August 20th 2026
Curated AI news and stories.
Alibaba's profit fell 76% on a doubling of AI capex, and investors sold the stock anyway
Alibaba reported fiscal first-quarter net income of 10.54 billion yuan, about $1.55 billion, down 76% from a year earlier, as rose 75% to 67.68 billion yuan, roughly $9.98 billion, almost entirely AI infrastructure. Revenue grew 9% to 268.95 billion yuan, and cloud revenue, the segment carrying the AI story, grew 45% to 48.44 billion yuan with up 133%. Free cash flow was negative 44.67 billion yuan, more than double the prior year's outflow. Hong Kong-listed shares fell about 5% on the print. The quarter is the clearest read yet on what "committing to AI" costs a company that already has the customers and the cloud business to sell it into: revenue climbing at a normal enterprise pace while the spending line goes vertical, and a market that priced in the growth story now pricing in the bill. CEO Eddie Wu called it commercialization of the company's "full-stack AI capabilities." The continued regardless, $162 billion in shares repurchased on the quarter. SourcesBB
Treasury yields gave back half of Wednesday's drop, and US stocks fell with them
The 10-year yield rose 4 basis points to 4.69% and the 30-year rose 4 basis points to 5.24% Thursday, undoing roughly half of the relief that followed Wednesday's Treasury buyback announcement, which had pulled the 30-year down to an intraday low of 5.18% from Tuesday's 5.333% close. The S&P 500 fell 0.87% to 7,641.16, the Nasdaq fell 1.00% to 26,067.17, and the Dow fell 1.32% to 52,759.21, dragged further by a 9% drop in Walmart despite the retailer beating earnings estimates on weaker-than-expected same-store sales growth. This is the same mechanism this morning's brief flagged in Seoul's rebound, just running the other direction on the other side of the market: a bond-buying operation moved the price of money for one session, and one session is what it bought. The chip trade that recovered in Asia on Thursday was built on Wednesday's US yield relief holding into the next session. It did not fully hold. SourcesB
Washington says hackers are using AI to write exploits for the controllers running water and power plants
CISA, the NSA, the FBI, the Department of Energy and the EPA issued a joint advisory Wednesday warning of active exploitation attempts against internet-exposed Siemens S7 series programmable logic controllers, the hardware that operates water treatment, energy and manufacturing equipment across the country. Attackers scan for exposed devices with commercial tools like Censys and ZoomEye, then feed the results to AI systems that generate working Python exploit scripts using the open-source snap7 library, without the attacker needing to understand the underlying industrial protocol. The advisory calls this "not a theoretical risk" and says AI "dramatically reduces the technical expertise and time required" to build a working intrusion tool. It follows a string of intrusions at water utilities in Minnesota, Michigan, Arkansas, Georgia and New Jersey that officials have linked to suspected Iranian-backed hackers. The specific claim to hold onto is narrower than "AI hacking infrastructure": the model is not finding the vulnerability, the scanner already did that. It is collapsing the distance between "we found an exposed " and "we have code that talks to it," which used to be the part that required a specialist. SourcesAB
Anthropic is preparing to loosen the enterprise data retention rule it announced in June
Anthropic plans to let enterprise customers using its most capable models keep the required 30 days of retained data on their own cloud infrastructure instead of Anthropic's, Bloomberg reported Thursday, rolling back part of a policy it introduced in June to guard against model-assisted cyberattacks. The change has reportedly been in development for months, developed in coordination with more than 100 enterprise customers including Salesforce, and is expected to ship later this year. It lands one day after OpenAI committed to zero data retention outright for its own frontier customers. Read together, both labs are conceding the same point from different directions: a policy written for security concerns is a sales obstacle the moment a competitor offers less friction, and "we hold your data to catch misuse" only survives as long as nobody enterprise-grade is willing to hold less of it. SourcesBC
Unitree's founder pushed the humanoid "ChatGPT moment" further out, a day after the stock quintupled
Wang Xingxing told the World Robot Conference in Beijing on Thursday that robots are two to three years from their defining breakthrough if progress moves fast, and five to ten years out if it does not, a more cautious call than his own forecast at the same conference a year ago, when he put the moment at under five years. He defined it as a robot completing roughly 80% of everyday tasks in an unfamiliar home from voice or text instructions alone, and said generalizing across unfamiliar tasks and precise manipulation remain unsolved. The remarks came one day after Unitree's Shanghai debut, where shares closed 460% above the offer price, more than five times the level at which the $905 million IPO was priced. The founder with the most reason to talk his own sector up just moved his own goalpost back, on the same stage, the day after the market handed him the best possible proof that nobody is waiting for the goalpost to arrive before buying in. SourcesBB
