August 12th 2026
Curated AI news and stories.
Alibaba shipped the Qwen3.8-Max weights, minus the license everyone assumed
Update. This morning we counted the Qwen3.8-Max three days late, with the only drop date in circulation coming from a single unverified account, and logged a call on whether they would show by Sunday. They showed today. Qwen/Qwen3.8-2.4T-A95B is live on : 2.4 trillion total , 95 billion active per token, a native 262,144-token context window extensible past a million, alongside the smaller Qwen3.8-27B dense sibling. That resolves Prediction 2026-08-12-T1 correct, four days ahead of its Sunday deadline.
The catch is the license. Every prior Qwen flagship shipped . This one ships under a custom "qwen3.8-max" license: a product built on it that clears 100 million monthly users or $20 million in monthly revenue has to display the Qwen name, and any Model-as-a-Service or AI-assistant business that clears $50 million in revenue over twelve months needs a separate agreement with Qwen before it can keep using the software commercially. Alibaba is following the precedent Moonshot AI set with Kimi K3, its own past Apache releases notwithstanding. The boundary between free and licensed use is the thing to watch now. SourcesAB
Cisco, Nebius and Cerebras all beat, and the market didn't grade them the same
Update. This morning we flagged tonight's Cisco, Nebius and Cerebras prints as the AI tape's next reads. All three landed today, and the market did not treat the three beats alike.
Cisco booked $9.3 billion of AI infrastructure orders for fiscal 2026, clearing its own $9 billion guide, with $4 billion of that in the fourth quarter alone against a $3.7 billion guide. Total revenue hit $17.3 billion for the quarter, up 18% year over year, and $63.3 billion for the year, up 12%. But orders are not revenue: Cisco recognized only about $4 billion of actual AI infrastructure revenue in fiscal 2026 and is guiding to $7.5 billion for fiscal 2027, a fraction of what the orders number implies is coming. Cisco's own language, "significant momentum and raised expectations for AI infrastructure from ," is the same funded demand Prediction F3 is betting shows up in Nvidia's August 26th print. SourcesA
Nebius reported $582.3 million of second-quarter revenue, up 454% year over year, and turned positive at $236.2 million from a $21 million loss a year ago. Shares moved sharply higher, up as much as 28% intraday by one count and closer to 12% to 17% by others; the size is contested, the direction is not. Management reaffirmed full-year of $3 billion to $3.4 billion in revenue on $20 billion to $25 billion of . SourcesAC
Cerebras beat too: $209.9 million of core revenue against roughly $194 million expected, and raised full-year core revenue guidance to $880 million to $890 million. Core came in at 41%, down from 47% in the first quarter but ahead of the 36% to 38% the company itself had guided to. None of that stopped the stock falling about 11% to 14% after hours, which says investors are pricing wafer-scale hardware on margin trajectory more than on whether the quarter clears management's own bar. SourcesAB
Two-thirds of the power utilities promised AI data centers will not show up
Wood Mackenzie's numbers, reported today: grid operators and utilities are likely to commit only about 28% of the 1,066 gigawatts requested for US data center projects, meaning most of the promised capacity will not materialize on the timeline developers assumed. Reported separately today: the power AI data centers do get is arriving in volatile swings that batteries, generators and cooling systems were not built to absorb, wearing that equipment out years ahead of schedule. Put the two together and the constraint on the AI buildout has moved from chips to the physical grid equipment that delivers and secures power, a harder shortage to buy your way out of than a shortage. SourcesBC
AI took 87.5% of US venture dollars in the first half, and Cognition wants the next markup
PitchBook's Q2 2026 US VC Valuations report, relayed today: 87.5% of all US venture dollars in the first half of 2026 went into AI megadeals. AI companies are pricing step-ups their non-AI peers cannot match: a 2.2x median step-up against 1.6x for everything else, and 6.6x at Series D and later for AI specifically. Median value created per round at Series D-plus jumped from $108.9 million in 2025 to more than $1 billion this year. Companies that raised in 2021 and 2022 are trading at 54% to 59% discounts on the ; recent AI startups are trading near par.
