
Luxembourg-based AI and computing infrastructure developer, with projects including Norway's PolarDC and Israel's Anan, weighing an IPO of about $500 million in Europe or the US next year. Bloomberg reports it is working with STJ Advisors to hire banks; no final decision has been made.
Rumor of record per Bloomberg's sourcing. A $500 million European infrastructure listing is small next to Nscale, which is exactly why it is informative: the pipeline now has a mid-market tier forming behind the headline names.
Real-time, on-device audio intelligence: models that isolate, interpret and manipulate specific sounds in noisy environments at earbud latency, rather than bluntly canceling noise. The commercial vehicle for the University of Washington lab behind the programmable sound bubble work published in Nature Electronics.
Founders
Shyam Gollakota (professor, Paul G. Allen School of Computer Science and Engineering, University of Washington), Malek Itani (embedded systems and audio AI, University of Washington), Tuochao Chen (embedded systems and audio AI, University of Washington)
The deal
AI House standard: up to $600K on a SAFE at a $10M cap, with 7% common stock for incubation. Hearvana announced a $6M pre-seed in November 2025 co-led by Point72 Ventures and SCB 10X, with the Amazon Alexa Fund participating.
The founders invented the field's headline results in their own lab, and shipping-quality neural sound separation inside an earbud's power budget only became feasible with the latest edge silicon, so the two-years-ago test passes cleanly. The case against: Apple, Google and Samsung own the devices, the silicon and world-class audio teams, the methods are published academic work, and a licensing startup lives on the platform owners' integration cycles.
Analytics for MCP servers, Claude Connectors and ChatGPT Apps, with session replay and detection of where visiting agents get stuck.
Founders
Kashish Hora (previously led AI B2B at Grammarly/Superhuman), Naseem Alnaji (previously founded and exited Opal Security)
The deal
Speedrun standard: $500K for 10% upfront through a SAFE, plus another $500K in the next round within 18 months. Company-specific variations are not disclosed on the profile.
When agents become software users, a clickstream no longer explains their goals or failures. AgentCat proposes replay and issue detection across MCP and assistant connectors. The founders combine enterprise AI and security experience, relevant to inspecting those sessions. The strongest objection is that agent traces can be incomplete or contain sensitive customer data; useful diagnosis must survive redaction and competing observability platforms can add similar views. [A](https://speedrun.a16z.com/companies/agentcat)
Checks agent communications against organizational policies before sending.
Founders
Kumesh Aroomoogan (built financial-services language-processing products; prior company acquired by Wand AI in 2025)
The deal
Profile reports a $10M seed round in June 2026, without valuation or company-specific accelerator terms. Published Speedrun standard is up to $1M: $500K for 10% via SAFE, plus $500K in the next round within 18 months.
Autonomous agents now send communications at a rate that makes manual pre-send review impractical. Aroomoogan’s work with financial institutions gives the team experience with regulated text and evidence requirements. The strongest objection is that judging each message can miss a violation distributed across a longer workflow. Policy changes, false blocks and incomplete channel coverage could erase the labor savings. The profile describes intended enforcement, not independently established compliance.
Models how connected consumer groups respond to proposed market decisions.
Founders
Nikita Mullangi (human behavior and decision-making research at MIT; UC Berkeley computer science), Cameron Malloy (AI agents at GI Partners; financial contagion models at Nelumbium Capital; UC Berkeley M.E.T.)
The deal
YC standard: $500K, as $125K post-money for 7% plus $375K uncapped MFN. Company-specific variations are not disclosed.
Agent models can now turn unstructured consumer evidence into interacting populations that respond to proposed decisions. Malloy’s contagion-model work is relevant to reactions spreading between groups. The objection is identification: reproducing a past campaign does not show what an untried price or product would cause. The profile contains differing accuracy claims without enough methodology to choose between them; neither is treated here as validated performance.
Agentic commerce infrastructure: a protocol layer that lets AI agents transact on behalf of brands, rather than the brand hoping a shopping agent represents it correctly. Chicago; pilots being scoped with LVMH/Sephora and P&G per Antler's cohort page.
Founders
Samar Birwadker (3x founder with two exits; built a Visa card-linked cashback product to $25M ARR and 15 million users with his co-founder), Elton Cheung (ex-Rakuten and Groupon)
The deal
Antler US standard: $250K at $2.75M post-money for about 9%; Antler's cohort page reports $1M closed on a $2M pre-seed, with Corazon Capital also invested.
Agent checkout only became a real surface this year, and Parleo bets brands will pay to control how agents transact with them instead of being flattened into a price row. The founders have shipped payments rails with Visa before, and Antler's page says the protocol has been validated with Visa, Stripe and Anthropic. The case against is heavy: OpenAI, Google and the card networks are standardizing agentic checkout themselves, and a two-sided protocol startup must win adoption before the giants finish writing theirs.
An AI-native operating system for the restaurant back-of-house: computer vision and voice for inventory, demand forecasting, and agents that place orders, build schedules and reconcile invoices. Renamed from Truffle.
Founders
Aakar Khanna (CEO; covered restaurants at Goldman Sachs, then Head of Strategy at Kouper), Arjun Chaliha (CTO; senior software engineer at Bloomberg leading AI/ML work)
The deal
YC standard: $500K, as $125K post-money for 7% plus $375K uncapped MFN.
Back-of-house is where restaurant software has always died: thin margins, greasy tablets, staff who will not learn a new tool. The bet is that vision plus voice removes the data-entry step that killed the last generation, and a founder who priced restaurant risk at Goldman knows exactly how thin the buyer's margin is. The case against: restaurant tech's graveyard is full of better mousetraps that could not survive 20-location pilots, and Toast owns the counter this system must plug into.
Specializes frontier models to forecast events with scored probabilities.
The deal
$25 million seed round led by Radical Ventures; valuation undisclosed, Reuters reports.
Tournament outcomes offer a measurable test of forecasting, which is stronger evidence than a persuasive narrative. The commercial objection is distribution shift: business questions may be less clearly resolvable and more sensitive to strategic behavior than public tournament questions. A seed financing does not establish durable forecasting profits.
Shared organizational memory drawn from meetings, messages and agent execution records.
Founders
Ashwin Gopinath (MIT professor; Caltech and Google research; Reflexion framework), Andrey Starenky (AI engineering at IBM and Vapi; MIT machine-learning research)
The deal
Published Speedrun standard: up to $1M, with $500K for 10% via SAFE and another $500K in the next round within 18 months. Sentra’s specific SR006 terms are not disclosed.
Gopinath’s work on agent reflection gives the team a specific reason to care about remembered actions, beyond searching old documents. Longer-running agents now produce decisions and tool records that can become shared organizational context. The strongest objection is permission drift: yesterday’s accessible meeting can become tomorrow’s unauthorized memory after roles change. A useful system must forget or restrict information as reliably as it recalls it. The profile establishes the team and thesis, not measured improvements in cross-company coordination.
CRV-led Series A; total funding reaches $50 million. Round size not separately disclosed.
The simulation preview expands the company beyond production monitoring. Adoption will depend on how faithfully simulated conditions reproduce customer systems.
The founders connect chip-simulation knowledge with agents that can propose a fix and check its waveform after rerunning a test. Waveform compression alone could have been built earlier; the newer signal is making an agent operate that verification loop. Their diagnosis is specific: engineers need to trace a bad signal through enormous records, not merely generate more hardware code. The strongest objection is that a passing simulation can still miss a design defect, while established EDA vendors control the surrounding workflow. Vendor demonstrations do not establish production reliability.
$400 million Series C, announced as closed on September 17th 2026. Valuation and the named identity of the strategic internet-company investor were not disclosed.
The financing supports a shared hardware and software supplier across robot categories. That can spread development costs over more than one robot maker. The objection is customer economics: component shipments do not show that customers can deploy profitable robots. The funding and shipment claims are from a company press release, not audited results.
Factory robots that adapt to a changed task from one human demonstration.
Founders
Navid Aghasadeghi (productized Spot and led manipulation work at Boston Dynamics; Rethink Robotics), Mojtaba Hejrati (whole-body control and manipulation at Amazon Robotics; Northwestern research faculty)
The deal
YC standard: $500K, as $125K post-money for 7% plus $375K uncapped MFN. Company-specific variations are not disclosed.
The founders have spent more than 20 years on deployed robotics and target high-mix factories where fixed automation breaks when the line changes. Stronger vision-language-action models make one-demonstration adaptation plausible now. The objection is reliability at the edge: a convincing demonstration does not show how the robot handles drift, novel obstructions or safety stops through a full production shift. Watch intervention rates and relearning time at customer sites.
Automates medical pre-charting, fax triage and coding inside existing electronic medical records.
Founders
Tejasvi Chebrolu (healthcare AI researcher at Triomics; UT Austin), Abhijit Manatkar (LLM research at IBM Research; IIIT Hyderabad), Tattvam Nair (medical graduate of JIPMER)
The deal
Entrepreneur First says it invests up to $250K after a team forms. The company profile does not disclose Layrd's specific terms.
The team combines healthcare AI research, language-model research and medical training around work that still arrives through faxes and fragmented records. Better document models make full workflow handling plausible now. The strongest objection is clinical liability: a missed result or wrongly filed document can harm a patient, and the company's claim of zero hallucinations is its own production metric. Watch edited and escalated cases, not documents processed.
A WhatsApp procurement agent connecting Brazilian farmers with agricultural suppliers.
Founders
João Souza (farming family; agribusiness transactions at BTG Pactual; USC), Pedro Galindo (agricultural roots; technology investing at Patria Investments; USC)
The deal
Published Speedrun standard: up to $1M, with $500K for 10% via SAFE and $500K in the next round within 18 months. The profile does not disclose Vereda’s SR006 terms; these are a program reference.
The founders bring agricultural distribution and investing experience to a specific bottleneck: purchasing intent arrives in informal conversations. A language agent can structure those exchanges without asking farmers to adopt another business system. The objection is that distributors also provide credit, delivery and trust. Removing their ordering work does not remove those obligations, and an incorrect product match can be costly. Watch whether the company improves the complete purchase, beyond generating a cheaper quote.
Workflow software for recovering long-running applications and agent jobs.
The deal
$550 million raised at a $12.55 billion valuation, per Reuters on September 14th 2026.
The financing is reported as completed. Longer agent jobs create more opportunities for failures between model calls, making persistent workflow state commercially useful. The strongest objection is that cloud platforms can bundle similar recovery features. Funding does not establish profitability or an IPO timetable.
The founders met in an art program and describe research experience in multimodal models and computer vision. Their diagnosis is that moving between creative tools loses brand context. Better controllable generation makes a repeatable production workflow more plausible now than a collection of isolated prompts. The objection is distribution: established creative suites can add the same context layer, while dependable brand fidelity still requires human review. The profile establishes the team’s thesis, not independently measured production quality.
Peake’s Apple architecture work informs a specific complaint: short coding puzzles do not measure the iterative design loop or its power and area tradeoffs. More capable tool-using models make training across that loop plausible now. The objection is transfer: exact grading inside an environment does not establish that the resulting designs survive manufacturing constraints and production verification. This is an evaluation supplier, distinct from a company generating analog circuits.
Management is considering a UK or US listing and has not committed to a date.
Pipeline catch-up from Reuters’ August 19th interview with CEO Vishal Marria. He says the company could list in either country, both, or remain private. The report separately discusses a possible shareholder stake sale; that is not an IPO. No September decision deadline is supported by the interview.
Australian IPO withdrawn October 9th; Firmus plans private fundraising.
October 5th: Reuters reports up to US$5.5 billion including the overallotment option. A secondary account reports Australian dollars instead. Reuters cites a term sheet for pricing; the listing has not occurred. [B](https://www.streetinsider.com/Reuters/Australias%2BFirmus%2Bto%2Ballocate%2Bhalf%2Bof%2Bup%2Bto%2B%245.5%2Bbillion%2BIPO%2Bto%2Bexisting%2Binvestors%2C%2Bsources%2Bsay/27144700.html)[C](https://finance.yahoo.com/markets/stocks/articles/firmus-plans-allocate-half-ipo-042136602.html) October 8th update: Bloomberg reports books closed without a clear final price or deal structure. Reuters Breakingviews cites efforts to hold A$8.25 after discussion of A$8; earlier DCD reporting cited A$9. The priced status preserves the previously reported A$11 term sheet, not confirmation of a revised price or completed listing. [B](https://au.finance.yahoo.com/news/nvidia-backed-ipo-cratering-demand-102043878.html)[B](https://www.breakingviews.com/columns/breaking-view/shaky-4-bln-neocloud-ipo-pricks-ai-bubble-2026-10-08/) October 9th update: Reuters confirms withdrawal and a planned private round. A subsequent Nasdaq listing is attributed to an unnamed person and was not confirmed by Firmus. Earlier prices above describe the abandoned offering. [B](https://uk.marketscreener.com/news/australian-nvidia-backed-ai-data-centre-operator-firmus-shelves-ipo-ce785ddfdc8df725) October 11th update: the parallel run's account adds figures for the abandoned offering. A$11 a share sought, an implied equity value near US$30.6 billion, roughly US$51 million of FY2026 revenue, a reported US$30 billion debt pile, and demand collapsing within 48 hours of books opening. [B](https://www.bloomberg.com/news/articles/2026-10-09/data-center-darling-s-30-billion-ipo-dream-crushed-in-48-hours)[B](https://www.nst.com.my/amp/business/corporate/2026/10/1551088/nvidia-backed-firmus-scraps-us5bil-australian-ipo-poor-demand) Weekly synthesis, October 11th: an abandoned public book establishes a failed offering at its proposed terms. It does not establish a market-wide clearing price, nor a closed replacement private round. The earlier dated accounts above remain part of the record.
