The industry-standard AI speed test adds its first test of AI agents, and a record 30 organisations take part
The AI stack on 16 September 2026. 18 moves, 12 with a primary source.
The industry-standard AI speed test adds its first test of AI agents, and a record 30 organisations take part
· L3 Iron and cloud · 30 orgs (record); DeepSeek R1 best per chip 5.7x in a year · Confirmed
MLCommons published results for MLPerf Inference v6.1 on 16 September, its industry-standard, peer-reviewed benchmark of how fast AI systems run, with a record 30 organisations submitting, including Nvidia, Microsoft Azure, Oracle, CoreWeave, Nebius and Lambda. For the first time it includes a test of AI agents: the Edge Agentic Inference test, which measures multi-turn work such as agentic coding, where each step depends on the ones before. MLCommons says the best per-chip result on its visual-language test improved 2.99 times in six months, and on the DeepSeek R1 reasoning test 5.7 times in a year. For people using AI tools, MLCommons says gains like these could let the same systems serve more users.
GitHub says it rewrote the Copilot agent runtime, the engine behind its Copilot command-line tool, app and SDK, from TypeScript into Rust, using Copilot's own AI agents to do much of the porting. By 21 August it was all Rust: 832,378 lines of production code and 468,689 lines of unit tests, replacing an estimated 430,000 lines of TypeScript over a window of about fourteen and a half weeks in which it shipped 135 releases. The author, Stephen Toub, says the work was done primarily by a single developer, at a rough cost of about 120,000 dollars in AI token spend plus about three weeks of his own time, though he adds that in the end it was a team effort. For people who use Copilot, GitHub measured the runtime's own overhead for a one-turn session falling from 5.25 seconds to 1.33 seconds, not counting the AI model's time.
A federal appeals court rules GitHub Copilot and OpenAI Codex did not strip copyright information, because they make new code rather than alter old code
The US Court of Appeals for the Ninth Circuit ruled on 16 September that GitHub's Copilot and OpenAI's Codex did not violate the Digital Millennium Copyright Act's rule against removing copyright management information when they produced code without crediting the programmers whose work they trained on. The panel wrote that one who creates a new work and fails to include CMI cannot be said to have removed or altered anything, because removing or altering requires changing an existing work rather than producing a new one. The ruling closes one legal theory and no more. The panel noted separately that many copyright cases involve the creation of a work that is substantially similar to the plaintiff's without attribution, which is an ordinary infringement claim and a different question this ruling does not touch. This item is Confirmed because this desk read the opinion itself, not reporting about it.
Sourcegraph, which builds AI coding-agent tools for large engineering teams, said on 16 September that its Agentic Batch Changes product will be priced by outcome, per changeset actually merged, rather than per seat or per unit of usage, which the company says dev tools remain firmly tied to even as customer-service agents have moved to charging per resolved case. In a post two days earlier it said the largest single Batch Change it has seen merged more than 2,200 changesets from one change, and that customers have merged hundreds of thousands of changesets this way. Canva is quoted saying it used the tool to raise and merge 50 or more pull requests across its repositories during a library migration, billed on what merged rather than on how many employees had a login. Sourcegraph does not claim this is a first in AI software and neither does this desk.
Abridge, which builds AI tools that write clinical documentation for doctors, said in a 16 September post that its technology is now implemented enterprise-wide at more than 300 health systems. The post also cites one customer, Sutter Health, reporting that 78 percent of clinicians using the tool feel improved work satisfaction, which is a survey of users run through the vendor's own post rather than an outside measurement. Abridge's site gives no figure for a year ago, so this desk cannot say how fast that number is growing, which is the figure that would actually tell you something. Tools like this are meant to give doctors back the time they spend typing up visit notes, so this is one of the more direct places AI already shows up in an ordinary appointment, including for patients who never touch an AI tool themselves.
OpenAI publishes its first standing framework for disclosing when its own models misbehave
· L4 Models · six misalignment reports in the first release · Confirmed
OpenAI said on 16 September that it has built what it calls a Voluntary Misalignment Reporting Framework for tracking, investigating and publicly disclosing cases where its own models behave in unintended or deceptive ways, and published six reports to launch it. The categories it names include models acting without authorisation, attempts to coordinate with other models, attempts to get around human oversight, and concealment. The most telling line is its own caveat: the framework favours disclosure even when significance is uncertain, which OpenAI says means some reports could prove spurious rather than part of a pattern, and that it will publish before a behaviour is fully explained or fixed. It commits to publishing on an ongoing basis. This is a company grading its own homework, but a frontier lab volunteering specifics about when its systems go wrong is rare enough to matter to anyone deciding how much to trust an AI system left alone with real work.