SpaceX tried to buy Cognition, and was told no
SpaceX approached Cognition, the maker of the Devin coding , about an acquisition as it works to close the gap with OpenAI, Anthropic and Google, Bloomberg reported Tuesday, five days after SpaceX closed its $60 billion purchase of Cursor. Cognition CEO Scott Wu disputed the story on X, saying the company "is not for sale" and that the two sides have not been in acquisition talks; people familiar told Bloomberg and TechCrunch the acquisition discussions are dead but that the companies are still discussing Cognition running on SpaceX compute. Read the two claims together and they are compatible: an approach that never reached a term sheet is easy to deny. The structural fact is the interesting part. SpaceX bought the most popular AI code editor last week and immediately went after the best-known coding agent, which says the plan is not one product, it is the whole category. Cognition's enterprise list, Mercedes-Benz, Citi and Goldman Sachs among them, fills the hole in Cursor's consumer-developer base. SourcesBB
Anthropic put $250 million on chips that will not exist until 2027
Fractile, a London inference-chip startup founded in 2022 by Oxford roboticist Walter Goodwin, is in advanced talks to raise about $600 million at a $6.5 billion pre-money , Bloomberg reported Wednesday, after reaching an initial agreement to sell Anthropic roughly $250 million of its chips. The valuation is more than six times the $1 billion Fractile raised at in May, and the chips are not expected to be usable until 2027. The load-bearing claim here is not Fractile's architecture, it is that Anthropic wants a supplier that is not Nvidia badly enough to commit a quarter-billion dollars to unshipped silicon. Every frontier lab now carries the same concentration risk: one vendor, one interconnect, one allocation queue. A $250 million order is cheap insurance against that, and for Fractile it converts a research bet into a company with a customer. What to watch is whether the contract survives first silicon, because most first chips miss their targets. SourcesBB
Korea's chipmakers took back in one session what they lost in two
The KOSPI closed Thursday at 6,852.58, up 5.89%, with SK Hynix up 12.73% to 1,691,000 won and Samsung up 9.49%, reversing the two-day selloff that had triggered a on Wednesday. The drivers were the same two that started the recovery in Wednesday's US session: SK Hynix's 40 trillion won buyback-and-cancellation, approved after Wednesday's close in Seoul, and lower US Treasury yields following the Treasury Department's doubled buyback operations. Speculation about a stronger Samsung shareholder-return announcement later this month did the rest. Foreign investors bought a net 1.7 trillion won, about $1.2 billion, while retail sold 2.3 trillion won into the bounce. Three sessions, a 5.7% fall and a 5.9% recovery, and nothing about memory supply or demand changed on any of them. The trade is being priced off the cost of money and the size of buybacks, which is worth remembering the next time a single day's move gets narrated as a verdict on AI demand. SourcesBB
OpenAI promised enterprises zero data retention, and a monitor that reads patterns instead of prompts
OpenAI committed Tuesday to zero data retention for enterprise API customers on its , and said it is testing a system called Private Safety Processing with early customers including Microsoft and Databricks. The problem the second system exists to solve is created by the first: retention-free processing evaluates each request in isolation, and the abuse OpenAI says it now sees most, from both malicious users and compromised agents, only becomes visible across many interactions over time. Private Safety Processing is supposed to detect those patterns without giving OpenAI staff access to the underlying prompts or responses; a technical paper and broader release are promised for September. Judge it when the paper lands. A privacy guarantee that depends on an unpublished mechanism is a claim, not a property, and the specific thing to look for is whether the pattern detector can be audited without becoming the very access it promises not to have. TechCrunch reads the move as chasing Anthropic's enterprise position, which the revenue numbers this week make a reasonable reading. SourcesAB
China is slowing exports of the materials Taiwan's optics and aerospace run on
China has been delaying customs clearance on germanium, quartz-based materials and neodymium magnets bound for Taiwan, Nikkei reported Wednesday, stretching delivery times for Taiwanese optical and aerospace suppliers and creating supply bottlenecks in the island's component industry. The delays reportedly date to 2025 and are tightening. Germanium sits inside infrared optics and fiber; high-purity quartz sits inside semiconductor tooling and crucibles; magnets sit inside actuators and motors. None of these is a headline product and all of them are upstream of things Taiwan sells to the world, including the equipment its chip industry runs on. This is the same playbook Beijing has run against Japan and the US, applied to the economy whose everyone else's AI buildout depends on, and it lands the same week the first licensed H200s entered the mainland. Chips flow one way across the strait while the materials to build chipmaking tools slow in the other direction. SourcesBC
