August 26th 2026
Curated AI news and stories.
Moonshot wants 30% of what American clouds make on Kimi
Reuters reported this morning that Moonshot AI is in early talks with Microsoft, Amazon and Google on revenue-sharing agreements for hosting Kimi K3 on Azure, AWS and Google Cloud, seeking up to a 30% share of the revenue K3 services generate there. Still unresolved: how the split is computed, data access rights, and auditing of usage. All four companies declined comment, and the talks may die. If they close, it is the first major revenue-sharing pact between a Chinese AI firm and a US cloud, and it rewrites the bargain: until now the American clouds have served Chinese open models without paying their makers anything. The context makes it stranger. Treasury Secretary Scott Bessent has threatened to blacklist Moonshot, and US officials have accused it of stealing from Anthropic's and illegally acquiring Nvidia chips, both of which Moonshot disputes. Meanwhile its pre-IPO round is due to make its final close tomorrow at a reported $50 billion pre-money, ahead of a Hong Kong filing targeted by September 30th. A revenue line paid by three US hyperscalers would be the best page in that prospectus. SourcesBB
Alibaba open-weighted a preview of Qwen4 overnight
Qwen3.8-Flash-Next went live on this morning, in standard and checkpoints under Alibaba's Qwen community license, and the company frames it as an early look at the architecture of the coming Qwen4 family. The design is the story: 125 billion total with about 6 billion activated per token, plus 51 billion parameters of n-gram embeddings that spend capacity on lookup instead of compute, a hybrid attention stack pairing Gated DeltaNet with Qwen's sparse attention, and 262,144 tokens of native context extensible to a million. It reads text, images and video, thinks by default, and Alibaba's own numbers put it at 62.5 on SWE-bench Pro and 91.7 on GPQA Diamond, vendor figures with the usual discount. The claim under the specs is that flagship-adjacent capability now serves at small-model cost, and that claim lands on receptive buyers: AT&T said Monday it routes 40% of its AI queries to open models and wants 70%. Within hours the card was the top trending model on Hugging Face. SourcesAB
Nvidia reports tonight, and options price the calmest print in five years
The seven-session slide ended Tuesday with a 2.2% bounce, and tonight after the close Nvidia reports against a of $92.07 billion in revenue and $2.09 a share, roughly double the $46.74 billion and $1.04 of a year ago. The company's own guide was $91 billion plus or minus 2%, built on an explicit assumption of zero China data-center revenue, which makes any resumed H20 sales, previously floated at $2 to 5 billion for the quarter, pure upside against . TradingKey puts the options-implied move at about 5.4%, which it calls the smallest into an Nvidia print since August 2021, a striking amount of calm for a stock that spent seven sessions de-risking. Raymond James raised its target from $330 to $352 on Tuesday. The number that decides the next month of this market is the October-quarter guide, and the number to read behind it is , where the memory costs already repricing Amazon's device shelf either show up or get explained away. SourcesBB
Anthropic will pitch investors a $30 trillion market
Anthropic is expected to tell IPO investors its exceeds $30 trillion, the Wall Street Journal reported Tuesday via Reuters, a figure computed from "the full scope of work that could be completed with AI models" and built to top the $28.5 trillion SpaceX claimed in its May filing. The same reporting cites 2028 revenue projections of roughly $190 to 200 billion, against the confidential filed June 1st and an October listing that investors reportedly expect. still showed no public Anthropic filing this morning. Read the TAM construction literally, because investors will: a market sized from all the work AI could do is not a software market, it is the wage bill. A company that prices itself against labor is telling you whose income statement its growth comes out of, and it is not the incumbents' software budgets. SourcesB
Amazon is closing Mechanical Turk, the platform that labeled the training data
