ENGIE will help supply up to 568 megawatts of Texas wind power for Oracle, one of several suppliers
The AI stack on 17 September 2026. 12 moves, 8 with a primary source.
The EPA repeals most of its 2024 carbon limits for fossil-fuel power plants, pointing to demand from AI data centres
· L1 Power and land · Most 2024 limits repealed; in force 16 Nov 2026 · Confirmed
The US Environmental Protection Agency published a final rule on 17 September repealing most of its 2024 Carbon Pollution Standards, the federal limits on greenhouse gas emissions from fossil-fuel power plants. It removes the emission guidelines for existing fossil-fuel steam power plants, the carbon-capture standards for coal plants making a large modification, and the carbon-capture standards for new gas turbines that run most of the time. Explaining how the outlook for electricity demand has changed, the rule says large users, "particularly those related to data center demand for artificial intelligence (AI) applications," are primarily driving a nearly 2.6-fold rise in projected growth of peak demand. The repeal takes effect on 16 November 2026, and in a separate proposal the EPA asks for comment on rescinding all greenhouse gas limits for these plants. For people living near these plants, fewer federal carbon controls will apply to them.
Nebius tells customers its on-demand Nvidia GPU prices will rise about 17 to 21 percent from 1 October
· L3 Iron and cloud · H100 +~17%; B300 +~21%; B300 +~56% over two rounds · Reported
Nebius, a cloud company that rents out Nvidia chips for AI work, has told customers it will raise prices for its on-demand GPU cloud services from 1 October, according to Proactive, a financial news service: about 17 percent more for Nvidia H100 instances, about 21 percent more for the newer Nvidia B300, and 25 percent more for AMD EPYC Genoa processors. The report says this is Nebius's second round of increases in a few months, taking the total rise for B300 instances to about 56 percent. This desk could not read a matching notice on Nebius's own site, so the item runs as Reported. For most people nothing changes directly; this is a price paid by companies that rent these chips.
ENGIE said on 17 September that, through a portfolio of wind energy resources serving the ERCOT market, it will help supply up to 568 megawatts of renewable electricity for Oracle's Texas operations. Other suppliers announced their own shares of Oracle's Texas wind buying separately, and this item's story brings those documents together. For people in Texas, it is one large buyer of computing locking in power from wind for its operations there.
Crusoe raises $3.9 billion at a $30.9 billion valuation to build AI data centres
· L3 Iron and cloud · $3.9bn initial closing at $30.9bn post-money; over $140bn contracted value · Confirmed
Crusoe, which builds and runs data centres purpose-built for AI computing, said on 17 September that it has closed the first part of a 3.9 billion dollar Series F at a 30.9 billion dollar post-money valuation. The round was co-led by Atreides Management, Mubadala Capital and Valor Equity Partners, with Founders Fund, GIC, Nvidia, the Qatar Investment Authority, Radical Ventures and TPG also named, and Crusoe calls it an oversubscribed initial closing of a round not yet complete. The company's own release also claims over 140 billion dollars of total contracted value across its platform, a figure it does not break down. That lands in the same week as Nscale's filing, which shows what such a number can hide: Nscale's comparable figure is 103.4 billion, of which 2.6 billion is live. Crusoe publishes no equivalent split. The confirmed valuation also lands within a billion of a 30 billion figure anonymous sources gave a financial outlet two weeks earlier.
India's IT and Electronics Minister Ashwini Vaishnaw announced the second phase of the national semiconductor mission, Semicon 2.0, at SEMICON India 2026 in New Delhi on 17 September, with an outlay of 1,27,500 crore rupees and a target of at least 200 chip-design startups and companies. The comparison he drew is the useful part: under the first phase more than 105 startups attempted chip design, and about 20 of those secured venture funding, worth roughly 800 crore rupees between them. So the target is not quite a doubling of attempts, and the first round's conversion rate into funded companies was about one in five. This desk could not locate a government document carrying these figures, so the item runs as Reported on the minister's on-record remarks as carried by Indian business press. India is trying to become a designer and maker of chips rather than only a buyer of other countries', and this is the more concrete of the two phases so far.
OpenAI launches a legal-research version of GPT-6 Astra, with early access at six large law firms
· L4 Models · 54.0% vs 38.7% on a 200-question benchmark, OpenAI's own run · Confirmed
OpenAI released Astra for Law on 17 September, a version of its GPT-6 Astra model set up for legal research, with access to what OpenAI calls a corpus of more than 230 million legal URLs. That includes case law from the Free Law Project's CourtListener, which OpenAI says covers more than 99.9 percent of published US precedential case law. On Vals AI's Legal Research Bench, using a 200-question private validation set, OpenAI reports its own run scored 54.0 percent against 38.7 percent for general-purpose Astra searching the web unaided. Those are OpenAI's numbers on someone else's benchmark, not a score Vals AI published. Early access firms named in the release are Sullivan & Cromwell, Ropes & Gray, Cooley, Skadden, Wachtell, Lipton, Rosen & Katz, and Latham & Watkins. Harvey, the legal AI company this site covered at a 15.5 billion dollar valuation earlier this month, appears as an API customer that will build on the model rather than as a rival. If you work at or hire a large law firm, another major vendor now sells a dedicated legal research tool, which over time tends to move what firms pay for one.