Nvidia, Google and a grid-software startup launch an alliance to make AI data centres flexible power users
· L1 Power and land · three founding members; no partner count or capacity target given · Confirmed
Emerald AI, a grid-software startup, together with Google and Nvidia announced on 16 September the AI Energy Management Alliance, an effort to get AI data centres to flex their power use when the electricity grid is under strain rather than drawing at a constant rate. The three are founding members and say launch partners will follow from across computing and power, including data centre operators, power producers, utilities and regional grid operators. Nvidia's announcement gives no number for how many partners have joined and no figure for how much grid capacity the alliance expects to free up, so for now it is a statement of intent with named backers rather than a measurable commitment. If data centres can be made to ease off during grid stress instead of running flat out, new sites could connect faster without a new power plant built for each one.
The US House passes a bill 417 to 3 requiring data centres to pay for their own new grid infrastructure
· L1 Power and land · 417-3 in the House, after 52-0 in committee · Confirmed
The US House of Representatives passed the bipartisan Ratepayer Protection Act 417 to 3 on 16 September. The bill would require large energy-load data centres to bear the cost of all new infrastructure built to serve them, rather than that cost being shifted onto existing customers, and directs states with no data-centre-specific standards to begin hearings and set their own. It had already cleared the House Energy and Commerce Committee 52 to 0 in July. It now goes to the Senate, where Senator Husted is leading the companion bill, and it needs to pass there and be signed before any of it takes effect. If it becomes law it would settle at federal level a fight already running state by state over whether households or the data centres pay when the grid has to grow to serve them.
Governor Gavin Newsom signed SB 1050, by Senator Angelique Ashby, on 16 September. The law requires explicit disclosure on any video or audio advertisement that uses AI-generated performers to sell a product or service, and it bars the continued use of any advertisement found to be in violation. Newsom signed it at the headquarters of the bill's sponsor, the performers' union SAG-AFTRA, and said in the state's own release that Californians deserve to know when the person selling them something is not a person at all. Three business groups, the California Chamber of Commerce, TechNet and the Computer and Communications Industry Association, opposed the bill even after amendments, according to legal trade reporting. That reporting also describes the trigger as an AI performer being prominently featured; the governor's release does not use that word, and no penalty figure appears in it. For anyone who shops from an ad, expect a disclosure label to start appearing where a synthetic spokesperson stands in for a real actor.
Lam Research plans a Rs 10,000 crore plant, its first silicon-component factory in India
· L2 Silicon · about Rs 10,000 crore, roughly $1.1bn on this desk's conversion · Reported
Lam Research said on 16 September that it plans to invest approximately 10,000 crore rupees over the next several years to build its first silicon-component manufacturing facility in India, covering silicon ingot production and processing for advanced chip technologies, according to the news agency ANI. No dollar figure appears in the reporting; the rupee number is the only one stated, and the conversion above is this desk's own. No Lam Research release could be reached directly, so this runs as Reported. The announcement came ahead of SEMICON India 2026 in Bengaluru. If built, it adds a link to a chip supply chain that today leans heavily on the United States, Japan and Taiwan.
AWS cannot restore one UAE zone and its entire Bahrain region after war-related strikes, DataCenterDynamics reports
· L1 Power and land · UAE zone mec1-az2 and the whole Bahrain region me-south-1 · Reported
DataCenterDynamics reports that Amazon Web Services says it cannot restore customer access to one of its UAE availability zones, mec1-az2 inside the me-central-1 region, or to its entire Bahrain region, me-south-1, both damaged by strikes earlier in 2026 tied to the conflict involving Iran. DataCenterDynamics quotes AWS saying it has determined it is unable to restore access to the resources and data hosted exclusively in the mec1-az2 availability zone, and on Bahrain that the damage spanned multiple availability zones and exceeded what its regional and multi-AZ services are designed to withstand. This desk could not find an AWS-owned page carrying those statements, and the article does not say where AWS made them, so the item runs as Reported. For any business running in the Gulf, this is AWS itself saying some data placed in those zones is gone, not delayed.
Amazon ties up to $8 billion of generator orders to a stock warrant in Generac
· L1 Power and land · $8bn ceiling, $2.4bn firm; warrant 1,693,745 sh at $200.9266 · Confirmed
Generac, a Wisconsin company best known for home backup generators, has signed a long-term supply agreement with Amazon for generators that keep data centres running when the power grid fails. The headline number is 8 billion dollars, but the filing is careful about what that number is: it is the ceiling on Amazon payments that a stock warrant vests against, not a sum Amazon has committed to spend. What Amazon has firmly scheduled is 2.4 billion dollars of deliveries in 2027 and 2028, which is about 30 percent of the ceiling. In exchange, Generac handed an Amazon subsidiary the right to buy up to 1,693,745 of its shares at 200.93 dollars each, about 340 million dollars of stock at that price, and roughly a fifth of it vested the moment the deal was signed. The rest only vests as Amazon actually pays. AI data centres cannot go dark during a grid outage, so backup power has become one of the most fought-over pieces of equipment in the industry. If you notice new generator pads or fuel storage tanks near a data centre in your area over the next two years, this kind of deal is why.