Amazon will shut Mechanical Turk on September 30th, ending 21 years of the marketplace Jeff Bezos once called "artificial artificial intelligence," where more than 500,000 people at its peak did the micro-work of the machine-learning era: labeling images, transcribing audio, rating outputs. The company says the decision followed "an internal assessment of its programs and services"; the platform stopped accepting new customers a month ago. The proximate cause is visible in the customer list, with Scale AI, Mercor and Prolific having taken the data business upmarket to credentialed experts, and the models themselves now doing much of what a HIT used to buy. What is quietly retiring is the bottom rung: the place a person with no credentials anywhere on earth could sell judgment to the AI economy by the task. The editorial takes this one up. SourcesBB
OpenAI published its first chip numbers, on Nvidia earnings eve
OpenAI released the first measured results for Jalapeño, the chip it designed with Broadcom, run on SemiAnalysis's public InferenceX with GPT-OSS 120B, DeepSeek R1 and Kimi K2 as workloads. The claims: 1.5 to 1.9 times more work per watt than the commercial systems compared, end-to-end 1.7 to 3.6 times lower, and 2.1 to 4.1 times faster on highly interactive workloads, with hardware lead Richard Ho saying the part serves more work per unit of power while returning responses faster. The comparison set includes Nvidia Blackwell systems, per TechCrunch. Deployment stays small through year-end, with volume planned for 2027. A first-party benchmark from the chip's own designer earns every discount it gets, and the timing earns a separate read: Nvidia's single largest inference customer published per-watt wins over Nvidia silicon the morning of Nvidia's print. SourcesAB
The SEC's Situational Awareness probe reached four banks
The has sent subpoenas to Goldman Sachs, JPMorgan Chase, Citigroup and Bank of America seeking records of Situational Awareness's trades, its borrowing, and its communications with lenders, CNBC reported Tuesday. The fund ran leverage as high as 400% before July's AI selloff forced margin calls, a distressed sale of its public book to Citadel, and a fall in assets from about $45 billion to about $15 billion. The investigation is at its earliest stages and nobody has been accused of wrongdoing. The question has sharpened, though: it is no longer just whether the fastest-growing fund in history had controls to match its size, it is whether the prime brokers who financed 400% leverage on a concentrated AI book did. Four subpoenas to four banks says the SEC is asking the second question. SourcesB
Unitree has given back 200 billion yuan, and the bubble postmortems have started
Unitree fell another 10% Monday to 605.10 yuan by Gasgoo's count, with Reuters putting the close at 603.08, and steadied Tuesday, leaving the stock about 45% below its debut peak and roughly 200 billion yuan of market value gone in four sessions, from an intraday peak near 444.9 billion to about 240 billion. The fundamentals under the fall: 2025 revenue of 1.699 billion yuan, first-half 2026 adjusted net profit of 40 million yuan, and a ratio still above 140 after the decline. Reuters' postmortem collected the quotes the way up never did. Fund manager Dong Baozhen: "all bubbles are doomed to burst." VC Abraham Zhang: the debut "was not fuelled by a rosy prospect, but a desire by some to pump up the shares so as to dump them later." A week ago this listing was the starting gun for China's IPO wave. It still is; what changed is what the runners now know about the finish. SourcesBB
The Robot Games closed on football, fighting and a reordered medal table
The second World Humanoid Robot Games wrapped up Wednesday in Beijing with the 5-a-side football final, the freestyle fighting final and the closing ceremony, after a Tuesday given to dancesport: street dance, ballroom and cheerleading finals drew 108 registered teams against 29 at last year's games, alongside table tennis, tug of war, weightlifting and the standing high jump. The competitive story inverted last year's. AgiBot claimed the overall medal lead after the weekend by its own accounting, Beijing's state-backed innovation center swept the track golds, and Unitree, last year's medal leader, had no athletics medals through Sunday, the same week its stock gave back 200 billion yuan. A full closing medal table had not reached English-language wires this morning. China Daily's tally of the event itself: 2,056 robots, 666 teams, 16 countries, 51 events. SourcesB
Nvidia bought into the company holding 15 gigawatts of Texas power