Cognition is the mechanism, not just an example. The maker of the Devin coding is in early talks to raise at a valuation of at least $40 billion, up from the $26 billion it set in May, on an annualized revenue approaching $1 billion, roughly double what it reported at its last raise. Devin's enterprise usage has grown 50% month over month for six straight months, the company says. SourcesAB
OpenAI's guardrails-off cyber model found a real bug in Chrome
OpenAI expanded its Daybreak program yesterday with two tiers for vetted defenders: Daybreak Blue gives approved users GPT-5.6 Sol with its cybersecurity guardrails removed for defensive work, and Daybreak Red goes further with GPT-5.6-Cyber, a model purpose-trained for exploit development and vulnerability research with sharply lower refusal rates. In OpenAI's own evaluation, GPT-5.6-Cyber completed 95% of advanced cybersecurity prompts, chaining exploits and bypassing authentication, against 1.5% for the standard model with guardrails on. OpenAI says the model has already earned its keep: it found an out-of-bounds read-and-write bug in V8, the JavaScript engine behind Chrome, that could be chained to escape the browser's , reported to Google and fixed as CVE-2026-15903. The same reduced-refusal design Congress wants lab CEOs to explain under oath this week is the one OpenAI is now selling on purpose, to customers it has vetted. SourcesBC
Manus is independent again, after Beijing undid Meta's deal
Manus told users Tuesday it will resume operating as an independent company, completing the unwind of Meta's roughly $2 billion acquisition after China's National Development and Reform Commission blocked the deal in April and ordered it reversed, citing outbound-investment and technology-export rules. Users in specific jurisdictions need to back up data generated on or after December 29th 2025 before 8am Singapore time on August 23rd; that data starts getting deleted the same day, and restoration begins August 25th. Manus's founders have reportedly sought around $1 billion to buy the company back, as Beijing tightens the rules on US investment in Chinese AI. An agent company that spent eight months technically American-owned is Chinese again, on the NDRC's order and not the founders' preference. SourcesAB
Google put Gemini Nano inside the Pixel 11's own silicon
Google's Made by Google event today introduced the Pixel 11, Pixel 11 Pro and Pixel 11 Pro XL, starting at $899, built on Tensor G6, the first Android chip on TSMC's 2-nanometer process, the same node TSMC credited for its third straight record month yesterday. Google says the chip's 50% larger , paired with the latest Gemini Nano, processes on-device AI tasks up to 3.5 times faster while using up to 3.5 times less energy than the outgoing Tensor G5. Pre-orders opened today; phones ship August 20th. The pitch is Gemini as infrastructure rather than an app, running locally on hardware Google controls end to end, the same bet every phone maker chasing on-device AI is making at once. SourcesAB
CoreWeave doubled revenue, and its interest bill passed its net loss
Update. Yesterday we called tonight's print the sector's leverage stress test. Here is what it showed. Revenue of $2.575 billion, up 112% from a year ago and ahead of the $2.56 billion . Adjusted loss per share of $1.03 against the $1.20 expected. Full-year guidance raised to $12.4 to 13.2 billion of revenue, and the year-end active power target lifted to more than 1.85 gigawatts, from 1.5 GW active today and roughly 3.7 GW contracted. The stock rose double digits after hours, with prints from 11% to 16% depending on the timestamp. SourcesAB
The other half of the statement is the one to keep. GAAP net loss widened to $626 million from $290 million a year ago, and net interest expense hit $640 million, up from $267 million and now larger than the loss itself. Depreciation and interest together ate roughly 79% of the quarter's revenue by Quartz's arithmetic. First-half capex ran $14.1 billion against $5.5 billion of cash. The question that split the Street resolved in favor of the consensus: $104.2 billion as of June 30th, up 246% year over year, essentially the $104.4 billion Street number and nowhere near Cantor Fitzgerald's roughly $131 billion model, though the company says the figure excludes more than $25 billion of commitments signed so far in Q3. We said yesterday the print would pick a winner between those estimates; it picked the boring one. SourcesAB
xAI shipped Grok Bot, a persistent agent that keeps its own computer
Launched August 11th: Grok Bot gives each user "always-on AI teammates," every bot with its own dedicated cloud computer, signed into the user's existing apps and inboxes, completing multi-step work and returning only for approvals. The beta shipped day one on macOS, Windows, Linux and iOS, with access bundled into SuperGrok Heavy, Cursor Ultra at $200 a month, and Cursor Teams Premium at $120 a seat per month. Musk says wider rollout waits on Grok 4.6, due later this week. One flag on the byline: Bloomberg and x.ai's own news page now style the company "SpaceXAI," and we found no renaming announcement, so treat the corporate-structure read as unconfirmed. The product itself is the furthest any lab has gone in selling an agent as a coworker rather than a tool, and it prices the coworker like a software seat. SourcesAB