Huawei said on its own website that an Atlas 950 SuperCluster with 256,000 cards is already being deployed, part of a computing architecture it calls Peerium and unveiled at Huawei Connect 2026 in Shanghai. A successor, the Atlas 960, built on near-packaged optics, is under testing. This is a separate claim from Huawei's already-reported plan to bring its next AI chip forward: that one was about when a chip arrives, this one is about how many of today's chips it can wire together so they behave as one machine. Huawei gives no performance figure for the cluster and no customer for it. Bigger clusters acting as one computer generally mean models can be trained faster or answer more at once, which matters to the companies and governments buying the hardware rather than to anyone directly.
Anthropic says Claude now leads more than a quarter of its own AI research work
· L4 Models · 26% at the "AI leads" level; 1 in 47,000 agent decisions blocked · Confirmed
Anthropic published a report saying Claude leads 26 percent of its own AI research and engineering work, where leads means finishing most of a task end to end from a high-level prompt while a person supervises. More than 90 percent of the work sits at or above a broader band it calls AI collaborates, and that band contains the 26 percent rather than stacking on top of it. Anthropic says Claude is not operating fully autonomously on any measured part of that work, and that a monitor reviewed over a billion decisions from its research and engineering agents during August 2026 and blocked about 1 in 47,000. The page carries no publication date, so the date above is when this desk found it. This is self-measurement, not an audit, but it is a rare case of a frontier lab publishing a number for how much of its own research its model does, which is the number to watch for anyone judging how close these companies are to automating their own progress.
Seoul-area data centres are being rejected from the power grid at a 98 percent rate
· L1 Power and land · 10 of 522 applications approved (1.9%); 279 ruled unsuppliable · Reported
Korean reporting on material South Korea's energy ministry gave the National Assembly says that of 522 applications to connect data centres to the grid in the Seoul metropolitan area between August 2024 and March 2026, 10 have final approval, worth 1,010 megawatts, a rate of 1.9 percent. The funnel is where the story is: 279 applications, 53.4 percent, were ruled unsuppliable at the first technical review, 243 passed it, only 24 reached the main review, and 10 came out the other end. The same report, citing the trade outlet EPNC, puts data centre power demand in the region at 7,343 megawatts by 2027 against 4,718 megawatts of available supply, a shortfall it puts at about 36 percent. This desk read the Korean reporting directly but could not reach the underlying government submission, so it runs as Reported. For anyone planning an AI data centre near Seoul, a grid this closed pushes projects to other parts of Korea or out of the country.
Huawei says its next AI chip is ready three quarters early, targeting early 2027
· L2 Silicon · 960DT: Q4 2027 → Q1 2027; 960PR: one quarter early to Q3 2027 · Reported
Huawei's rotating chairman Tao Wang told the company's Huawei Connect conference that its next AI chip, the Ascend 960DT, is ready three quarters ahead of plan and will launch in the first quarter of 2027, with a second chip, the Ascend 960PR, arriving one quarter early. He said performance doubles, but gave no benchmark, no memory figure and no power number to support that, and Huawei's own website did not carry a readable version of the remarks at the time of writing, so this rests on Chinese trade press reporting from the keynote. If accurate, it is a sign China's leading chipmaker is pulling its own AI hardware roadmap forward as it works around restrictions on buying chips from Nvidia and other US companies.
Taiwan's economy ministry showed a silicon photonics engine called 3.2T at the Taiwan Innotech Expo in Taipei, developed by the government-backed Industrial Technology Research Institute with the ministry's support. The ministry says it doubles the transmission capacity of the previous 1.6T generation and can move about 400 gigabytes of data a second. Silicon photonics moves data with light instead of electrical signals, and the ministry pitches it at AI computing. The institute has lined up 20 companies across chip design, manufacturing, testing and packaging around it, and it was one of 64 technologies the ministry put on show. For ordinary people nothing changes yet; this is a demonstration, not a product on sale, and a sign that Taiwan is aiming its chip industry at AI computing.
Anthropic opens a special access programme letting biologists use its AI models for a wider range of research
· L4 Models · Access to Mythos, Opus, Sonnet; grants renew every 6-12 months · Confirmed
Anthropic has launched a Life Sciences Verification Program that gives vetted life-science organisations access to its Mythos, Opus and Sonnet models with safeguards it describes as more permissive for biology-related work. Biology is normally restricted more tightly because, as Anthropic puts it, it is often not possible to tell valid work, such as studying a virus to develop a vaccine, from an attempt to make a virus spread more easily. Applicants are checked on their research credentials, security standards and ethics oversight. Anthropic says dozens of organisations joined through early access and it expects to enrol hundreds in the first week. Higher-risk grants must be renewed every six months and standard grants once a year, and usage data is kept for 30 days so it can be monitored. For most people nothing changes directly, but it is a live test of whether an AI company can speed up medical research without opening the door to misuse.