Reuters traced Unitree's robot dogs to Army-funded American research
Unitree based the designs of its most successful quadrupeds on breakthroughs financed by the US Army Research Laboratory and allied military programs, Reuters reported this week, citing a former US defense technology official and three researchers involved. Ben Katz of MIT's Biomimetic Robotics Lab says the dimensions of Unitree's Go series match the Mini Cheetah he helped build "to the millimeter"; a former University of Pennsylvania researcher calls the Army-funded work "basically, the first Unitree robot that had any kind of scale"; and founder Wang Xingxing's 2016 master's thesis cites the MIT lab's pioneering work directly. Reuters is explicit that Unitree did nothing improper: the research was published openly, on purpose, because that is how government-funded science is supposed to work. That is what makes the story uncomfortable rather than scandalous. The US funded the foundational work, published it, built no company on it, and watched a Hangzhou startup turn it into an 85% market share and a $48 billion listing. The FCC's July import ban is the after-the-fact answer to a pipeline problem the ban does not touch. SourcesBB
Xiaomi put its first humanoid on a public stage
The 2026 World Robot Conference opened in Beijing on Wednesday with about 3,000 products on display, and the debut that mattered was Xiaomi's: a 1.7-meter humanoid built on what the company calls a universal framework spanning the hardware and the model behind it. Xiaomi led with factory numbers: in trials at an automotive plant the robot tightened nuts, a two-handed coordination task, at a 90.2% initial success rate, pushed to 98% with continued testing, and it is slated for deployment in Xiaomi's own smart-manufacturing lines. That framing is the tell. Unitree listed this week on backflips and retail euphoria; Xiaomi is pitching a machine that earns its keep on an assembly line it happens to own, where the employer can tolerate the robot's learning curve. The World Humanoid Robot Games start Saturday in the same city. SourcesBB
Three of 1,357 FDA-cleared AI devices were tested on whether patients got better
Of the 1,357 AI-enabled medical devices the FDA has authorized for patient care, 34 were linked to a registered prospective clinical trial and exactly three were evaluated on patient-centered outcomes like mortality, readmission or quality of life, according to an analysis published Tuesday in PLOS Digital Health. The rest, 99.8%, were cleared without anyone measuring whether patients do better with the device than without it. The mechanism is the pathway: a device is cleared by showing substantial equivalence to a device already on the market, which lets each new clearance chain off the last one, with no outcome trial anywhere in the chain. Radiology accounts for 78% of the devices. The claim being regulated is "this reads an image about as well as the previous one," and the claim being marketed is "this improves care." Those are different claims, only the first was ever tested, and 1,354 devices sit in the gap between them. SourcesBB
Flock built a police search that starts from how you drive
Flock Safety is testing a tool called OS Investigate with a small group of police departments that lets officers search for vehicles and people by movement pattern alone: no plate, no name, no suspect, no crime required as a starting point. Reporting by 404 Media, which obtained the tool's code, shows prompts that surface vehicles seen most often in a neighborhood over 14 days, or vehicles that visited three or more shops in three days, and the system can then pull dates of birth, Social Security numbers, relatives and home addresses from commercial identity databases to turn a car into a person. The ACLU called it AI-generated suspicion. The technical change is small, the same network Flock already runs, plus a query layer. The categorical change is the direction of the search. Every prior version started from a suspect and asked where they went; this starts from a pattern and asks who fits it, and "parked near three shops" is a pattern most of us fit some week. SourcesBB
Nvidia is playing matchmaker between GPU owners and Nordic data centers
Nvidia has been introducing companies that hold its to Nordic data-center operators with spare capacity, CNBC reported Wednesday, including approaching at least one operator to scout potential offtakers on its behalf. CFO Colette Kress acknowledged the practice in June, and the company now runs a marketplace called Compute MatchMake for exactly this. The Nordics are pulling in projects on cheap power, available land and free cooling. The interesting question is why the matchmaking is needed at all. If every purchased GPU had a rack, power and a workload waiting, there would be nothing to match. Brokers emerge when chips and capacity stop finding each other on their own, and allocation bought ahead of deployment plans is how they lose each other. Nvidia brokering its own aftermarket is good business and also a small telltale about how much of the two-year order book represents scheduled compute rather than scheduled need. SourcesBB