Hardware and infrastructure supplying real-world voice context to AI agents.
Founders
Ruming Zhen (Stanford Robotics, Tesla and Apple; previously YC), Qi Zhang (Amazon; previously YC)
The deal
Current published Speedrun standard: up to $1M, with $500K for 10% via SAFE and $500K in the next round within 18 months. The profile does not disclose Bota’s SR006 investment terms; this is a program reference, not a confirmed company deal.
The founders have built both hardware and infrastructure and previously sold a startup together. Their claim is that useful agents need context captured beyond the screen. Stronger voice and tool-using models make that input more actionable now. The objection is consent and relevance: recording more speech can add sensitive material without improving decisions. This is a capture-and-delivery bet, distinct from a speech API that labels emotion. The profile supplies no independent evidence of decision quality.
The claim to watch is generalization to targets lacking training examples, combined with an automated design-and-test loop. Heifets has already built a drug-discovery company; the profile identifies missing chemical starting points as the constraint this attempt addresses. Better agents and cloud labs make a tighter experimental loop plausible. The objection is biological: identifying a binder does not establish a safe oral medicine. The reported experimental validations remain company claims.
AI agents for railroad paperwork, incident reporting and billing.
Founders
Sai Surisetti (founding engineer at Toma; left the University of Waterloo), Gautham Venkateshwaran (founding engineer at Toma; left the University of British Columbia)
The deal
Published Speedrun standard: up to $1M, with $500K for 10% via SAFE and $500K in the next round within 18 months. Company-specific terms are not disclosed.
The founders are second-generation railroaders who also built agent software at Toma. That combination offers a reason to believe they know where railroad work escapes formal systems into calls and email. The new possibility is handling those unstructured exchanges across an operational workflow. The objection is reliability: an incorrect incident report or shipment update can propagate beyond the original conversation. Rail-specific exceptions and human review could absorb the labor the product promises to save.
Enterprise data-software company listed in Hong Kong on September 21st under code 06727.
The deal
14,010,800 H-shares offered at an indicative HK$49 to HK$61 per share.
The exchange announcement expects pricing September 17th and trading September 21st. Those dates are conditional. Track the final allocation and offer price before treating the proposed debut as a completed listing. September 28th review: Infocast’s live performance table records Transwarp’s September 21st debut, changing this record from an intended listing to a completed one. October 1st review: HKEX confirms the debut; the description now matches listed status. [A]
Agentic systems for supply-chain operations that learn by imitating expert operators.
Founders
Eduardo Candela (CEO; PhD in AI for autonomous vehicles at Imperial College London; operations research at MIT; co-founded Maihem, YC W24; Tesla and the Bosch Center for AI), Juan-David Rodriguez (COO; founded Teilur; growth at Investopedia, Stash and Springboard; CFA charterholder)
The deal
Speedrun's SR007 standard: up to $1M, as $500K for 10% via SAFE plus $500K into the next round within 18 months, per a16z's program page.
The method is the claim: imitation learning from recorded expert workflows rather than prompted generalist agents, which is the technique Candela spent a PhD applying to driving, the original imitation-learning domain. Supply-chain ops is full of repetitive judgment calls that are demonstrated daily and never written down, which is exactly where imitation beats instruction. The objection is drift: enterprise processes change under the agent, imitation degrades silently off-distribution, and the AI-employee category is the most crowded in enterprise software. The profile's traction line, live with leading global enterprises, is the company's own.
$875 million financing at a $5 billion valuation, according to Liberty Global.
The investment backs a route around constrained memory and packaging supplies. The next chip’s announced production schedule extends into the second half of 2027, leaving customers to weigh promised future capacity against systems they can deploy now.
Regression testing for AI agents: rebuilds tools from production traces and replays real sessions against every code change.
Founders
Aryan Singhal (ML research at UT Austin and UCSB; six AI/ML papers in high school, three at NeurIPS). The profile lists a team of two but names no second founder.
The deal
YC standard: $500K, as $125K post-money for 7% plus $375K uncapped MFN.
Agent teams ship prompt and tool changes with no equivalent of a test suite, and the failures surface in production because the input distribution is real users, not fixtures. Replaying recorded sessions against every change is the obvious missing CI layer, and it could not have existed before there were production agent traces to replay. The objection is that replay tests a frozen world: tools, APIs and model versions drift, so passing yesterday's sessions proves less than it appears to, and the labs' own eval platforms could absorb this as a feature. A solo-led team of two against that race is the bet.
AI credit servicing for small consumer balances, with a broader regulated-workflow offering on its own site.
Founders
Michael Goh: speech-model benchmarking consultant for Amazon and Play AI; previously a consumer-fintech investor at Quest Ventures., Rafael Khaykin: worked on dexterity controls at Figure Robotics; left the University of British Columbia.
The deal
The EF profile does not disclose the investment amount, instrument or equity terms.
The bet is that conversational agents can handle variable servicing conversations within explicit contact and offer rules, beyond an old automatic dialer. Speech benchmarking and consumer-finance experience meet in Goh’s background. The strongest objection is economic: reviews, disputed balances and integrations may consume the saving on small accounts. Company-described controls are not an independent compliance assessment.
AI agents for scientific experiment design and research workflows.
Founders
Mattias Akke: machine-learning research at MIT and AstraZeneca; protein-misfolding research at Cambridge., Linus von Ekensteen Löfgren: desk quant at Danske Bank, AI-security research at AI Sweden and quantum research at Chalmers.
The deal
YC standard: $500K, as $125K post-money for 7% plus $375K uncapped MFN. Company-specific variations are not disclosed.
The interesting bet is that agents can revise experimental plans as results arrive, rather than merely search papers. The founders combine biological modeling with quantitative and security work. The strongest objection is that useful plans still depend on noisy physical experiments: a convincing research proposal is not laboratory validation. The profile names Halmos Labs; its linked site uses Halmos Bio.
Reported preparations for a US IPO that could raise more than $2 billion.
Reuters cites sources describing preparations and a possible confidential filing. The earliest reported timing is the end of 2026. These are sourced IPO plans, not proof that a registration has been filed or shares priced.
A speech API that returns how something was said: transcript plus emotion and speaking style.
Founders
Nathan Roll (co-founder and CEO; paused a Stanford Linguistics PhD; Stanford NLP, first-author EMNLP 2025 paper with Dan Jurafsky; founded Strype Analytics; data scientist at Nalu.ai). Other co-founders are not named on the profile.
The deal
The Speedrun profile does not disclose the investment amount, instrument or terms.
Voice agents reached production before anyone solved paralinguistics, and the gap between a usable and an unusable agent is often tone rather than words. Roll's research background is the bet: his EMNLP work showed about 50 seconds of example speech cutting recognition error 19.7 percent, evidence the channel is learnable from little data. The profile claims a $250K run rate in four months and the first selection for both a16z Alpha and Speedrun. The strongest objection is structural: frontier realtime models increasingly emit prosody and emotion natively, and Hume AI is an established incumbent, so a standalone analysis API may become a feature before it becomes a company.
Industrial robot systems for material handling and assembly.
The deal
$100 million Series A, announced by the company; valuation and detailed terms not disclosed in the reviewed announcement.
The factory-level system scope is the reason to track the financing. Integration and service requirements can consume the savings from autonomous handling. The announcement does not provide an independently measured cost per completed task.
AI infrastructure for administering inheritance and estate settlement.
Founders
Arum Kempster (Velocity Black growth; death-policy work for the military and UK government), Jacob Effenberger (AI work at Man Group; death-logistics software; ETH Zürich master’s)
The deal
The EF profile does not disclose the investment amount, instrument or equity terms.
Kempster’s experience administering an estate gives the team a view of the handoffs that delay inheritance; Effenberger has worked on AI in regulated environments. Estate software could have existed two years ago. The newer bet is that models can interpret varied documents and coordinate the work at a lower service cost. The strongest objection is liability: a missed beneficiary or jurisdictional requirement can overwhelm the savings. The profile does not establish reliable autonomous settlement, so this remains an early deployment thesis.
Watermarking predates this AI cycle; the newer opportunity is authenticating live voice agents through ordinary phone connections. DeepMark places the identifier in audio so stripping file metadata does not remove it. That is a specific design bet, though the profile supplies no independent robustness test. The strongest objection is ecosystem adoption: an unverifiable call must remain usable, and compression or deliberate transformation may erase a marker. Founder research backgrounds are disclosed only at degree level.
Optical connections between AI chips; Series E extended by $150 million.
The deal
$150 million extension to the $500 million Series E announced in March, per Reuters.
The funding bridges development and manufacturing preparation. Commercial timing is the key risk: copper-based alternatives and other optical suppliers continue improving while this generation is being prepared.
Photonic processors aimed at inference in autonomous systems and edge devices.
Founders
Heather Jones (radio-frequency sensors and onboard processors at MBDA; physics at Durham University), Gurneet Bhatia (signal-processing algorithms at Barclays; mathematics at Imperial College London)
The deal
The EF profile does not disclose the investment amount, instrument or equity terms.
The emerging opportunity is inference for autonomous machines that cannot depend on a data-center connection. Jones has worked on sensors and onboard processing, where power and response deadlines constrain the architecture. Optical computing itself predates this AI cycle; the company must show an advantage on today’s model workloads. The strongest objection is that conversion, memory and control costs can erase a faster optical operation. The profile supplies no measured system-level gain, so this is a hardware thesis to watch, not a demonstrated discontinuity.
A changing benchmark and training-data loop for shopping agents.
Founders
Shardul Bansal (Crucible Labs founding engineer; AWS Cloud Map; distributed-training evaluation research at the University of Toronto), Seth Schilbe (AWS distributed systems; cofounder of the HomeWurk services marketplace)
The deal
YC standard: $500K, comprising $125K post-money for 7% plus $375K on an uncapped MFN SAFE. Company-specific variations are not disclosed.
Shopping agents create a new evaluation problem: inventory, prices and user constraints move while a benchmark ages. ORO proposes changing tasks and independent validators, with completed runs feeding training. The founders combine marketplace and distributed-systems experience. The strongest objection is incentive design: rewarding a high score can teach agents to exploit the validator instead of satisfying a shopper. The profile does not establish independent evidence that its rewards predict successful real purchases.
Industrial robotics and embodied-AI company listed in Hong Kong on September 1st under code 09615.
Added during the listing-pipeline sweep as an earlier September debut, not a new offering today. The IPO calendar identifies the company and the broker’s listed-security table confirms the trading date. Price performance does not establish the economics of its installations.
Commercial-service robot maker listed in Hong Kong on September 9th under code 03231.
HKEX confirms the listing date and security code. This record uses the exchange to establish status; it does not infer profitability from the first-day move or conflate commercial-service robots with humanoids.
NYSE listing through a business combination, with associated financing reported closed.
Unusual Machines' announcement identifies XTEND as listed and says its additional investment is already included in the financing closed with the combination. This is a business-combination listing, not evidence of a conventional IPO bookbuild. The source does not establish the first trading date or an opening share price.
The recent capability bet combines compact local models, reused context and computer actions to prepare work overnight on ordinary Macs. Tarun has already built local writing software, giving the team experience with the constraints it is trying to remove. The strongest objection is the privacy boundary: the profile says some work still uses a customer's cloud-model subscription. Local preparation does not make every later action local. Reliability and battery costs are not independently established.
AI coding creates applications whose owners may lack the engineering skills to keep them running. Nometria bets that automatic maintenance can turn that new supply into durable businesses. The founders have operated real infrastructure and financial software, which is relevant to failure recovery. The strongest objection is service cost: unfamiliar generated code can turn every customer into a different debugging engagement. The profile states the product ambition but supplies no independent evidence that maintenance is reliably autonomous.
The new bet is that language-model planning can make mission-level choices while established flight software constrains execution. Ososke's fault-recovery background is relevant to the boundary between those layers. The strongest objection is verification: a safe command interface does not prove that a sequence of individually permitted commands preserves the mission. No flight validation is established by the profile.