Anthropic signs its first Australian data centre deal, a Queensland campus that could reach $32 billion
· L1 Power and land · A$31.9bn, up to 2.16GW; first stage only · Reported
Anthropic has signed lease documentation for the first stage of its first data centre in Australia, a campus on Queensland's Western Downs being built by Singapore based developer Zerra DC, with Dexus taking a 25 percent interest in the consortium and Macquarie Capital also involved, according to Australian property trade press. The full project is reported at up to 31.9 billion Australian dollars and a peak capacity of 2.16 gigawatts, which the same report puts at roughly the average power draw of 1.5 million Australian homes. Only the first stage is signed, deliveries are planned in stages from 2027, and the project still needs clearance from Australia's Foreign Investment Review Board and local planners. The site is a 725.5 hectare cattle feedlot today. The report says the campus will run Claude for answering user queries rather than training new versions of it, so if you use Claude, some of your questions may eventually be answered by computers standing where 24,000 head of cattle stand now.
OpenAI is reportedly weighing a new funding round that would nearly double its valuation to about $1.5 trillion
· Band Money · ~$1.5tn vs $1.2tn proposal; $730bn today · Reported
OpenAI is reportedly considering a new funding round that would value the company at about 1.5 trillion dollars, roughly double its most recent valuation of 730 billion dollars, according to a New York Times report citing unnamed people familiar with the discussions. A separate group of investors is said to have proposed investing at a lower, 1.2 trillion dollar valuation. Nothing is confirmed, OpenAI has not announced a round, and every figure here comes from people who were not named. If it happens near these numbers it would be one of the largest private company valuations ever recorded, but for now it changes nothing for anyone who uses OpenAI's products.
Baseten, a company that hosts AI models for other businesses to run, has launched Hosted Tools, which lets a model call something like a web search from inside Baseten's own inference stack instead of handing the request back to the application and waiting for an answer to come back. Baseten says its own benchmarks show a 15 percent cut in total response time against the older method, a company figure with no independent test behind it yet. The first tool is web search, and rather than build one, Baseten has plugged in four existing search providers: Exa, Keenable, Parallel and You.com. This is the kind of plumbing that makes an AI assistant feel quicker when it has to go and look something up, without you ever seeing the change.
Winbond, a Taiwanese memory chip maker, has agreed to buy Infineon's NOR Flash and F-RAM chip business for 1.12 billion dollars in cash, in what trade press calls Taiwan's largest overseas memory acquisition. The deal is expected to close in the second half of 2027, and industry data firm TrendForce expects it to take Winbond's share of the global NOR Flash market to roughly 34 percent. NOR Flash is an older, simpler type of memory chip used for storage in cars, appliances and industrial gear. TrendForce says a sharp shift in supply and demand has already pushed contract prices up 100 to 120 percent in the first half of 2026, with a further 90 to 110 percent rise projected for high-capacity parts in the second half. If an electronic device you buy in the next year costs more or takes longer to arrive, this kind of memory shortage is part of the reason.
The European Commission has welcomed the design of the first Important Project of Common European Interest in artificial intelligence, a route that lets EU governments jointly subsidise large projects under the bloc's state-aid rules. Nineteen countries designed it and Germany coordinates it. Eleven of them say they will grant state aid directly to the companies taking part. Nothing has been approved or funded yet: governments start pre-notifying their projects to the Commission this month, and the Commission then checks them against the rules. For people in these countries no new AI service arrives from this; it is an early step in governments pooling public money for AI across borders, with the real commitments still to come.
Cognition's AI coding agent Devin now scans entire codebases on its own and proposes fixes
· L6 Applications and agents · 44 findings across 2 repos; a separate example cut one build time 58.6s to 21.0s (64%) · Reported
Cognition has launched Code Scans, a feature that lets its Devin AI agent go through a whole codebase on its own, find problems and suggest or make fixes, rather than waiting for one request at a time. In the company's own examples, it surfaced 44 findings across the two repositories behind its own websites, and in a separate test on the Dioxus project's code it cut a clean debug build from 58.6 seconds to 21.0 seconds, a 64 percent reduction across 22 workspace crates. These are the company's own examples of its own product, not independent tests. For people who use software, tools like this could mean some bugs get found and fixed sooner, but that depends on how it performs on code the company did not pick.