Cursor's cloud agents now wake on events and hold goals
Cursor's Tuesday changelog gives its cloud agents Subscriptions, standing attachments to an event source, a pull request, a Slack thread, a schedule, that wake the agent when something happens, plus a /goal command that holds an objective like "fix all flaky tests and make CI green" across sessions until it is met. Agents now auto-subscribe to pull requests they create and drive them through CI failures and bot comments; subagents get isolated VMs with a clean copy of the project each. This is the same week Warp shipped its agent factory and three days after Cursor launched a GitHub rival, and the shared direction is precise: the agent stops being a tool you invoke and becomes a process that runs until a condition is true. The component that decides whether this works is not the model, it is the goal checker, because an agent that can hold a goal it cannot correctly verify will hold it wrong indefinitely. SourcesA
Google Cloud is hiring an army of forward-deployed engineers
Google Cloud has 59 open roles for across the US, London, Paris and Hong Kong, CIO Dive reported this week, building teams that embed inside enterprise customers to turn agent prototypes into production systems, with senior total compensation reaching $700,000. CEO Thomas Kurian says demand for engineers to help customers "embrace agent development" is growing very rapidly. Sit this next to what the is simultaneously selling: agents that do the work of engineers. The honest reading is not hypocrisy, it is a measurement of where the technology actually is. If enterprise AI deployed itself, the margin would be pure software; instead the fastest-growing role at the vendor is a human who goes to the customer's building and makes it work, the Palantir model at Google scale. The size of the forward-deployed payroll is one of the better live gauges of the gap between agent demos and agent deployments, and right now it is growing. SourcesB
Amazon will fly drone deliveries in nearly 500 cities by year-end
Amazon said Wednesday that Prime Air will expand from 11 metro areas to nearly 500 US cities and towns by the end of 2026, a sixfold jump, with Chicago, Atlanta, Cleveland, Syracuse and Boise next. Each site covers about 175 square miles; items under five pounds, which Amazon says covers more than 60% of its most-purchased products, arrive in as little as 30 minutes. The constraint that matters here was never the aircraft, it was the FAA's beyond-visual-line-of-sight approvals and the per-site economics of a launch footprint, and a company does not announce a 45-fold city expansion unless it believes both have been solved in a repeatable template. Wing and Zipline are running the same race with different vehicles. Autonomy over sidewalks and streets keeps stalling on edge cases; autonomy at 200 feet has fewer pedestrians, and it is scaling first. SourcesAB
Munich Re bought At-Bay for $575 million, a 57% markdown
Munich Re agreed Wednesday to acquire At-Bay, the US cyber insurer that pairs policies with managed detection and response for mid-sized companies, for $575 million, closing in early 2027. At-Bay holds a top-ten position in US cyber insurance with $278 million in gross written premiums, so the price is roughly two times premiums, and 57% below the $1.35 billion valuation At-Bay carried after its 2021 Series D. The world's largest reinsurer buying a primary cyber , in the same month AI-assisted intrusion campaigns became a weekly story, is a statement about where it thinks cyber risk pricing is heading: in-house, data-driven and vertically integrated from reinsurance down to the endpoint sensor. The markdown is the other half of the story. Insurtech valuations were set in 2021 on software multiples; the exit is priced like an insurance book, because that is what it is. SourcesAB
Grok 4.6 landed on Amazon Bedrock
Grok 4.6, the flagship model from Elon Musk's xAI, now branded SpaceXAI, became generally available on Amazon Bedrock this week: a 500,000-token , text and image input, and four configurable reasoning levels, in every region Bedrock serves. On the xAI API it prices at $2 per million input and $6 output below 200,000 prompt tokens, doubling above that. The distribution is the story. AWS customers can now run the Musk model inside the same compliance envelope as Claude and Amazon's own models, one procurement signature away, which matters for a lab whose enterprise sales motion barely exists. It also means Amazon is hosting the model at the center of active CSAM litigation and a First Amendment challenge to Minnesota's nudification ban, a ruling on which is due by Monday. Platforms have opinions whether they state them or not; Bedrock's is that Grok is a product like any other. SourcesAA
Frontier models recovered 3% of research ideas on a blind test