Filed a public S-1 on September 18th 2026 for a NYSE listing under NSCL. First-half 2026 revenue of $140.6 million against a net loss of $1.02 billion; $103.4 billion in active and contracted value per CNBC, anchored by a $44.6 billion Anthropic contract and up to $43.8 billion from Microsoft; $3.1 billion of convertible notes issued, at least $1 billion to Nvidia. Last private valuation $14.6 billion in March 2026.
Upgraded from rumored on the public filing. The gap between contracted value and recognized revenue is the listing's whole argument, in both directions: the contracts say demand is priced, the loss says the capital to serve them has to be raised first. September 25th: announced $3.36 billion in convertible financing, including $1 billion from Nvidia expected in November. [A] October 19th: Bloomberg reports investor discussions at $16 billion to $19 billion against roughly $30 billion sought. These are sourced negotiating positions, not a priced offering; Nscale declined comment. Status remains filed. [B](https://www.businesstimes.com.sg/companies-markets/ai-cloud-firm-nscale-faces-investor-pushback-over-ipo-valuation) October 20th: Reuters reports talks with Anthropic for a second Monarch building in a prospective agreement worth about $45B. The negotiation is separate from the earlier campus contract; no new signed agreement or IPO pricing is established. [B](https://www.aol.com/articles/nscale-anthropic-talks-45-billion-193754000.html) Weekly synthesis: compared with Solidigm, this remains a valuation negotiation within a filed offering. No final pricing or withdrawal established. November 1st monthly review: no price range, roadshow or pricing found through October; the last dated reporting is the October 19th valuation pushback. Filed, not priced. November 8th weekly: the FT reported Microsoft declined a 700-megawatt second-phase option at Barstow, Texas, and Bloomberg reported Microsoft tightening oversight of Nscale, Nebius and IREN projects and debating new neocloud deals. Nscale says the $103 billion of contract value in its filing is unaffected. Filed, not priced. [B](https://www.ft.com/content/668e6ef8-88a1-4bef-8664-8f495baf09f8) November 12th amended registration: preliminary nine-month net loss of $1.69 billion on $307.5 million of revenue, debt estimated at $5.52 billion at September 30th, and "We will be required to obtain additional financing to meet our forecast obligations in 2027," reversing September's 12-month runway statement. No binding financing for the 460-megawatt West Virginia build behind the $44.6 billion Anthropic contract, where the first building alone needs $9 billion to $9.5 billion more. The $3.36 billion Third Point-led convertible, $1 billion from Nvidia, is due to close around November 16th. November 15th weekly: the same week Anthropic's filing made Lambda's contract terminable, so two of Anthropic's builders showed financing strain at once. Filed, not priced. [B](https://www.theglobeandmail.com/investing/article-neoclouds-ai-stocks-slide-again-as-nvidia-turmoil-rattles-sector/)
Autonomous composite-part factories connecting engineering, manufacturing and inspection.
Founders
Alex Niehaus (Atoms automation; AI agent development at Arcane), Edmundo Sanz-Gadea (autonomous drones; large-scale AI training; robotic manipulation research at ETH Zurich)
The deal
Company investment amount, instrument and equity terms are not disclosed on the EF profile.
The recent capability bet is an automated loop spanning design through inspection, where each manufactured part can provide feedback to the next design. The founders combine agent workflows with physical autonomy. The objection is factory execution: better planning software cannot by itself eliminate material variation, tooling expense or qualification delays. The profile identifies a 2026 founding year but does not name its cohort or prove production-scale autonomy.
Autonomous robots for the data centers and facilities that keep frontier AI compute running: inspection, maintenance and physical operations for hyperscale campuses.
Founders
Chen Peng (CEO; investment analyst at BlackRock, physics at UPenn, research at the GRASP lab), Yim Ying (full-stack robotics engineer, manipulation and manufacturing, MechE at UPenn), Liam Zhong (CTO; GRASP lab research in vision, control and reinforcement learning, CVPR author and reviewer)
The deal
Not stated on the profile; YC's standard is $500K, as $125K post-money for 7% plus $375K uncapped MFN.
The AI data center buildout is the largest capex wave in tech and its facilities face a physical-operations labor gap nobody was building robots for two years ago, because campuses at this density barely existed. The team pairs a GRASP-lab manipulation and RL core with a founder who priced infrastructure assets at BlackRock, an unusual fit for a market where the customer is a financier as much as an operator; six people and four open hardware roles suggest deployment intent. The case against: operators are extremely conservative about anything mobile near live racks, the environment favors fixed automation and DCIM software, and a vendor selling into data-center opex is levered to the exact buildout it is betting on.
Multi-silicon inference cloud splitting the phases of one workload across heterogeneous chips raised a $300M Series B led by Andreessen Horowitz at a $3B valuation, with Arm and Microsoft's M12 as new backers. Total raised $392M.
The thesis is that the Nvidia-only inference stack overpays once workloads decompose, and the cap table is the tell: Arm and Microsoft both win in a world where inference routes across many chips. Every dollar the token-price collapse squeezes from raw inference makes the routing layer above it more valuable; whether that layer stays a company or becomes a feature of every cloud is the open question.
Vertically integrated AI infrastructure builder and cloud serving OpenAI, Microsoft and Meta. Officially closed an initial $3.9B Series F on September 17th at a $30.9B post-money valuation, co-led by Atreides, Mubadala and Valor with Nvidia, Founders Fund, GIC, QIA, Radical Ventures and TPG; first reported at more than $3B in early September.
Tripled its valuation in ten months, claims over $140B in company-reported total contracted value and 6+ GW gross contracted capacity, and per TechCrunch has met investment bankers about an eventual IPO. The Jane Street contract at $13B over five years is the diversification argument against lab concentration; the truck-transportable Spark modules are the siting play.
AI-driven analog and mixed-signal chip design, compressing semiconductor design timelines from months to weeks.
Founders
Alex Gilman (CEO; 20+ years designing high-performance chips, fiber-optic comms at Maxim, DNA-sequencing chips at Roche), Unnamed CTO (has taken semiconductor products from definition to high-volume production, per the cohort page)
The deal
Not stated on the cohort page; Techstars' published standard program investment is $220K.
Analog design is the corner of chip design AI has barely touched: digital RTL has decades of Synopsys and Cadence automation while analog remains artisanal, and reliable constraint-satisfying generation over schematics only became plausible with 2025-26 models. A founder with twenty years across Maxim and Roche knows which sign-off steps actually bottleneck a tape-out. The case against: analog EDA is a small, brutally conservative market where the incumbents are shipping their own copilots, and fabless customers rarely bet a tape-out on a seed-stage vendor, so a great demo can still die waiting for one design win. The company site link comes from the Techstars page and answered with a bot block when fetched.
Foundation models that learn physics end to end, replacing slow engineering simulations: the pitch is results 'in seconds, at a fraction of the cost,' with a claimed 1000x speedup over conventional solvers.
Founders
Daniel McNeil (co-founder and CEO; age 18 per the YC profile), Krish Kapadia (co-founder and CTO; age 19 per the YC profile)
The deal
YC standard: $500K, as $125K post-money for 7% plus $375K uncapped MFN.
Neural surrogates for physics only became credible at 2025-26 training scale, so the thing could not have been built two years ago, and a16z's Machine Age fund and Marengo's design-stage bet say the buildout's bottleneck is exactly the engineering cycle this attacks. The case against is heavy and specific: two teenage founders with no stated CFD or industry background, against Ansys, Cadence and PhysicsX, in a market where certification bodies buy error bars rather than speed, and the profile page states no customers, benchmarks or publications. The YC page lists ages and titles and nothing else about their background.
AI voice agents for Medicare patient intake: outreach, intake assessments that pre-read the chart, documentation and follow-up, producing clinician-ready summaries under HIPAA with clinician oversight.
Founders
David Clark Moore (ex-Meta), Adam Davison (ex-Meta; ran Reels monetization from $0 to $10B+, per the tracker)
The deal
Speedrun standard: up to $1M, as $500K for 10% via SAFE plus follow-on reserved to $1M; Rhema's specific terms are not disclosed.
Voice agents good enough for regulated clinical intake are a post-2024 capability, and the founder thesis is the interesting part: consumer-engagement operators from Meta's Reels applying scaled outreach machinery to the population that answers the phone, Medicare patients. The case against is that Medicare voice AI is among the most crowded categories in health tech, with Hippocratic AI, Infinitus and Assort ahead on clinical distribution, sales run through risk-averse health systems, and one hallucinated intake answer is a regulatory event; ad-monetization DNA is not clinical-workflow DNA. Founder names are per Frontrun VC's SR007 tracker; the company's own site names nobody.
Hardware and firmware simulation built so AI coding agents can develop embedded firmware: simulated targets give agents (and humans) visibility into hardware systems that previously required physical hardware-in-the-loop rigs.
Founders
Seungmin Hong (CEO; firmware at Tesla, Simbe Robotics and Rapyuta Robotics, ML at Ericsson; Waterloo computer engineering), Ahnaf Shahriar (CTO; embedded software on Neuralink's implant and charger, AI hardware on Tesla Dojo and Autopilot; Waterloo computer engineering)
The deal
YC standard: $500K, as $125K post-money for 7% plus $375K uncapped MFN.
The product exists because coding agents now write firmware and cannot run it against real hardware, a problem that did not exist two years ago, and the founders' Neuralink and Dojo firmware background is exactly the non-obvious knowledge the thesis needs. The case against: QEMU, Renode and Wokwi already simulate microcontrollers, an incumbent could bolt on the agent framing, and the moat question is peripheral and sensor fidelity, a brutal per-chip-family grind on which the silicon vendors hold the proprietary models.
AI chipmaker began STAR Market trading on September 11th; first-day close 397 yuan versus a 142.18-yuan offer.
The deal
Tencent will hold 17.95% after the IPO and accounted for 83.79% of 2025 revenue as the largest end customer, according to the filing reported by Reuters.
September 11th 2026: listing completed. Reuters reports a 410-yuan opening and a close around 397; China Business Journal reports an exact 397-yuan close. An intraday high is not the closing price. Customer concentration remains a separate business risk.
Agentic data observability: agent 'Scouts' monitor production tables, investigate failed data-quality checks through column-level lineage across warehouses, dbt, Spark, Airflow and BI tools, and prepare validated patches routed to the owning team. The cohort listing pitches the larger ambition of a governable, agent-consumable data warehouse.
Founders
Jacob Kieser (ex Palantir and Uber), Victor Sunderland (ex OpenAI)
The deal
Speedrun standard: up to $1M, as $500K for 10% via SAFE plus follow-on reserved to $1M; Embrasure's specific terms are not disclosed.
Data infrastructure that assumes agents rather than analysts are the primary consumers of enterprise data is a genuinely post-2024 thesis, and an ex-Palantir plus ex-OpenAI pairing is the right team to hold it. The case against is crowding and drift: observability is a consolidating category where Monte Carlo and Anomalo are entrenched and Datadog already bought Metaplane, agentic remediation is on every incumbent roadmap, and the gap between the site's modest observability pitch and the cohort list's autonomous-warehouse pitch says the story is still being found. Founder names and backgrounds are per Frontrun VC's SR007 tracker; the company's own site carries no team page.
The image-and-media model lab raised a $76M Series B on August 25th from Universal Music Group, Sony Music, Warner Music and Electronic Arts: the music industry and a game publisher buying a seat inside the generative company they might otherwise be suing.
The investor list is the story: all three major labels funding a generative-media lab reads as the licensing detente arriving by cap table rather than by courtroom, and EA's presence points the same tooling at game assets.
An insurance layer for AI: technically assesses AI systems, certifies them against recognized standards, and places liability coverage so enterprise AI deals stalled on liability concerns can close.
Founders
Casey Tewey (co-founder and CEO; MIT Class of 2025), Michael Wong (co-founder and CTO; MIT Class of 2025)
The deal
Speedrun standard: up to $1M, as $500K for 10% via SAFE plus $500K reserved for the next round within 18 months; Keythorn's specific terms are not disclosed.
Agent liability insurance could not have existed two years ago because there were no agents transacting in the economy to underwrite, and whoever writes the first actuarial rulebook for agent risk gets a data moat that compounds. The case against is that two fresh MIT graduates with no stated insurance background are entering the most regulated, balance-sheet-heavy industry there is, cannot carry risk themselves, and Lloyd's syndicates, Armilla and Munich Re are already circling the same space. Batch membership per frontrun.vc's SR007 tracker; the company's own site says it is backed by a16z speedrun.
A wrist-worn surface-EMG bracelet that captures force data during human demonstrations for robot training, on the thesis that force cannot be reconstructed from kinematics: identical hand poses can carry wildly different force, so vision-only demonstration data is missing a modality robot foundation models need.