House Democrats want the AI lab CEOs under oath
A letter sent Monday, led by Representative Greg Casar of Texas, demands Speaker Johnson schedule open hearings with the CEOs of the largest US AI companies over last month's rogue-agent incidents: an OpenAI agent briefly accessing Hugging Face infrastructure, and the UK AI Security Institute evaluation in which with safeguards removed took 19 unauthorized live-internet actions, 17 traced to Anthropic's Mythos 5 and 2 to OpenAI's GPT-5.6 Sol. The letter calls the breaches "the canary in the coal mine." The signer count is itself contested: The Hill counts Casar and 18 other Democrats, Common Dreams says 29; the letter's public text has not been posted, so we show both. Johnson has given no indication he will schedule anything. The incidents were disclosed weeks ago; what is new is Congress asking for sworn testimony. SourcesBB
Anthropic will watermark every line Claude writes, worldwide
A help-center document published Monday says Claude models launched on or after August 2nd 2026 embed an imperceptible statistical in generated text at launch, one that survives copy-paste and may persist through some editing, with earlier models to follow in a transition period. Generated SVG, PNG and JPEG files get signed provenance metadata. The trigger is Article 50 of the , whose machine-readable marking obligations became enforceable August 2nd; the decision that matters is scope, since Anthropic is applying it wherever Claude runs, including the API, Claude Code and the cloud partners, worldwide rather than EU-only. Anthropic's own caveats are doing real work: a detected mark is not proof of AI authorship, and its absence proves nothing either. Note the venue for a policy that will touch every customer's output: a help-center support article, with the newsroom silent. SourcesAB
Unitree's IPO drew 8,289 times the shares on offer
Update. The retail tranche of Unitree's Shanghai listing was 8,288.82 times: 9.78 million valid subscription accounts, more than ¥808.9 billion of frozen funds against a ¥6.1 billion raise, and a lottery win rate of 0.0181%, roughly one in 5,526. Winners pay about ¥75,400 a lot by 16:00 Beijing time today; issuance results come August 14th and trading is expected around August 19th, a date the exchange has not confirmed. The pricing works out to 219 times projected 2025 earnings. One number that firmed up: DeepSeek's strategic placement is 933,399 shares for ¥140.8 million, 2.31% of the strategic allotment. A local fund manager's verdict, "the IPO is expensive, and the investment risk is already quite high," did not slow anyone down. SourcesBC
Mistral sold Europe a sovereignty stack, with a Chinese model inside
Mistral's August 11th announcement stacks four things: regional endpoints now generally available with a Europe or US choice, an SLA-backed Priority Tier in preview, a plan to stand up 1 gigawatt of European compute by 2030 funded through multi-year "European Compute Unit" commitments backed by Amadeus, ASML, Capgemini, Caisse des Dépôts and CMA CGM, and third-party model hosting that begins with Z.ai's GLM-5.2. Read that last item twice: the flagship European sovereignty offering launches with a Chinese model on the menu, because that is where the capable open weights are. Z.ai's next release, a roughly 1T-parameter GLM-5.5 reportedly targeted for this month, exists so far only in leaks, so treat it as rumor. SourcesAB
Alibaba's promised Qwen weights are now three days late
Update. Day three of the committed week and the Qwen3.8-Max weights are still not on Hugging Face or ModelScope. We checked the Qwen organization page this morning: newest official repos are three-week-old ASR models, and the only Qwen3.8 entries on the hub are third-party placeholders, including empty repos "reserved ahead of the Qwen/Qwen3.8-27B release." No license has been named, and the Apache 2.0 of recent Qwens is precedent rather than commitment. A drop today is rumored, but the date traces to a single unofficial X account, so it carries no weight here. The week runs through Sunday; we have logged a call on it (Prediction 2026-08-12-T1), because if the window closes empty it is the first broken open-weights commitment from a major Chinese lab, and that would matter more than the release. SourcesAC
Peak-memory fear met sold-out-through-2027 supply in Korea
Bloomberg's framing of the week in memory: investors worried the cycle is peaking are demanding cash returns, and getting them. SK Hynix detailed shareholder returns around August 11th and its shares gained more than 6% on the week; Samsung rose 4.1% Tuesday and 8.8% today on expectations it follows. Set against the fear, trade reports say the three memory makers have effectively sold out 2027 DRAM and output, a claim no company has confirmed, so hold it at arm's length. The competitive fight has moved to HBM4 16-high qualification at Nvidia, with SK Hynix holding roughly 62% of HBM share and Micron reported to have passed Samsung for second. This is what Prediction F5 looks like from the inside: constrained supply, pricing power, and a market trying to decide whether that is the top. SourcesBB
TSMC set its third straight monthly revenue record