A new called Reconstruction, circulated this week as a preprint, hands a model a paper's bibliography and asks it to recover the paper's core idea, a contamination-resistant proxy for hypothesis generation, and frontier models score 3% to 15% working alone. A multi-agent tournament design, generating candidate hypotheses and judging them against each other in Swiss-tournament rounds, reaches 42% on the same task. Both numbers are informative. The solo score says the models that summarize literature impressively do not, on their own, take the step from evidence to idea; the tournament score says a lot of the missing capability is recoverable through orchestration and selection rather than a better base model. This is the same shape as the protein-binder result on Tuesday: the wins come from running many candidates through a filter that can tell good from bad. Generation is cheap. The judge is the scarce part. The paper has not been peer reviewed. SourcesB
Vercel open-sourced fx, a coding agent that cold-starts in 10 microseconds
Vercel Labs released fx, an internal coding agent now open under : a single native binary written in Zig, 6.3 MiB, single-digit megabytes of memory at baseline, cold-starting in about 10 microseconds, model- and provider-agnostic, with no telemetry and sessions stored locally. It reads as a Unix tool rather than an IDE in the terminal: fx ask for headless tasks, an ACP mode for connecting to editors, a WebAssembly SDK for embedding. The dimensions are the argument. Every mainstream coding agent is a Node or Python process that assumes it is the main event; fx is built to be embedded, spawned by the thousand inside sandboxes and other agents' loops, where startup time and memory footprint are the budget. Whether the agent is good is almost secondary to what its existence says: the is becoming infrastructure, cheap enough to put anywhere, and the model behind it is a config field. SourcesAA
DFlash 2 gets 20% more tokens out of every verification pass
Inco released 2, the successor to the Z Lab speculative-decoding system, with draft models out for Qwen3.8-27B and Meta's Muse Glimmer 30B. The technique drafts tokens in parallel blocks and has the target model verify them, output provably identical to standard decoding; version 2 claims over 20% more accepted tokens per verification pass for about 1% added , 2.7 to 3.4 times autoregressive throughput in SGLang at batch size 1, and 70 tokens per second from Qwen3.8-27B on a MacBook M5 Max. The provable-equivalence property is why this matters more than most tricks: there is no quality trade to argue about, only a speed dividend, which makes adoption a pure engineering decision. Local agents on consumer hardware are the direct beneficiary. A 27-billion-parameter model at 70 tokens per second on a laptop is past the threshold where an on-device coding agent stops feeling like a compromise. SourcesAA
Google is giving US college students a free year of AI Pro
Google opened a free 12-month AI Pro subscription, normally $19.99 a month, to US college students who verify with a school email, bundling higher Gemini limits, Gemini Spark, Gemini in Gmail and Docs, 5TB of storage, and a new student hub with flashcards, practice quizzes and study notebooks. Students outside the US get a year of the cheaper AI Plus tier. OpenAI, Anthropic and Perplexity have all run versions of this play, and Perplexity's India numbers published this week show how it ends: most free users leave when the bill starts, and the ones who stay are the business. The student cohort is the one worth paying most for, because the tool a person learns to think with at twenty is the one they demand at their first job. That is how a spreadsheet company becomes an institution. The price of the giveaway is a rounding error against inference budgets; the switching costs it builds are not. SourcesAB
Alexa+ is now free on Fire TV, no Prime required
Amazon made Alexa+, its LLM-rebuilt assistant that cost $19.99 a month without Prime, free on current-generation Fire TV devices and compatible Hisense and Panasonic sets in the US, upgraded automatically with conversational search and smart-home control on screen. The free tier stops at the television: Echo speakers and the advanced home features still need Prime or a paid plan. Read alongside Google's student giveaway the same day, the consumer-assistant market has stopped pricing the assistant at all. The model bill is real, so a free assistant is being paid for somewhere else, and on a TV the somewhere else is obvious: engagement, commerce and the ad-supported surfaces the assistant steers you toward. The assistant is becoming the interface everyone gives away to sell what sits behind it. Search, maps and email each went the same way in their decades. SourcesBB
Prevalent AI took its first outside money in nine years
Prevalent AI, a London security-data company founded in 2017 by former GCHQ and Darktrace leaders, raised $22 million from Integrity Growth Partners, its first outside capital ever, to expand into the US and extend its platform beyond cybersecurity. The company builds a over fragmented enterprise security data so that human teams and AI agents can query risk with context; it says it has been profitable since its first customer and doubled in the past year. A nine-year bootstrap raising now is a signal about timing, not desperation: agent deployments are stalling on exactly the problem Prevalent sells a fix for, models grounded in enterprise data that is scattered, stale and contradictory. The quiet, profitable version of this company existed for years without a funding round. The agent wave is what made the data-context layer suddenly worth venture money. SourcesBB