Nvidia took a direct equity stake in Lancium, the Blackstone-backed developer whose Texas portfolio spans 4 gigawatts of leased capacity and a powered-land pipeline above 15 gigawatts, and made that portfolio a deployment platform for its full AI-factory stack. The announcement leans on two Nvidia systems: DSX MaxLPS, which claims up to 40% more per unit of power budget, and DSX Flex, which manages loads against grid conditions. Neither side disclosed the check; earlier reports of up to $3 billion for about 20% were not confirmed and stay unconfirmed. Place it on the week's map: Texas froze new data-center approvals, PJM wants new loads to bring their own generation, and the binding constraint on compute has moved from chips to interconnection. Nvidia already capitalizes its model customers and its clouds. Now it capitalizes the power layer too. SourcesAB
o3 left ChatGPT on schedule
OpenAI removed o3 from ChatGPT on Wednesday, closing the 90-day sunset announced May 28th. The keeps the model until December 11th with gpt-5.6-sol as the designated replacement, o3-pro stays in ChatGPT for Pro, Team, Enterprise and Edu plans, and o3-mini follows on October 1st. The bookkeeping fact worth keeping: with GPT-4o, GPT-4.1, GPT-4.5, o4-mini and now o3 all retired this year, OpenAI has withdrawn more models in 2026 than in all previous years combined. The model that introduced most of the public to visible reasoning lasted 16 months in the product. Deprecation at that cadence is a real cost for anyone who built on the retired behavior, and it compounds the argument enterprises keep making with their routers: depend on the interface, never the model. SourcesA
Seoul rose into the print, on its chipmakers' own money
The Kospi closed Wednesday at 6,808.21, up 0.97% for a second straight gain, with Samsung Electronics up 1.75% to 261,500 won and SK Hynix up 0.6% to 1,688,000 won. The support is partly self-supplied: SK Hynix has been executing its 40 trillion won ($28.6 billion) buyback-and-cancel program since last Thursday, and Samsung repurchased shares this week as well. Institutions net bought 760.37 billion won of Korean shares Wednesday while foreigners and retail sold a combined 2.36 trillion. After Monday's 3% washout, the memory complex has spent two sessions buying back its own tape ahead of the one earnings call that decides whether the demand story carries into 2027. SourcesBA
OpenAI banned a Russian operation running a fake Israeli think tank
OpenAI disrupted a covert influence operation run from Russia through VPNs, which used ChatGPT to write English-language posts for Substack, Telegram, X, Facebook and LinkedIn promoting the "International Burke Institute," a fabricated Israel-based think tank whose "sovereignty index" ranked Russia favorably. Operators explicitly prompted the model to scrub linguistic tells of Russian authorship, and 34 of the 36 articles attributed to the institute's experts between September 2025 and May 2026 were copied from other sources, some under fake author credits. Reach was minimal; the posts drew few views. The detail that stays useful is the shape: not bot swarms, but a fake institution with a fake metric, laundered through real platforms, built by people who understood that the credibility layer, not the prose, is what a model cannot generate. SourcesAB
Google took Gemini vertical: lawyers and banks the same day
Google Cloud launched Gemini Enterprise for Legal in preview Tuesday, with Cleary Gottlieb, Freshfields, Weil and Williams & Connolly as launch customers, and a financial-services counterpart the same day; healthcare and life sciences are named as next. The legal product bundles four things: packaged skills for contract review, redlining, regulatory scanning and legal research; connectors into document and case systems that inherit existing permissions; that complete work rather than suggest it; and a partner layer spanning the big consultancies and legal-tech vendors. Pricing was not stated. The strategic read is simple: the enterprise checkbook opens by profession, not by model, and the startups that proved legal AI revenue exists now have a selling the same category with the customer's cloud bill as the wedge. SourcesAB
The encrypted reasoning a frontier API returns can be decrypted by anyone