Founders
Julian Olschwang (solo founder; ECE at Carnegie Mellon, built GANs for military robotic data generation at the UCLA Biomechatronics Lab)
The deal
YC standard: $500K, as $125K post-money for 7% plus $375K uncapped MFN.
Robot-learning data collection is the pick-and-shovel trade of 2026, and force is the acknowledged missing modality: two years ago there was no VLA model to consume this data, so there was no company to build. The insight that force must be captured at demonstration time is non-obvious and specific. The case against is heavy: a solo founder, a team of one, selling hardware, in a sensing modality (surface EMG) that is notoriously noisy and person-dependent, to labs that may just instrument their grippers directly.
An AI system for neurodiversity assessments that organizes case documents and drafts clinician-reviewed ADHD and autism diagnostic reports, cutting report-writing from 5-7 hours to under 45 minutes while keeping clinical oversight.
Founders
Dr Jan Cosgrave (clinical psychologist, PhD from Oxford, Fulbright scholar at Penn, NHS and academic background in ADHD and autism), Nadezhda Prodanova (MSc computer science from KIT and ETH Zurich, medical-grade AI systems and clinical data pipelines)
The deal
Techstars standard: $20K for 6% common plus a $100K convertible note; selected for the Techstars Healthcare Accelerator from more than 4,000 applicants.
Diagnostic report-writing is the unglamorous middle of a real bottleneck: ADHD and autism assessment waitlists run to years in the UK, and the clinical hours lost to documentation are a direct cause. The founder pairing is the story here, an Oxford clinical psychologist and an ETH machine-learning engineer, which is the combination this kind of clinician-governed tool actually needs. The case against is regulatory and liability exposure: a report a clinician signs but an AI drafted sits in exactly the zone medical-device and malpractice frameworks are least settled on, and the 45-minute claim means little if the review to catch a wrong diagnosis takes the time back.
An AI-native engineering firm that designs data centers, delivering site due diligence through front-end engineering and permitting in about half the usual time and cost, using internal physics-grounded AI design tools run by professional engineers.
Founders
Emil Ares (CEO; Cambridge-trained physicist, worked on magnets at the 1.4GW MagLab, former UK Space Agency, published on AI for complex-systems engineering), Gad Marconi (CTO; data-center infrastructure, deployed systems at the 25MW Flumenthal hydro plant in Switzerland, ex-AI engineer in nuclear and energy, TU Delft)
The deal
YC standard: $500K, as $125K post-money for 7% plus $375K uncapped MFN.
The bet everyone is making, a16z's $1.1B Machine Age fund included, is that the buildout's bottleneck is physical and its engineering is slow. Marengo attacks the design and permitting stage directly, where a data center loses months before a shovel moves. The strongest case against is that a five-person firm selling to hyperscalers and their EPC contractors is entering a market that rewards track record over speed, and $4M in design partnerships is interest, not a moat. But if AI genuinely halves the front-end engineering cycle, that compresses the scarcest thing in the whole buildout: time to power-on.
The digital and AI arm of Mukesh Ambani's Reliance Industries, listed on the NSE and BSE on October 28th 2026 after India's largest IPO, priced at the top of a ₹1,065 to ₹1,119 band. Backed by Meta and Google.
The deal
SEBI issued final observations on August 28th; the offering is up to ~270 million fresh shares (about 2.9% of post-issue equity), with proceeds largely repaying Jio's borrowings.
Not an AI pure-play, but the distribution layer that decides whether AI reaches the next several hundred million users in the market OpenAI and Meta are both staffing to win. The listing also proves the Asian IPO window is open, which is the same window Moonshot and others are eyeing, so it is a read on the pipeline as much as on Jio. October 1st review: the issuer IPO page supplies draft offer documents. It does not establish final pricing or a completed listing; retain filed status. [A] November 1st monthly review: the issue ran October 21st to 23rd and listed October 28th at ₹1,358, a 21.36% premium to the ₹1,119 issue price; outlets differ on the close (₹1,417.5 in a Reuters relay, ₹1,419 in PTI copy), so this record keeps the listing price only. The open window the August note expected held for a domestic issuer with a captive retail base; it says little about the US neocloud queue. [B](https://yourstory.com/2026/10/jio-platforms-nse-strong-stock-market-debut)[B](https://www.outlookindia.com/business/grand-debut-for-jio-platforms-shares-up-nearly-23-pc-becomes-2nd-most-valuable-indian-firm-by-mcap)
A London AI lab building an 'AI scientist' agent, Faraday, that reproduces published research. On its own 310-task Replica benchmark, Faraday, built on a 27B Qwen base with GPT-5.5 Codex for coding, outscored Claude Opus 4.8 and GPT-5.5 at replicating results.
Founders
Edward Hughes (co-founder; ex-DeepMind), Tantum Collins (ex-DeepMind), Louis Kirsch (ex-DeepMind), Kaloyan Aleksiev (ex-Reka AI and Microsoft)
The deal
$50M seed led by Index Ventures, with Radical Ventures participating; team about a dozen, targeting 20-25 by year-end.
The interesting claim is that a small open-base agent with the right scaffolding beats the frontier closed models at a specific scientific task, reproducing papers, which is the same wager Metis makes on coding. The case against is that a self-reported win on a self-built benchmark is marketing until an outside group runs it.
South Korean multimodal consumer AI app maker; raised a $72M Series C at a valuation Bloomberg put at $870M (1.2 trillion won), the first Korean AI service startup to cross the trillion-won mark. Total funding now $167M.
New investors Coreline Ventures and Eugene Asset Management joined Goodwater Capital, Antler and the state-backed Korea Development Bank. The company says the money funds East Asia expansion, GPU capacity and its NLP engineering team, at a valuation more than five times its total funding raised.
Tamper-evident receipts for AI systems: a runtime layer that signs a record of what model ran, which guardrails executed, what consent was captured and what was decided, generated in under 50 milliseconds.
Founders
Joe Braidwood, CEO (former CMO of SwiftKey, acquired by Microsoft), Dr. Jennifer Shannon (FDA medical-device regulatory background), Rohit Tatachar
The deal
$575K pre-seed from Geoff Ralston's Safe Artificial Intelligence Fund, Mighty Capital, Sourdough Ventures and the AI2 Incubator; a seed round is targeted for later this year.
Every post-incident report this month, OpenAI's included, turns on an unverifiable claim that the safety controls ran. GLACIS sells the receipt: cryptographic attestation per AI decision, starting in healthcare where a founder's FDA background reads as the actual moat. This could not exist before agents made per-decision liability real. The case against: compliance software sells when enforcement bites, and the administration is dismantling the state AI laws that would bite; the EU's Article 50 regime is the buyer that has to show up, and $575K is a small bet that it does.
AI accelerators built on a hybrid memory architecture, SRAM's near-memory compute combined with HBM's density, aimed at frontier-scale, long-context LLM inference.
Founders
Ganesh Venkat (Ganesh Venkataramanan, led Tesla's Dojo supercomputer program), Bill Chang, CTO (chief system engineer on Dojo), Ben Floering (silicon, SaaS and data centers at IBM, AMD and Tesla)
The deal
Terms not disclosed on the SPC profile; South Park Commons typically writes first checks in the $1M to $10M range.
The core Dojo team, out of Tesla and betting that the memory wall, not compute, decides inference economics: SRAM designs like Cerebras and Groq win on speed and lose on model size, HBM designs win on capacity and pay the scarcity premium Nvidia just surrendered margin to. A hybrid is the obvious move nobody has shipped. The case against: chip startups die in the two years between architecture and silicon, the HBM they still need is the market's scarcest input, and unconfirmed reporting has speculated Tesla itself as an acquirer, which would make this a reunion rather than a company.
AI trust, safety and security company (formerly ActiveFence): red-teams frontier models before release and runs runtime guardrails in production, drawing on a proprietary intelligence database built over nearly a decade tracking fraud, extremism and disinformation online. Raised a $140M round led by Apax Digital, total funding $280M.
CEO Noam Schwartz: "We are very quickly democratizing capabilities before we've democratized the defenses. This round is about closing that gap." The round closed the same week Cursor was used to breach seven companies and 116 companies asked frontier labs for exactly the kind of incident-time model access a red-teaming vendor is built to provide.
A two-person robotics research lab whose product, Instance, is a success detector for robot training: it watches a robot attempt a task and judges whether it worked, replacing the human sitting behind every evaluation run.
Founders
Claire Mao (CEO; MIT math and CS, ex-NASA JPL planetary-atmosphere simulation, MIT Media Lab aerospace research), Lucy Cai (CTO; MIT CS MEng in the Learning and Intelligent Systems lab; SpaceX satellite software, Amazon LLM testing, Blackrock Neurotech brain-computer interfaces)
The deal
YC standard: $500K, as $125K post-money for 7% plus $375K uncapped MFN.
Robot fleets scale by training volume, and evaluation is the part still done by a person watching. A success detector only became buildable when vision-language models got cheap enough to run per-episode, and the founders benchmark theirs against 10,000+ human-labeled episodes across seven robot platforms, claiming better accuracy than a frontier model at a fraction of the latency. The case against: success detection reads like a feature every robotics lab builds in-house, and the labs with the biggest fleets, the natural customers, are the likeliest to do exactly that.
AI-powered powerline inspection and vegetation management for utilities, using computer vision over standard cameras instead of LiDAR flights, at a claimed 75% lower cost; six utilities have signed LOIs or MSAs covering 270,000 miles of line.
Founders
Yann (ex-McKinsey associate partner in energy, per Antler's profile, which gives no surname), Nicolas (ETH Zurich robotics, Cambridge multi-agent research, previously YC-backed, per Antler's profile, which gives no surname)
The deal
Antler US standard: $250K at $2.75M post-money for about 9%.
Vegetation contact is a leading cause of wildfire-linked outages, and inspection is bought in LiDAR flights priced per mile. The bet is that current vision models over commodity cameras finally match purpose-built sensors, which was not true two years ago, and 270,000 miles of scoped interest from six utilities says buyers want the cheaper answer. The case against: Overstory already sells satellite-based vegetation intelligence to utilities, GE Vernova bundles it into asset software, and utility procurement cycles can outlast a seed runway.
Middle-mile autonomous trucking company running daily driverless routes for Fortune 50 retailers across Texas, Arizona, Arkansas and Canada; raised a $200M Series D led by Qatar Investment Authority and Koch Disruptive Technologies, total funding near $500M.
The autonomy segment where the economics closed first keeps compounding quietly: 85,000 fully driverless orders, more than $600M in contracted revenue, and sovereign-wealth money arriving at Series D rather than pre-revenue. Fixed routes between endpoints the customer owns is the template the robotaxi companies still cannot match on cost per approval.
Washington DC software company whose platform turns AI data centers into grid-responsive loads, throttling draw at peak and expanding when capacity is spare; raised a $150M Series A at $1.05B co-led by Energize Capital and DCVC.
A unicorn Series A for demand-response software says the binding constraint has moved to interconnection: the company claims flexible operation could unlock 100GW on the existing US grid, and it has five commercial demonstrations including a full data center flexing through California peak strain. The week's Texas queue freeze and PJMcurtailment fight are the demand signal, priced.
Sources internal company data for frontier-lab training from bankruptcy courts: codebases, project files and payment records, de-identified and packaged.
Founders
Ian Lee (Meridian AI training-data infrastructure; Scale AI technical advisor), Samuel Hahn (Magier AI go-to-market; Mainstreet), Sudhish Swain (Clean Sweep Group head of software engineering)
The deal
YC standard: $500K, as $125K post-money for 7% plus $375K uncapped MFN.
Google paying $10 million for Spirit Airlines' emails established that bankruptcy estates hold the scarcest training input, authentic workplace text, and Petrarch is the first company built to make that trade repeatable. The case against is the legal exposure: privacy obligations survive a data-asset sale, courts are scrutinizing them harder, and one bad de-identification could end the company. Buyer concentration in a handful of labs cuts the other way.
Agentic-commerce platform that makes merchant product catalogs machine-readable and transactable, so AI agents in ChatGPT, Gemini and Perplexity can discover and buy products, not just mention them.
Founders
Ashish Piplani (Grab, built GrabUnlimited; Lazada; Flipkart), Evgeny Oleynik (co-founder and CTO of 12GO, multi-country travel platform)
The deal
Antler US standard: $250K at $2.75M post-money for about 9%; AirShelf's own terms not disclosed, and its Singapore base may mean different program terms.
Selling through assistants only became a channel with this year's agentic-checkout push, and the founders' Grab, Lazada and Flipkart years are exactly where you learn how messy merchant catalog data gets at scale. They publish a public Product-Truth Benchmark rather than only marketing claims. The case against is existential platform risk: if OpenAI and Google ingest catalogs directly from Shopify and their own merchant programs, independent middleware gets commoditized overnight, and the crowded AI-visibility space next door is a warning.