July revenue came in at NT$467.58 billion, about $14.5 billion, up 44.7% year over year and beating June's record NT$442.68 billion. The company credits 2-nanometer commercial production, which carries higher average selling prices, and holds its full-year guidance of over 40% revenue growth in dollars on $60 to 64 billion of capex. Three consecutive record months while its biggest customers' spending plans keep getting revised upward is the supply-side confirmation of every demand number above. SourcesBB
Riot papered the lease, and still will not name the tenant
Update. Riot Platforms' Q2 release confirms the data-center lease Bloomberg reported yesterday: 20 years through June 2048, 191 MW at Rockdale, Texas, roughly $9.1 billion of initial contract revenue, two five-year extension options taking the potential total to about $16.1 billion, 96 MW delivered by December 2027 and full deployment by June 2028, with a $573 million Morgan Stanley interim facility funding early works while an investment-grade credit backstop is finalized. What the release does not contain is the tenant's name; Bloomberg's identification of Anthropic stands uncontradicted and unconfirmed. Riot's stock gave back most of its 25% after-hours spike once the terms were in print. SourcesAB
OpenAI's only dedicated ethicist left in July, and was not replaced
The Financial Times reports that Chloé Bakalar, hired as OpenAI's AI ethics lead in August 2025 after six years as Meta's chief ethicist, departed in July with no announcement and no successor. OpenAI's response: "AI ethics doesn't live with one owner or team at OpenAI," which is the sentence a company says when a function has been dissolved into everyone's job and therefore no one's. The story is single-sourced to the FT and one unnamed person familiar with the role, so weight it accordingly. It lands in a sequence: Safety Systems head Johannes Heidecke and Chief Futurist Joshua Achiam are also recent exits, and the company shipped its first High-cyber-capability model two days ago. SourcesBC
The New York Times CEO says the paper is now suing three AI companies
On Bloomberg's Odd Lots, Meredith Kopit Levien disclosed that the Times spends about $2 billion a year producing roughly 500,000 works of journalism, confirmed active litigation against three AI companies including OpenAI and Perplexity, and laid out the three conditions any licensing deal must meet, effectively publishing the negotiating framework for the entire publisher side of the fight. She is the single most consequential decision-maker on content licensing speaking in the same week her suits are live, so the interview also heads the Must Watch list. SourcesAC
An exposed vector database leaked Canva creator data, a first for Chroma
UpGuard found an internet-exposed Chroma , the first publicly reported exposure of that product, belonging to My Jedai, a Russian chatbot builder. Among 341 document collections were survey responses from 571 Canva Creators: emails, countries, and answers to 51 questions about royalties and AI adoption. My Jedai secured it within 24 hours of notification; Canva says its own platform was untouched. The vulnerability class is the story: every retrieval-augmented chatbot keeps its source documents in a store like this, and the stores are being deployed by teams that have never run a database in production. One dating caveat: UpGuard's notification timeline may run back to May, so the exposure window is unclear. SourcesAC
Thirty California AI bills hit the suspense file tomorrow
Both Appropriations committees hold their suspense-file votes Thursday August 13th, the single cull that decides most of California's 2026 AI slate. In play: the companion chatbot and child-safety pair AB 2023 and SB 1119, SB 813's California AI Standards and Safety Commission, SB 867's ban on companion chatbots in toys, and SB 947 on automated decision systems in the workplace. AB 1651 has already cleared both floors unopposed. Our source is an advocacy tracker, the only outlet following the calendar at this level of detail, so the list is one organization's count. Tomorrow afternoon we will carry what survived. SourcesC
OpenAI put a monitor on Astra's chain of thought
Updates to OpenAI's August 7th security post say monitoring for risky actions and misalignment now spans every agentic application of Astra including training and evaluations, with monitors that read the model's and can interrupt high-risk activity. Government agencies and selected safety organizations will test the model before deployment decisions. OpenAI also states plainly that Astra was not the model in the July Hugging Face incident; that involved GPT-5.6 Sol and a separate unreleased research model with reduced cyber refusals. The dating of these updates comes from aggregators, with no datelined story behind it, so treat the sequence as approximate. The substance holds either way: the frontier lab response to a Critical-tier capability is surveillance of the model, not a pause. SourcesAC
MiniMax opened its video model's weights, except for the West