A paper making the rounds this week, from authors including Alexander Panfilov, Maksym Andriushchenko and former Google DeepMind researcher Ilia Shumailov, shows that the encrypted reasoning blocks major providers return for conversation resumption are interchangeable across sessions, users and models within a provider's ecosystem. Inject a stronger model's encrypted trace into a weaker model's conversation and the weaker model decrypts it for you, defeating the anti-distillation logic of hiding reasoning at Anthropic, OpenAI and Google alike. From 315,320 publicly shared reasoning blocks the authors recovered 367 pieces of personal information and 182 credentials, and they show the blocks also work as a prompt-injection carrier. Disclosure was responsible and mitigations are proposed. The industry hid the to protect it, then handed it back to every user in a sealed envelope any sibling model will open. SourcesA
Anandkumar's next act is a physics AI with no transformer in it
Caltech's Anima Anandkumar, formerly Nvidia's senior director of AI research, and Benedikt Jenik unveiled Accelerated Understanding, an enterprise AI company built on rather than transformers, in a Reuters exclusive. The two walked away from Prometheus, the Bezos-backed physical-AI startup, to do it. The company claims its system ingested 5 trillion data points in a single prompt in testing, a company-supplied figure with no independent verification, and targets chip design, robotics, weather prediction and geological analysis. The interesting bet is architectural: neural operators learn mappings between functions rather than sequences of tokens, which is a different claim about what intelligence over physical systems requires. Every frontier lab is scaling one architecture. The hedge against that monoculture is now being funded as a company. SourcesB
The five-hour caps came back for the $20 tier
OpenAI's five-hour usage windows returned for Plus subscribers on Codex and ChatGPT Work on Tuesday, ending six uncapped weeks that began July 13th. Pro plans at $100 and $200 stay uncapped, and OpenAI's Thibault Sottiaux framed the change as a way to "smoothen the load on our compute." Set it next to Monday's adoption numbers, where fewer than 1% of individual subscribers touch the agent products at all: the constraint on agents at the consumer price point is not demand and not capability, it is that serving long-running agent sessions at $20 a month loses money whenever people actually use them. Rationing at the cheap tier is what the unit economics of agents look like from below. SourcesB
Cerebras slipped under its IPO price, and ARK kept buying
Cerebras traded between $180.44 and $192.08 Tuesday, closing near $184, below the $185 it priced at in May and about 14% under where it stood before its August 12th report, a quarter in which revenue grew 74.3% year over year and full-year guidance rose to $880 to 890 million, the second raise since listing. The selling has one prominent counterparty: Cathie Wood's ARK funds have bought more than $60 million of the stock in August, most recently 93,290 shares. The first AI-hardware listing of the class of 2026 is now the class's first test of what a broken issue price does to the ones behind it, with insider unlocking through the fall and SB Energy's next in line. SourcesBB
Claude's memory now spans chat and Cowork
Anthropic unified Claude's memory across chat and Cowork, so what the assistant learns in one surface carries into the other, with memory organized into topics a user can view and edit and a setting that walls off sensitive subjects. It is a small release with a large direction: the labs have concluded that memory, not raw capability, is where assistant lock-in actually forms, because a model that knows your context is expensive to leave even when a rival's benchmark is better. Instinct priced that same insight as a perpetual license to user data last week. Anthropic is pricing it as a settings page. The gap between those two answers is where the assistant market's trust fight happens. SourcesB
Ox Alpha's free week ends today, with nobody's name on it
OpenRouter's listing for the now carries an end date: "Going away August 26, 2026." Six days of free frontier-scale service, a million-token , input and tool calling, heavy traffic from agent harnesses, and the provider is still anonymous, with the tokenizer evidence pointing at Microsoft's model lineage and the timing theories still pointing at Zhipu. If the model disappears today unclaimed, someone ran a week-long public load test on real developer workloads, paid for it in free tokens, and kept the results. That data changed hands in exactly one direction. SourcesA