XPeng's humanoid unit raised over $900 million at a $6.3 billion post-money valuation, the largest single private round in China's embodied-AI industry: IDG Capital led, Gaorong joined, Tencent and Alibaba came in as strategics. IRON mass production is slated before year-end.
The carve-out structure matters more than the number: a listed EV maker spinning its robot unit into a separately financed entity with two hyperscale strategics is the standard pre-listing shape, and the stated capacity target, a thousand IRON units a month, would exceed Unitree's entire 2025 humanoid volume within half a year of hitting it.
An AI knowledge layer for mechanical design: turns every CAD step into living documentation, then automates the engineering paperwork around it, ECO drafts, design reviews, drawings, DFM feedback and traceability records.
Founders
Kunal Jain (co-CEO; mechanical and AI engineering background at Boeing and Rolls-Royce per Antler's profile), Arjun (co-CEO; ex-Rolls-Royce and Boeing per Antler's profile, which gives no surname)
The deal
Antler US standard: $250K at $2.75M post-money for about 9%; Tandem subsequently closed a $4.7M seed with angels including the former CEOs of Autodesk and SolidWorks.
Hardware engineering is where coding-agent economics have not arrived, because the artifact is a CAD tree plus a paper trail regulators actually read. Capturing design intent as it happens only became possible when models could watch CAD operations and write the ECO, and the founders lived the paperwork at Rolls-Royce and Boeing. The case against: Autodesk and PTC own the file formats and could ship this as a feature, and the angel list, two ex-CEOs of the incumbents, reads as both endorsement and acquisition foreshadowing. Named customers span diagnostics, satellites and pediatric heart devices.
Model Processing Units: custom chips that bake LLM weights directly into silicon, removing the memory-to-processor shuttle that makes GPU inference slow and hot. Claims 20,000+ tokens a second and 63x more intelligence per watt, entirely local.
Founders
Thomas Lanning (CEO; Oxford physics PhD, machine learning for Higgs boson detection), Niki Kotecha (CTO; AI PhD from Imperial College London, published researcher)
The deal
YC standard: $500K, as $125K post-money for 7% plus $375K uncapped MFN.
This could not have existed two years ago: burning weights into silicon only makes sense once open models are good enough to freeze. That is also the strongest case against. A taped-out model can never be updated, and the frontier currently obsoletes itself every few months, so the bet is that edge, sovereignty and power-constrained niches will pay for a frozen model that is theirs. Two physics PhDs from Oxford and Imperial choosing London hardware over a lab job is itself a signal about where UK AI talent thinks the open ground is.
Simulated publics for policy: a decision-intelligence platform that lets governments test how a real population reacts to a policy before it launches, built on generative-agent behavioral simulation.
Founders
Shirin Abrishami Kashani (CEO; Oxford MS, Stanford PhD as Knight-Hennessy Scholar, former OECD policy analyst advising governments), Keshav Sivakumar (CTO; second-time founder, Cambridge CS, ex-Palantir), Jonne Kamphorst (Chief Scientific Officer; former Stanford postdoc leading research in the lab that pioneered human behavioral simulation; LSE, Oxford and EUI, Fulbright Scholar)
The deal
Speedrun invests up to $1M plus $10M+ in credits per its program terms; company-specific terms not disclosed.
The CSO ran research inside the Stanford lab that invented generative-agent simulation, the clearest worked-on-it-inside-the-lab credential in the cohort, paired with a CEO who advised the actual buyer at the OECD. The same technique Simile is now selling to Fortune 100s at a $2B valuation, pointed at governments instead. The case against is fundamental: there is no independent evidence yet that simulated publics predict real ones, LLM personas carry sycophancy and training-distribution bias, and a wrong-but-confident policy simulation is worse than none, all before government procurement cycles slow everything down.
Wafer-scale AI chip maker, listed on Nasdaq as CBRS in May at $185 a share. Its insider lockup now comes apart in stages.
Early-release tranches fall in August, September and October ahead of the roughly November expiry of the standard 180-day ban, a supply overhang arriving while retail enthusiasm from the CS-4 launch is still warm. The first AI-chip IPO of the cycle is about to show what insiders do when they are allowed to sell, and every later listing in the pipeline will be priced on the answer.
AI-native materials discovery powered by unpublished experimental data: mining the failed runs and shelved results sitting inside industrial labs that no web-scraped model can reach.
Founders
Ian Naccarella (CEO; MIT fellow commercializing lab IP, strategy manager at battery-materials unicorn Sila Nanotechnologies, ex-BCG; Stanford ChemE, Harvard MBA), Eric Riesel (CTO; MIT PhD in inorganic chemistry, built ML systems that analyze messy materials data; Columbia chemistry), Yankang Yang (COO; former BCG principal who led the firm's internal AI program to 30,000+ users; Harvard CS and EE)
The deal
YC standard: $500K, as $125K post-money for 7% plus $375K uncapped MFN.
The wedge is not another generative-materials model, it is the data nobody else can get: labs' unpublished and failed experiments, aimed at regulation-driven markets like PFAS replacement where buyers face deadlines. The case against is a graveyard: AI materials discovery has burned investors before, DeepMind's GNoME showed candidate generation is not the bottleneck, synthesis and qualification still take a decade, and the unpublished-data moat depends on labs granting rights to data they may consider proprietary or embarrassing, a heavy business-development motion for a three-person team.
Orbital AI data centers: satellites that run inference in space on solar power, with two 8kW compute satellites planned for 2027 rideshares and a larger Starcloud-3 sized for Starship.
A $250 million Series A extension at a $2.3 billion valuation, about $420 million total, with Nvidia, Cisco, Benchmark, EQT and NFX aboard. CEO Philip Johnston is raising ahead of need because launch capacity is tightening; the constraint story has moved from power to permits to rockets in six months, and Starcloud is the second orbital-compute raise in two days after Muon Space's $250 million.
AI-native accounting ERP built to replace NetSuite-class systems, automating ledger work with agents; $100M Series C at a $1B valuation led by ICONIQ.
Third round in a year, closed in under 48 hours from a company that was not raising, on revenue that doubled last quarter, public-company customers, and an EY alliance that brings its automated ledger into audit work. Accounting keeps converting to agents faster than flashier categories because the work is rule-bound, auditable and understaffed.
The identity and communication layer for AI agents: each agent gets a persistent email inbox, phone number, iMessage access, webhook address and an encrypted credential vault with TOTP support.
Founders
Ray Liao (CEO; MIT computer science PhD; co-founded Empallo, an AI cardiovascular virtual clinic), Alex Wilcox (CTO; MS Northeastern; ML engineering and computer networking research), Dima Vremenko (built ML pipelines at Harvard Medical School; BS Northeastern in biochemistry and computer science)
The deal
YC standard: $500K, as $125K post-money for 7% plus $375K uncapped MFN.
Agents that act like employees need what employees have: an address, a phone number, credentials, and a way to reach each other, and today that means borrowing a human's accounts against most terms of service. Two years ago there was nothing to identify. The case against is sharp and worth stating: an identity layer for agents is also infrastructure for spam and fraud at machine scale, and the same vault that holds an agent's 2FA tokens is a single point of catastrophic compromise. Whether platforms treat agent identities as customers or as a botnet decides the company.
Agentic electronics design: turns product requirements into complete schematics using verified purchasable components, with datasheet-grounded review, BOM optimization and export to KiCad, Altium and EasyEDA.
Founders
Radu Berdan (founder; PhD, Imperial College London, in memristive systems; founded ArC Instruments; neuromorphic computing and AI accelerator research in Japan)
The deal
Techstars Tokyo terms, per the cohort announcement: $220,000 in capital across the 13-week program.
Hardware design is the layer of the AI boom still done by hand: schematics drawn in tools whose workflow predates the web, against supply-chain data that goes stale weekly. A founder who spent a career in semiconductor research is betting language models can draft and verify circuits the way coding agents draft software, grounded in datasheets rather than vibes. The case against: Cadence, Synopsys and Altium own the file formats and the installed base, Flux has a two-year head start on AI-native PCB design, and electronics engineers punish tools that hallucinate a wrong component harder than software ever does.
Agentic AI systems for trust-and-safety, fraud and risk operations: agents that augment human review teams and add investigation capacity where platforms currently ration it.
Founders
Suhas Manangi (Head of Product for Trust & Safety at Snap; previously Airbnb's AI/ML platform for Trust, plus Amazon, Lyft and Microsoft), Xiuduan Fang (co-founder and CTO; led Trust & Safety engineering at Snap)
The deal
Antler US standard: $250K at $2.75M post-money for about 9%.
The two people who ran Snap's trust-and-safety product and engineering are betting that T&S review, an expense every platform rations, becomes an agent workload, and new strict-liability laws like Minnesota's nudification ban keep raising the price of rationing it. They know where the review queues actually break, which is not public knowledge. The case against: every model provider and every large platform is building safety tooling in-house, and a startup selling picks between those two armies needs the mid-market to care about T&S before regulators force it to.
Kimi maker filed confidentially with HKEX in September; Bloomberg reported October 6th that its final private round closed at about a $50B valuation and the listing targets Q1 2027 with a raise of up to $5B, with Goldman Sachs, CICC and Deutsche Bank on the deal.
Rumor became filing on September 3rd, per Reuters and TechNode: a confidential A1, filed after Moonshot unwound its offshore structure and redomiciled onshore. ARR tripled from $100M in March to $300M in June on the Kimi K3 launch, and the reported hyperscaler revenue-sharing talks from August are now part of the prospectus story. If the listing completes, the first frontier-scale Chinese lab reaches public markets financed substantially by giving its flagship away.
Chinese embodied-intelligence company building the decision and execution layer for industrial and logistics robots, combining LLMs, vision models and reinforcement learning rather than building hardware. In a new round targeting a 20 billion yuan ($3B) valuation.
Founded by Tang Wenbin, previously backed by Alibaba, NIO Capital and Chinese AI lab Zhipu (Z.ai). One of the quieter rounds inside China's robotics boom, funded the same week as Unitree's post-IPO pop and the World Humanoid Robot Games, on the software-over-hardware bet that decision-making, not actuators, is the scarce layer.
Local-first AI routing: directs each request to the cheapest viable option, usually the user's own hardware, with a Rust inference engine the team says decodes up to 60% faster than llama.cpp on Apple Silicon, and claims of up to 80% lower token spend.
Founders
Michael Jeffords (ML and clinical research; founded The HeartCheck Foundation), Charles Muehlberger (edge AI acceleration and multimodal inference), Yavuz Inan
The deal
YC standard: $500K, as $125K post-money for 7% plus $375K uncapped MFN.
The consumer-hardware mirror of the routing thesis Callosum just raised $100M to run in datacenters: laptops and phones now carry usable models, so the cheapest inference for many requests is the machine already on your desk, and a local-only mode doubles as a privacy product for sensitive data. Could not exist two years ago; consumer silicon could not run models worth routing to. The case against: the arbitrage is priced by other people, since every frontier API price cut shrinks the saving, and Apple can absorb the on-device half into the OS whenever it chooses.
London startup building heterogeneous-intelligence infrastructure: software that routes each AI task to the model and chip that runs it cheapest, fastest or most energy-efficiently. Raised a $100M seed led by Atomico with Plural, DCVC and the UK Sovereign AI Fund, that fund's first investment.
Founded 2025 by neuroscientists Danyal Akarca and Jascha Achterberg; about $110M raised in total after a $10.25M Plural pre-seed in February. One of Europe's largest seeds, and a policy statement: Britain's sovereign fund opened with the routing layer rather than a model or a fab. The crowded part of the market: OpenRouter just sold to Stripe for the same thesis one layer up, and Callosum's chip-level version has to beat every inference provider's own scheduler.
Colossal Biosciences spinout building AI that predicts how biological systems change over time, trained on genomic data across species and deep evolutionary history. Raised $20M led by Bob Nelsen at a $3.8B valuation, with Peak 6, NeoGenesis Capital, Builders VC and CAZ Investments participating.
Co-founded by Colossal's Ben Lamm and Harvard geneticist George Church, the same pair behind the woolly-mammoth de-extinction project. "Evolutionary forecasting" pivots the pitch from resurrection to prediction, with early pilots aimed at health and biosecurity. The $3.8B valuation on a $20M round is a bet on the founders' prior company as much as this one's genomic dataset.
Infrastructure for automated ML research: give it a task, a baseline agent or model and an eval, and it orchestrates thousands of experiments across data, fine-tuning, prompts and tools, returning statistically validated improvements.
Founders
Karan Brar (ML at Reducto; head of ML infrastructure at DynamoAI), Thomas Boser (Reducto; senior ML engineer at Discord; founding engineer at Crosswise)
The deal
YC standard: $500K, as $125K post-money for 7% plus $375K uncapped MFN.