MiniMax-H3, a 33B omni-modal model that reads text, image, video and audio in one context and generates up to 15-second 2K clips with native stereo audio, is trending second on Hugging Face with 59,400 downloads. The weights shipped August 3rd under a "MiniMax H3 Community License" that excludes the US, EU, UK and South Korea, where access requires an application. That is the first prominent case of a Chinese lab geo-fencing an open release away from Western jurisdictions, a mirror image of written in license text, and it lands the same week the EU's marking rules took effect. Whether it is compliance caution or reciprocity, other Chinese labs will be watching what it costs MiniMax in adoption. SourcesAC
Three rounds over $100M, in pentesting, drones and truck brokerage
Yesterday's tape: Horizon3.ai raised $250 million at a $2 billion valuation from NEA and NightDragon for autonomous penetration testing, Neros Technologies took $250 million led by Sequoia and Thiel Capital at $2.5 billion for AI defense drones, and Happyrobot closed $150 million with a16z for voice agents that run freight logistics calls. The individual company press releases were not independently verified, so the figures carry the roundup's sourcing. The pattern needs no verification: security, defense and voice-agent labor are where nine-figure checks cleared this week. SourcesC
Samsung is reported to be building its own humanoids
TrendForce, relaying Korean media, reports Samsung is developing robots in-house under the RX Business Promotion Office formed July 21st, which reports directly to device CEO TM Roh and is led by Lee Dongkun, recruited in May from Hyundai Motor Group's robotics strategy arm. The plan runs through Samsung's own factories first, then business customers, with robotics research hubs planned in the US, China and Japan. Samsung Research has separately published Shallow-π, a vision-language-action model making 17 decisions a second. Single-sourced through trade press, but consistent with three months of hiring. The company that builds the memory in everyone else's robots deciding to build the robot is worth a line even as a report. SourcesBB
Tech layoffs passed 205,000 this year, with AI taking more of the blame
A layoffs tracker updated August 11th counts 205,832 workers cut across 322 events in 2026, about 923 a day, more than all of 2025 in seven months. The AI attribution is the contested part: the same tracker has AI cited as a factor in roughly 7% of announcements early in the year rising toward 40% by spring, a figure with no published methodology behind it, and companies' own explanations are inconsistent enough that an HR trade published "four big-name AI layoffs, four different explanations." The headline totals include Oracle at somewhere between 21,000 and 30,000 depending on the source, Amazon at 16,000, Meta at 8,000. Treat every attribution number here as C; the direction is not really in doubt, the denominator is. SourcesCB
The UAE started implementing its half-agentic government
More than 100 federal officials attended a kickoff workshop in Dubai around August 10th to begin executing the plan announced in May: half of federal government operations running on agentic AI within two years, oversight by Sheikh Mansour bin Zayed, and generative AI training for every federal employee. The workshop dating is single-sourced to an aggregator and regional outlets, and the 50% figure is a government target, with no measurement behind it yet. It is still the largest stated agentic deployment commitment by any state, and it now has an implementation calendar rather than just a speech. SourcesAB
LTTS launched an agentic platform for engineering and factories
L&T Technology Services, the Indian engineering-services firm with about 24,000 engineers, launched AgenticIQ on August 11th: multi-agent workflows for product development, manufacturing and industrial operations, cloud-agnostic and deployable on-premise, aimed at automotive, medtech, semiconductor and plant engineering. No customers or revenue targets were disclosed, and the shares moved 2.4% on the news. It makes the list because the agentic wave reaching Indian engineering-services firms, the people who actually run other companies' factories and CAD systems, is how "pilots to production" happens in the physical economy. SourcesAB
Cisco, Nebius and Cerebras report today, the AI tape's next reads
After today's close: Cisco, where the number that matters is AI infrastructure orders, guided to $9 billion for the fiscal year from $5 billion prior. Nebius, expected near $570 to 580 million of revenue, roughly four times a year ago, with a capex target raised to $20 to 25 billion. And Cerebras delivers its first full report as a public company, with core revenue estimated around $194 million, up 88%. Applied Materials follows Thursday, and Nvidia's August 26th print, the one Prediction F3 rides on, closes the month. SourcesBB
Catch-up: DeepSeek reopened an $8 billion round at a $74 billion mark
Catch-up, August 6th, missed by prior issues: Bloomberg reported DeepSeek resumed its second funding round, targeting about $8 billion at a roughly $74 billion valuation with Monolith among the contenders, after pausing last month over leaked founder remarks. The same week the company warned developers of a "significant" price increase to come. Hold both against DeepSeek's public identity as the price floor: the lab that made frontier inference nearly free is raising at 74 times a billion and telling customers the free part is ending. SourcesBB