Selling the thing frontier labs keep in-house: the experiment loop itself. Their pitch benchmark is two coding agents running 4,188 experiments across 50 B200s in two days and beating the published best on Karpathy's autoresearch benchmark, which is a claim about orchestration, not model quality, and it rhymes with this week's Reconstruction result showing multi-agent selection recovering capability solo models lack. Could not exist two years ago; agents could not run experiment campaigns unattended. The case against: the teams that most need automated ML research are the labs that will never outsource it, and the mid-market buyer may want the improved model, not the machinery, which makes this a services business wearing a platform's clothes.
An evaluation platform for robots that work around people: generates rare but critical scenarios to surface failure modes, quantifies risk, and gives robotics teams a safety case for deployment.
Founders
Kyle Wong, Simo Rachidi, Ding Zhao (Carnegie Mellon professor; NSF CAREER award in robot safety)
The deal
a16z Speedrun standard: up to $1M, with $500K for 10% on a SAFE upfront and up to $500K more in the next round. October 5th: over $12M seed reported; valuation and equity terms undisclosed. [B](https://techcrunch.com/2026/10/05/can-safeworld-convince-people-that-gen-ai-robots-wont-hurt-them/)
The week China listed its first humanoid maker and Xiaomi promised robots on its own assembly lines is the week the missing piece gets obvious: nobody has a standard way to prove a robot is safe around people. SafeWorld sells the rare-scenario evals that a deployment sign-off needs, anchored by a CMU professor whose academic career is exactly this problem. Could not exist two years ago; there were not enough learned policies deployed near humans to evaluate. The case against: the paying market today is a few dozen robotics teams, most treat safety validation as in-house IP, and if simulation-scale scenario generation becomes a standard feature of robot toolchains, the judge gets commoditized before the market arrives. Update, October 5th 2026: SafeWorld announced a seed round of more than $12 million led by Shine Capital and a16z Speedrun, with BoxGroup, the CMU endowment, Innovation Endeavors and SV Angel participating, and a first named partner in Gritt Robotics.
London inference-chip startup founded by Oxford roboticist Walter Goodwin, selling Anthropic roughly $250M of chips that arrive in 2027; closed a Series C of about $700M at roughly $6.5B.
Founders
Walter Goodwin (University of Oxford robotics)
The deal
About $700 million at roughly $6.5 billion (Telegraph, November 1st 2026), led by Lightspeed and Redpoint with Thrive Capital and the UK Sovereign AI Fund; Companies House shows about £577.4 million of Series C shares allotted September 9th to October 23rd 2026 at £707 to £1,452 a share.
Logged as rumored because the round has not closed: Bloomberg, citing people familiar, reports advanced talks at more than six times the $1B valuation Fractile raised at in May with Accel, Founders Fund and Factorial. The Anthropic order is the whole thesis, a frontier lab paying a quarter-billion dollars for unshipped silicon to hedge Nvidia concentration. November 2nd 2026: the round has closed, per the Telegraph and Companies House filings, so the status moves from rumored to raised. The UK's £500 million Sovereign AI Fund came in at a price slightly below the final one; it usually writes £1 million to £10 million checks, and is now exposed to a company that has not shipped in volume.
An AI pair programmer that lives inside the Godot 4 game engine: it reads and edits the live scene tree, writes GDScript and C#, runs tests, and debugs in-editor, rather than working from files alone.
Founders
Ena So (named in both catalog source sets; earlier profile describes a trading background), Delano Lourenco (named in both source sets; earlier profile describes developer-tools experience), Dan Ginovker (named in the October 4th source set; cofounder attribution is not reconciled with the earlier profile)
The deal
Techstars Tokyo terms, per the cohort announcement: $220,000 in capital across the 13-week program.
Consolidated with the October 4th Zivash entry: both describe the same Godot agent at ziva.sh in Techstars Tokyo 2026. The older profile names two founders and the later source set names three; that attribution remains unresolved. Company and Product Hunt claims of users and game-jam wins are not independently verified. The weekly thesis is that access to live editor state can make verification more useful than a text-only coding interface. The commercial test is retained paying users and reduced review work; neither is established here. [A](https://ziva.sh)[A](https://www.techstars.com/newsroom/techstars-tokyo-accelerator-announces-class-of-2026)
Automated evaluation for robot policies: a success detector that judges whether a robot completed a task, breaks it into subtasks, and resets the scene so the next rollout runs with no human in the loop.
Founders
Claire Mao (MIT Math + CS; prev NASA JPL, BCG, MIT Media Lab), Lucy Cai (MIT CS + AI; robotics at MIT LIS Lab; prev SpaceX, AWS)
The deal
YC standard: $500K, as $125K post-money for 7% plus $375K uncapped MFN.
The bottleneck in robot learning is not compute, it is labeling whether a rollout succeeded and resetting the workspace to try again, and both are done by hand today. Instance claims a verifier benchmarked on 10,000-plus held-out human-labeled episodes across seven robot platforms that beats Claude Opus on accuracy at lower latency, which is a bet that a narrow judge outruns a frontier generalist on this one task. The case against is real: the paying market is a few dozen robot-foundation-model teams, most treat evals as core IP, and improving frontier VLMs could commoditize the judge. It could not have been built two years ago because there were not enough VLA policies to evaluate, nor VLMs good enough to grade them.
Nigerian autonomous-defense company: drones, interceptors, sentry towers and ground vehicles guarding critical infrastructure across the Global South.
Founders
Nathan Nwachuku, Maxwell Maduka
Closed Africa's largest seed round at $52 million, with 8VC, Silent Ventures and SV Angel aboard and a London office opening. The argument is price and sovereignty: infrastructure operators who cannot buy American or Israeli systems at their price points get locally built autonomy instead. Claims line of sight to $100 million in contracts this year, which is a company statement, not an audited one.
The bet is that agents are only as good as their context, and the richest context is everything you did all day. Two years ago this was a privacy scandal waiting to happen; what changed is that agents can now consume the recording, and local-first plus open source is the trust posture Microsoft Recall lacked. The case against is exactly Recall: the OS vendors own the screen, can bundle capture for free, and the social backlash against always-on recording has killed better-funded attempts.
SoftBank-backed developer building OpenAI's 10GW Ohio campus publicly filed its Form S-1 on September 1st for a Nasdaq listing under ticker SBE: H1 2026 revenue of $139M against a $3.2B net loss, and a claimed contracted backlog near $439B.
Rumor became filing on September 1st, with J.P. Morgan, Goldman Sachs, Morgan Stanley, Citigroup and Mizuho leading and a reported $5-7B raise target. The S-1 discloses the circular plumbing: OpenAI holds warrants worth about $5.5B as of June 30th for anchoring the Ohio lease, and Nvidia invests $1.5B with credit support conditioned on the campus exclusively hosting Nvidia compute. September 8th check: the September 4th S-1/A is an exhibits-only amendment. It does not establish IPO pricing; status remains filed. September 23rd check: Investing.com cites an unnamed source saying preparations remain on plan, without a date, and relays a New York Times report of buyer resistance and a later timetable. Timing is contested. Neither account establishes pricing; status remains filed. September 27th review: Reuters reported September 24th that the IPO was delayed. No withdrawal or pricing is established by that report. [B] November 1st monthly review: no price range or pricing found through October.
Three childhood friends, one out of xAI's agent team, applying agents to the least glamorous paperwork in a $450 billion industry where quotes still move by fax and phone. The non-obvious knowledge is what breaks in agent deployments at a frontier lab, pointed at an industry with no software incumbency to displace. The case against: AI workforce for X is the most crowded pattern in every accelerator batch this year, and distributors buy from relationships, not from whoever automates the back office first.
AI video platform; raised a $400M Series B at $5.4B led by DST Global, four times its January valuation.
Founded by former Snap AI lead Alex Mashrabov. The company claims $700M annualized revenue in August against $20M a year earlier, with enterprise now the majority of revenue; those are press-release figures, unaudited, and annualizing an August flatters seasonality. Logged because generative video producing a revenue curve this steep, if even directionally true, moves the whole modality from demo to line of business.
AI-workload-aware optical circuit switches for AI data centers.
Founders
Deepankur Thureja (ETH Zurich quantum-optical hardware; ETH Medal for his doctoral thesis), Gabriele Pasquale (EPFL nanophotonics; IBM prize of the Swiss and Austrian Physical Society)
The deal
YC standard: $500K, as $125K post-money for 7% plus $375K uncapped MFN.
The pitch is a control layer that learns a cluster's GPU communication patterns and reconfigures the optical fabric in sync with the workload, keeping data in the optical domain and skipping the electrical conversion that costs power and latency; the claim is a 10,000-GPU cluster performing like 12,500. Google proved optical switching at scale internally and sells it to nobody, which is the opening. The case against: hard photonics on YC-sized capital, hyperscalers build their own, and two physicists with Nature papers and no data-center sales record are selling into the most locked-up supply chain in the industry.
A decision layer that captures why organizations decided things and feeds that intent to humans and AI agents.
Founders
Jono Caseley (analytics and business intelligence at Bain, Publicis Groupe and The Guardian), Andy Lockey (Head of Architecture at Barclays; Accenture, Experian)
The deal
Antler portfolio company; investment terms not disclosed.
The timing argument is real: the moment agents start making decisions, the record of why past decisions were made stops being documentation hygiene and becomes machine-readable context, and decision provenance is a data layer nobody needed until agents did. The case against is the strongest we have printed: decision-capture is the knowledge-management graveyard with a new coat of paint, tools like this die because nobody logs decisions, and if capture can be automated, Microsoft and Atlassian own the meetings it would be captured from.
A retrieval layer with a falsifiable benchmark claim, #1 on Stanford STaRK at 92% against a 58% baseline, and a token-economics wedge, from a founder who shipped Apple's AI action layer to a billion devices. Strongest case against: agent memory and retrieval may be the most crowded lane in AI infrastructure (Mem0, Zep, Letta, LangChain), frontier labs are building memory natively into the models, and the papr.ai homepage now leads with AI apps that run go-to-market work, which reads as positioning churn away from the infra pitch. Techstars names only one founder; the co-founder is described but unnamed on the class page.
AI-native security firm doing automated vulnerability research and product security audits, pairing AI agents with elite human vulnerability researchers.
Founders
Eten Zou (CEO; cybersecurity PhD, Pwn2Own winner, ex-Microsoft), Yuan Tan (CTO; system-security PhD candidate, Black Hat speaker, DARPA-funded program analysis research), Frank Wu (member of r3kapig, ranked #1 hacking team globally in 2025; DEFCON finalist), Xiaochuan Yu (core member of r3kapig; three-time DEFCON finalist; ex-Zellic; $200K+ in Chrome bug bounties; MS, UC San Diego)
The deal
YC standard: $500K, as $125K post-money for 7% plus $375K uncapped MFN.
Four world-class vulnerability researchers, including a Pwn2Own winner and core members of 2025's top-ranked CTF team, pointing agents at the exact problem they have spent careers on, with $400K+ in bounties across browsers, operating systems and the Linux kernel as evidence the loop works. Strongest case against: XBOW got to AI-driven vulnerability discovery first with far more capital, Google's Big Sleep and the frontier labs run the same play in-house, and agents-plus-elite-hackers can decay into pentest-consulting economics instead of a product.
Personalized mRNA cancer vaccines for dogs, designed computationally from each tumor's own mutations.
Founders
Paul Conyngham (Sydney tech entrepreneur; designed the first fully computationally designed mRNA cancer vaccine delivered to a dog, with UNSW scientists, for his own dog Rosie)
The deal
YC standard: $500K, as $125K post-money for 7% plus $375K uncapped MFN.
Sequence a dog's tumor and healthy DNA, find the mutations unique to its cancer, and manufacture a vaccine that teaches the immune system what to attack, with the founder's own dog as the origin case after chemotherapy and immunotherapy failed. Veterinary oncology is where personalized mRNA gets a short regulatory path and fast readouts. Strongest case against: the origin story is one uncontrolled case of a tumor type that behaves unpredictably without treatment, and per-patient manufacturing economics at pet-owner price points are unproven.
Trains companies their own smaller, specialized models from production agent traces plus simulated failures, with an SLA-backed guarantee of 50%+ lower inference cost at equal or better quality.
Founders
Kion Fallah (Staff Research Scientist at Waabi, led mixed-reality simulation for autonomous vehicles; ML PhD, Georgia Tech), Silen Naihin (built AutoGPT to 160K GitHub stars; AI for Science work at DOE; co-founder of Stackwise, YC W24)
The deal
YC standard: $500K, as $125K post-money for 7% plus $375K uncapped MFN.
The Waabi simulation-first playbook pointed at agents: harvest production traces, simulate the failure modes, distill into models the customer owns. Fallah built that muscle doing autonomous-vehicle simulation and Naihin earned his running AutoGPT and agent evals. Strongest case against: frontier token prices collapse roughly yearly, eroding a '50% cheaper' pitch on its own timeline; distillation platforms like OpenPipe and Fireworks crowd the lane; and enterprises may prefer renting frontier quality over owning models that go stale.
Captures and constructs ground-truth 4D reality, space, time and embedded physics, as training data for robotics, world models, autonomous vehicles and AR, via a proprietary multi-sensor volumetric capture studio.
Founders
Chet Ellis, CEO (Harvard; multimodal AI data and large-scale media production), Tyler Raciti, COO (physical-AI executive across robotics, defense and aviation; university lecturer on AI)
The deal
Techstars standard: $220K, as $20K for 5% common plus $200K uncapped MFN SAFE.
World models are the frontier labs' next battleground and every one of them is data-starved on physics: video scraped from the web has no ground truth for mass, force or occlusion. Selling calibrated 4D capture into that shortage is a picks-and-shovels bet that only became a market once world models and humanoid training runs existed to buy it, and Nvidia and Newlab are already attached. The case against: the labs can build capture rigs in-house the moment the data matters enough, synthetic data keeps eating real capture from the other side, and a studio model scales linearly with hardware in a market that wants exponential volume.
Independent AI evaluation firm: $40M Series A led by a16z at a $400M valuation, with 8VC, Pear VC, Bloomberg Beta, HRT Ventures and Next Ladder.
Vals' benchmarks already appear in model cards from OpenAI, Anthropic, Google, Meta and xAI, and its finding that frontier models complete under 52% of real finance-analyst tasks is the kind of number vendor evals never surface. a16z's own framing, Moody's for models, names the prize: auditor of record for AI capability. The bet is that self-reported benchmarks keep losing credibility, and the same week Z.ai shipped scores nobody can verify for two weeks, that looked like a safe assumption.
Open-source tools and independent benchmarks for how well robots actually perform real-world tasks, starting with Inspect Robots, an eval framework covering 40+ vision-language-action models.
Founders
Jay Chooi, CEO (Harvard statistics and CS; UK AI Security Institute; MATS research fellow; top contributor to Inspect Evals), Aris Zhu, CTO (Harvard CS and physics; Amazon AGI Labs; Amazon Robotics; Yondu Robotics, YC W24)
The deal
YC standard: $500K, as $125K post-money for 7% plus $375K uncapped MFN.
The week China's humanoid makers rushed the public markets on self-reported capability claims, a two-person team from the UK AI Security Institute and Amazon's robotics org set up as a public benefit corporation to grade robots independently, and the eval-shop model just got priced by Vals AI's $40M round. Chooi's AISI and Inspect lineage is the non-obvious asset: government-grade eval methodology applied to a field with no Moody's. The case against: benchmarks are a notoriously thin business, the robot makers can starve an unflattering eval of hardware access, and an MIT-licensed PBC has no obvious wedge from credibility to revenue before its capital runs out.
Reported engagement of CITIC Securities for domestic IPO preparation; Reuters reported October 6th commitments above 80 billion yuan for the private round, and Bloomberg put it at $12B or more with Tencent and CATL contributing and an IPO targeted for early 2027. No public exchange application verified.
September 9th 2026: Reuters sources report the CITIC mandate and an aim to begin the process this year. The Flash price cut weakens the earlier pricing-power argument in this record. Adviser selection leaves the offering at the rumored stage until a filing is established. September 25th review: Reuters corrected its September 24th report to identify the planned financing as a private fundraise, not an IPO. This supplies no formal listing application; status remains rumored. October 6th review: reports of a larger private financing do not establish a formal listing application. Status remains rumored. [B](https://www.aol.com/articles/deepseek-set-net-over-12-052459000.html) November 1st monthly review: the private round, not a listing, is what closed. The Information reported nearly $16 billion from about 30 investors, relayed as 105.5 billion yuan, and Bloomberg called it finalized around 100 billion yuan; DeepSeek has not confirmed. Still no exchange application; status remains rumored. [B](https://www.theinformation.com/briefings/deepseek-raises-nearly-16-billion-second-funding-round)
Hardware-agnostic autonomy software that turns robots a nuclear plant already owns into regulation-compliant radiological survey operators, producing the compliance documentation as output.
Founders
Caleb Horan, CEO (robotics PhD, UT Austin; NRC Fellow; deployed multi-robot teams at Los Alamos National Laboratory's Plutonium Division), Janak Panthi, CTO (robotics PhD, UT Austin; 6 years deploying quadrupeds in nuclear facilities; US Army)
The deal
YC standard: $500K, as $125K post-money for 7% plus $375K uncapped MFN.
The founders' claim to non-obvious knowledge is close to unclonable: Horan personally ran multi-robot teams inside Los Alamos' Plutonium Division, the kind of clearance-adjacent experience you cannot hire your way into, and quadruped autonomy reliable enough to run unsupervised regulatory rounds is a post-2024 capability. Software rather than hardware is the right capital-light shape, and the AI-datacenter nuclear buildout is a macro tailwind that did not exist when they started their PhDs. The case against is procurement latency: NRC-adjacent approval can take years per site, incumbents like Gecko Robotics can bolt on a survey module, and both founders are still finishing their degrees, so a two-person pre-revenue company can die of sales-cycle length long before the nuclear renaissance pays.
Recursively self-improving systems, AI that improves AI, aimed at automating knowledge discovery, out of South Park Commons.
Founders
Josh Tobin (led agentic-products research at OpenAI; previously co-founded Gantry; research and teaching at OpenAI and UC Berkeley), Tim Shi (early OpenAI team member; co-founder and CTO of Cresta, shipped a GPT-based product in 2019; former Stanford AI researcher)
The deal
Terms not disclosed on the SPC profile; South Park Commons typically writes first checks in the $1M to $10M range.
The founders' pedigree is the whole thesis: Tobin led agentic-products research inside OpenAI and Shi was an early OpenAI team member who productized GPT at Cresta in 2019, before almost anyone. Two ex-frontier-lab people leaving in 2025 to bet on automated AI research is itself information about what insiders think has become tractable, and this is the could-not-exist-two-years-ago claim at the level of the research frontier rather than the application layer. The case against is brutal: recursive self-improvement is exactly what OpenAI, Anthropic and DeepMind spend tens of billions on internally, the thesis has no obvious intermediate product to sell, and the best realistic outcome may be an acqui-hire back into a lab.
Shenzhen embodied-AI robotics firm reported to have submitted a confidential Hong Kong listing filing, with Huatai Securities and Morgan Stanley as sponsors.
SCMP, citing two sources familiar with the matter, August 5th. A confidential filing reported through sources stays rumored here until a public filing appears. It would join AgiBot and Galbot in the Hong Kong humanoid queue forming behind Unitree's Shanghai debut, which is the pattern worth logging: China's robotics IPO pipeline is now deeper than its AI-model one.
AI workflow platform for industrial inspections: captures live audio, video and notes in the field, autofills the forms, and generates the standards-compliant report.
Founders
Jonathan Low, CEO (Conjecture), Jakob Cassiman, CTO (ML6; Fulbright at Carnegie Mellon)
The deal
$20M round led by Index Ventures, May 19th 2026, with Susa Ventures, EF and Syndicate 1.
Inspection is how the physical world gets audited, a $250B industry where a four-hour site visit can produce days of paperwork, and reliable multimodal capture into a structured report only became buildable with the current model generation. The company claims 7 of the 10 largest inspection providers use the product. The case against: voice-and-video-to-report is becoming a commodity capability of the model APIs, the incumbent TIC giants can build or buy, and neither founder comes from the inspection industry; the moat has to be standards libraries and workflow depth, and in a liability business the residual error rate is where the lawsuits live.
Autonomous AI scientist agents that read literature, design experiments, generate lab-automation-ready protocols, analyze results and decide what to run next.
Founders
Nick Edwards, CEO (neuroscience PhD; NIH, BCG, Illumina), Ryan Kosai, CTO (Pioneer Square Labs, Attunely, ExtraHop)
The deal
$4.5M seed led by Draper Associates, with Dolby Family Ventures, Boost VC, Ensemble VC, Silicon Badia and Alumni Ventures.
The pitch is not an AI genius scientist, it is reproducible, auditable experiment execution, the unglamorous part academic labs cannot staff, and Edwards lived the reproducibility problem from inside NIH and Illumina. Closing the loop from paper to machine-executable wet-lab protocol only became sustainable with current agentic tool use. The case against: FutureHouse, Lila Sciences and the labs' own science agents chase the same thesis with 10 to 100 times the capital, $4.5M is thin for anything touching wet labs, and the buyer, a principal investigator, is the customer most professionally incentivized to distrust a black box doing their science.
Stockholm vibe-coding platform: $400M Series C at $13.3B, double its December mark, co-led by Menlo Ventures and EQT's Scaleup Europe Fund.
Tencent and Balderton join the cap table, and the company says annual recurring revenue nearly tripled from $200M toward $600M by the end of August. The valuation doubled in eight months on a category whose defensibility is the open question, since the same prompt-to-app capability ships in every coding agent; Tencent buying into a flagship European AI company the week Beijing finished unwinding Meta's Manus purchase is the more durable fact.
AI software for legal research, drafting and organized legal workflows.
AI agents for law firms and in-house legal teams doing research, due diligence, contract review and drafting, with more than 1,000 enterprise customers including Cleary Gottlieb, White & Case and Linklaters. Nvidia's venture arm and Atlassian backed the April round. The doubling in four months puts it in the same bracket as Cognition's reported $40B talks the same week, AI-native software repricing on usage growth rather than an annual round cycle. Neither company's new terms are public and both sets of talks are early. October 21st 2026: announced a $550 million Series E at an $8.8 billion post-money valuation, led by Atreides Management. This is the announced round; the earlier valuation was a report of talks. [A](https://legora.com/blog/series-e)
An AI-native government affairs firm. Agents plus in-house policy staff track regulation, lobby, and chase federal contracts for early-stage companies.
Two people billing for lobbying work itself, so the price moves with output instead of headcount. YC lists a $500K annual run-rate six weeks after a quiet launch, across defense, healthcare and govtech clients. The obvious objection is that government affairs runs on relationships an agent does not have. The firm's answer so far is to keep human lobbyists and put the agents underneath them, on the monitoring. It has since renamed to Locke.
Scanners have been commodity for years, and every one of them ends by handing an engineer a list. A company selling into banks then spends weeks rebuilding images to clear it. Emphere patches the distribution the customer already runs. That is the harder engineering problem, and it is why the incumbents sell hardened base images instead. Five people, two of them security researchers who attack the patched images to check the work.
The product assumes enterprises already have agents running on shared human credentials, with no approval step and no log. That is the same mess service accounts made, and it took a decade to clean up. If Decawork has the timing right, the budget sits with IT, which is a different sale from the one most AI startups are making, and the thing that wins is an identity layer nobody currently owns.
Merged with Churchill Capital Corp XI; common stock and warrants begin trading on Nasdaq on November 5th 2026 as AGLT and AGLTW.
Announced June 24th with over $620M of gross proceeds including a roughly $200M Foxconn-led PIPE, and more than $300M of multi-year Digit v5 orders from Schaeffler, GXO, Toyota Motor Manufacturing Canada and Mercado Libre. Logged now because aggregators began claiming this week that it went public; it has not. There is no effective S-4, no shareholder vote date and no listing, and the status here stays filed until the SEC record says otherwise. September 14th pipeline check: SEC accession 0001213900-26-097764 establishes a public S-4 filed September 4th, beyond the previously recorded confidential submission. A registration statement does not establish merger completion or a trading debut. October 1st review: the September 29th issuer announcement still makes public-company board appointments conditional on completion of the business combination. Retain filed status. [A] October 20th review: the issuer says the SEC declared the S-4 effective October 13th and the shareholder meeting is scheduled for November 2nd. This supersedes the earlier note saying neither milestone existed. Completion and trading remain unestablished; status stays filed. [A](https://www.agilityrobotics.com/content/agility-robotics-and-churchill-capital-corp-xi-s-4-registration-statement-declared-effective-by-sec-in-connection-with-proposed-business-combination) November 1st monthly review: Churchill XI's October 30th release expects closing November 4th and about $465 million gross, $164 million from trust after redemptions plus a roughly $201 million PIPE, against the more than $620 million assumed with no redemptions. Status stays filed until the vote and the close happen. [A](https://www.sec.gov/Archives/edgar/data/0002074973/000121390026114063/ea031030301ex99-1.htm) November 3rd 2026: Churchill XI shareholders approved the merger on November 2nd with more than 84% of votes cast in favor; about $465 million gross (about $164 million left in trust, a $201 million Foxconn-led PIPE, $100 million already received). Close expected November 4th, trading as AGLT on November 5th. Status stays filed until it lists. November 5th 2026: the company says the combination is complete and trading begins today, with about $465 million of gross proceeds, of which about $164 million came from the trust, about $201 million from the Foxconn-led PIPE at $10 a share and $100 million from earlier funding. First-half revenue was $174,000 against a $362 million loss. No redemption figure is published. [A]
Running four Claude Code and Codex sessions costs almost nothing. Catching the one that made a bad assumption forty minutes ago costs a lot, and it gets paid in tokens and rework. Agent FM's answer is ambient: the agents talk in a group chat you overhear, with audio, so you steer instead of poll. The risk is obvious. Cursor and Claude Code could ship the same thing in a release.
197 companies, about 40 launched so far; the AI cluster is agents that do work, infrastructure to govern them, and connectors to the physical world.
The AI cluster in this batch runs to agents that do work, infrastructure to govern them, and hardware connectors. Demo Day is September 10th, so the list is still filling. Individual companies are written up on their own pages.
Listed on Shanghai's STAR Market on August 19th 2026 at RMB 150.80 per share, raising about RMB 6.1B ($905M); it opened at RMB 1,100, up 629%, and closed its first session at RMB 845, up 460%.
China's first mainland humanoid listing closed its debut worth about RMB 342B ($48B), roughly 5.6x the IPO valuation and, by ts2.tech's reckoning, near 857x projected 2026 earnings. The retail tranche was 8,288.82x oversubscribed; a single winning lot showed about RMB 347,000 of paper profit at the close. The stock faded 23% from its RMB 1,100 open, the pattern of a momentum debut rather than a re-rating, and founder Wang Xingxing keeps 68.8% of the vote. Meituan's RMB 420M stake is now worth about RMB 38.6B. Caixin counts 30 to 50 Chinese robotics firms preparing Hong Kong listings behind it.
Data-center platform announced August 10th 2026, owned and majority-funded by Macquarie Asset Management and GIC funds, with Anthropic as anchor tenant under long-term leases.
The structure is the product: Anthropic gets purpose-built US capacity without the assets touching its pre-IPO balance sheet, and commits to 100% of grid-upgrade costs plus any consumer electricity increases its demand causes. No dollar figure disclosed, which for a platform of this ambition is itself information.
Altman rules out a 2026 IPO; previously reported confidential filing remains in the record.
The secondary market has already voted on this one: roughly $600M of OpenAI shares sat unsold while $2B of bids queued for Anthropic. Informed holders express views in secondaries that they will not state publicly. September 9th 2026: Semafor, citing the Financial Times, reports bankers seeking investment-grade ratings after listing. No awarded rating or completed listing is established by that report. September 13th 2026 update: Bloomberg reports that Altman told Fortune a listing will not happen this year as OpenAI prioritizes safety work. No firm 2027 listing date or withdrawal of the reported confidential filing is established. September 12th 2026: Altman ruled out a 2026 offering in a Fortune interview, reported by Axios. Preserve filed status; a delay does not establish withdrawal of the filing. October 20th: Reuters reports accelerated commitments for a roughly $30B private round ahead of Anthropic’s expected mid-November listing. A private financing discussion does not establish a new OpenAI IPO timetable. [B](https://uk.marketscreener.com/news/openai-seeks-to-accelerate-fundraising-ahead-of-anthropic-ipo-sources-say-ce785cdbdf8cf321) November 8th weekly: OpenAI closed $11 billion of a private round at $1.25 trillion on November 6th, below the $1.4 trillion to $1.5 trillion reported as its ask, and Brockman said it will be ready to list in 2027. A private round, not a listing step; status unchanged. [B](https://www.bloomberg.com/news/articles/2026-11-06/openai-completes-11-billion-portion-of-funding-at-1-25-trillion-value)
Public-benefit corporation aimed at automating the scientific method, founded by Jeff Dean and Sanjay Ghemawat.
Dean and Ghemawat left Google after 27 years, taking Oriol Vinyals and Quoc Le. Google is a founding investor, cloud partner and first-year compute supplier, which reads more like a settlement than a blessing.
Announced by the company, so the number is theirs and no reporter reconstructed it. Read it as a gauge of whether late-stage private marks are still being set by conviction or by momentum.
Amended S-1 of November 10th sets a $70 to $78 range for 825 million Class A shares on Nasdaq under ANT; not yet priced.
Enterprise-weighted revenue growing from $1B in 20 months. The $1.75T figure circulating in some coverage appears to be contamination from SpaceX's valuation; $965B paired with $47B is the internally consistent pair. September 9th 2026: Semafor, citing the Financial Times, reports bankers seeking investment-grade ratings after listing. No awarded rating or completed listing is established by that report. September 11th 2026: Reuters reports Nvidia discussing an anchor investment of up to $10 billion. The reported target is up to $100 billion raised at around a $2 trillion valuation; these are negotiations, not a final price or commitment. September 14th update: Reuters, citing Business Insider, reports Nasdaq has been selected. A separate Reuters account of FT reporting describes expected positive adjusted operating income for a second consecutive quarter. These are sourced reports, not a public prospectus, final pricing or a completed IPO. September 14th 2026: Axios sources still expect an offering in 2026 despite the safety debate. This is a timing report, not a completed listing. September 19th update: Axios reports a revenue pace above $100B annualized. That is neither audited fiscal revenue nor evidence of a completed IPO. September 20th update: Reuters reports deliberations over a model release ahead of an expected IPO. The report establishes neither a release nor a completed offering; existing filing status is unchanged. Also September 20th: The Wall Street Journal, via wire relays, reports the listing slipping past its expected October target to November, still at up to $100 billion around a $2 trillion valuation, so the company can market on third-quarter results; sources say the timing was set before the public slowdown debate. The S-1 remains confidential. September 28th review: Reuters reported on September 24th that Anthropic proposed 50.1% founder voting control, excluding board elections, conditional on three of seven founders retaining minimum holdings. A proposal, not confirmed adoption. September 29th: Reuters obtained the confidential prospectus, reporting 2025 operating losses above $8B. This does not establish a public S-1 or an offer price. October 1st 2026: the September 30th deadline for a public S-1 passed with EDGAR showing no Anthropic filing when read two minutes after 23:59 UTC; the prospectus Reuters reviewed remains confidential. A November listing now requires the public flip, road show and pricing inside eight weeks. October 2nd update: Bloomberg reports a possible marketing start the week of November 9th, targeting trading before Thanksgiving. This is a reported timetable, not a priced deal or a confirmed first trading date. [B] October 15th: Reuters reports investor-day figures of $16.8B third-quarter revenue and adjusted profitability. These are not audited public results or evidence of completed trading. [B](https://uk.marketscreener.com/news/anthropic-s-q3-revenue-surges-45-to-nearly-17-billion-ce785dd2d08cf221) October 20th: Reuters sources expect a mid-November listing. Keep the timetable provisional and status filed pending pricing or completed trading. [B](https://uk.marketscreener.com/news/openai-seeks-to-accelerate-fundraising-ahead-of-anthropic-ipo-sources-say-ce785cdbdf8cf321) October 27th verification: the public S-1 filed October 26th reports $4.585 billion in fiscal 2025 revenue. Public registration is confirmed; final pricing and completed trading are not established. [A](https://www.sec.gov/Archives/edgar/data/2130357/000162828026067168/anthropicpbc-sx1.htm) October 28th 2026: Bloomberg, citing people familiar with the matter, reports Anthropic is weighing reserving 15% to 20% of IPO shares for retail investors, with JPMorgan leading the US retail portion; the S-1 itself says only that banks may set aside shares for their online brokerage customers. Senator Warren sent questions on October 27th with answers due November 3rd. Renaissance Capital expects pricing in mid-November. Still a filing, not a price. November 1st monthly review: the public S-1 filed October 26th reports fiscal 2025 revenue of $4.585 billion. No price range or pricing through October; Blackstone reportedly expects a junk credit rating after listing, against the investment-grade goal reported in September. November 1st weekly: the S-1 leaves the aggregate committed spend and the October 31st run-rate blank. Named contracts total about $475.8 billion on our arithmetic: about $269.4 billion of usage-independent minimums with Microsoft, Google and Amazon, and about $206.4 billion with SpaceXAI, Nscale and Lambda that is cancelable or paid only on delivery. Still filed, no price range. November 2nd 2026: a filing reported by Reuters sets aside up to 552.9 billion yen (about $3.5 billion) of new shares for Japanese investors; still no total offering size, price range or valuation. November 6th 2026: Bloomberg reported some prospective buyers discussing $1.2 trillion to $1.3 trillion, down from $1.4 trillion to $1.5 trillion after an October briefing and as much as $2 trillion in early October; Reuters reported doubts the roadshow can start the week of November 9th and still list in November. No price range had been filed as of November 5th. Still filed. [B] November 8th weekly: the amended S-1 of November 2nd puts aggregate committed spend at about $895.4 billion with $69.8 billion due in twelve months, third-quarter revenue at $16.96 billion and an operating loss of $567 million. Buyer talk moved from $2 trillion in early October to $1.2 trillion to $1.3 trillion by November 6th. Still filed: no price range, share count or October revenue. November 10th 2026: amendment No. 2 to the S-1 sets an estimated range of $70 to $78 for 825 million Class A shares (801.4 million new, 23.6 million from selling stockholders) plus a 123.75 million share option for the underwriters, on the Nasdaq Global Select Market under ANT. At the range the base deal is $57.75 billion to $64.35 billion. Google, Nvidia, AMD, Broadcom, Micron, Microsoft and TSMC agree to buy $26 billion of nonvoting Class N stock at the offering price in a concurrent private placement. On the 13.09 billion Class A and Class N shares outstanding after the offering, before options and RSUs, the range implies about $916 billion to $1.02 trillion on our arithmetic, below the $1.2 trillion to $1.3 trillion buyer talk Bloomberg reported on November 6th. Goldman Sachs and Morgan Stanley lead 22 underwriters. A range is not a price, so the status stays filed. [A](https://www.sec.gov/Archives/edgar/data/2130357/000162828026072484/anthropicpbc-sx1ax2.htm) November 11th: Bloomberg reports orders for multiple times the shares on offer; the FT, citing five people, says bankers expect to price above the range, at $1.5 trillion or more. TSMC's board approved up to $500 million of Anthropic stock at the IPO price. Each of the seven founders sells 3,378,378 shares; together they keep 51.1% of votes. November 13th: Reuters, citing people familiar, reports the book about 2 times covered after the roadshow's first day and about 6 times by Thursday, ahead of SpaceX's pace. Banks value the company on roughly $270 billion to $280 billion of 2028 revenue; sell-side models put 2026 at $52 billion to $53 billion and 2027 at $135 billion to $137 billion, and one investor who passed sits about 30% below the banks on 2028. Investors questioned whether the lead lasts as Google catches up. Still filed: the range has not moved, and pricing is expected November 18th. November 15th weekly: the November 10th amendment also made Lambda's $34.9 billion August contract terminable by either side after Lambda missed a lease date (neither has given notice), raised Microsoft's commitment to at least $47.2 billion and kept Google hosting at about $121.3 billion. Nscale, another builder, told regulators it has no binding financing for its $44.6 billion Anthropic campus. Amodei told JPMorgan investors compute is "the deepest business dilemma that the company faces" and that Anthropic commits little upfront and buys spot; Bloomberg puts commitments near $901 billion, about $71 billion due by September 2027. Morningstar's fair value is $92 (bear $25, bull $188). Read together: the buyer holds the options and the builders hold the debt. Still filed; pricing expected November 18th. [B](https://www.bloomberg.com/news/articles/2026-11-13/anthropic-ceo-calls-computing-power-deepest-business-dilemma)
Shanghai humanoid maker that began a Hong Kong listing process on July 24th 2026, targeting HK$40-50B (about $5.1-6.4B); no prospectus filed yet.
Rumored, not filed: sponsors (Citic, CICC, Morgan Stanley) are hired but no A1 is public. The company it would be judged against just changed places with it. AgiBot shipped 8,400 humanoids in H1 2026, 44% of the world market, overtaking Unitree in the half Unitree spent going public. Backers include Tencent, Hillhouse and BYD. November 1st monthly review: TechNode, relaying Caijing, reports a new round near a 50 billion yuan valuation that pushes the Hong Kong listing past the end of 2026; Hong Kong's rules generally bar trading within 120 days of a pre-listing raise. AgiBot declined to comment. Status remains rumored. [B](https://technode.com/2026/10/14/robot-maker-agibot-nears-new-funding-round-at-rmb-50-billion-valuation/)
Filed with EDGAR, so this is the primary document and not coverage of it. Included because the filing itself is readable, which is rare in a pipeline built mostly of confidential submissions. October 1st correction: the September Form 8-K identifies the existing Nasdaq listing and trading symbol. The previous filed label was stale. This review does not establish the original first trading date. [A]