AI Industry Daily Briefing — September 19, 2026

The Executive Read

For a week, the industry’s safety debate ran on speeches — pledges to coordinate, essays proposing standards bodies, a King asking executives whether they had “sufficient means of control.” Friday and Saturday, three different institutions turned specific numbers loose on the same question. Anthropic and Accenture put a combined $2 billion behind a program that gives outside evaluators from Accenture’s Faculty unit employee-level access inside Anthropic — the ability to watch models train, sit in on deployment decisions, and talk directly to staff, rather than review whatever Anthropic chooses to show them. California Governor Gavin Newsom, not waiting on Washington, signed an executive order Friday giving his administration two months to draft a mandatory AI “kill switch” and independent on-site verification requirements for frontier labs, citing a recent security breach at Hugging Face as the kind of loss-of-control incident current law doesn’t clearly cover. And a decade-old copyright fight produced its own kind of hard evidence: unsealed filings in The New York Times’ suit against Microsoft and OpenAI quoted a Microsoft director of applied science calling AI scraping “an astonishing theft of unprecedented proportions” and, in his own words, “the largest theft of labor in human history” — a company’s own internal assessment, surfacing in litigation, of the industry it now leads. None of these three stories resolves anything by itself. But together they mark a shift from labs and governments describing what oversight should look like to specific dollar figures, specific legal deadlines, and specific internal admissions landing in the public record. Underneath the governance news, the infrastructure buildout kept running on its own schedule and its own money: Crusoe, a five-year-old company that started by burning stranded natural gas to mine cryptocurrency, raised $3.9 billion at a $30.9 billion valuation to keep building AI data centers, and Huawei used its annual conference in Shanghai to unveil a new AI chip cluster explicitly aimed at reducing China’s dependence on Nvidia. The pace-of-development argument hasn’t been settled by anyone. The capital keeps moving regardless.

---

Top AI Headlines

Anthropic and Accenture commit $2 billion combined to give outside evaluators access inside Anthropic

What happened. Anthropic and Accenture announced a partnership Friday, September 18, that the two companies call “embedded evaluation”: Accenture’s specialist AI unit, Faculty, will place evaluators inside Anthropic with what Anthropic describes as employee-level access — the ability to observe models during training, see the internal decisions that govern model development and deployment, communicate directly with Anthropic staff, and report incidents or identify blind spots as they find them. Anthropic and Accenture each expect to invest at least $1 billion over the next five years. Anthropic said the arrangement is non-exclusive and that it is in discussions with the AI-safety nonprofit METR and other evaluators about piloting similar embedded arrangements, funded independently of Anthropic. In its own announcement, Anthropic said “independent embedded evaluators do not reduce our accountability, but help to make it more verifiable. The safety of our models remains our responsibility.”

Why it matters. This moves the outside-verification idea Google DeepMind’s Demis Hassabis proposed in the abstract this week (Edition No. 18) from a policy essay into a funded, running arrangement — the first time a frontier lab has paid an outside organization at this scale for standing access to its internal operations, rather than a one-time audit or a published self-report like Anthropic’s own pace-of-development metrics two days earlier.

Business implication. Enterprises weighing safety claims from Anthropic against those of labs with no comparable arrangement now have a concrete example to ask for — not just published numbers, but paid, ongoing, internal access for an outside party; whether Accenture’s evaluators are independent enough to catch what Anthropic itself would miss is a question procurement teams can’t yet answer, since the program is one day old.

Sources: Anthropic, official announcement · TechCrunch · CNBC

---

California’s Newsom orders a two-month sprint toward a mandatory AI “kill switch”

What happened. Governor Gavin Newsom signed an executive order Friday, September 18, directing California’s Government Operations Agency to accelerate implementation of two AI safety laws already on the books — SB 813, which sets up a framework for independent AI-verification organizations, and AB 1405, which creates a state registry of AI auditors — and to convene outside experts who have two months to recommend how to strengthen state law further. The order names four specific measures for the state to pursue: requiring onsite independent verification inside AI labs, verifying labs’ own safety frameworks and risk assessments, building a “kill switch” mechanism to force an emergency shutdown of a frontier model, and updating California’s definition of a reportable safety incident to explicitly include “loss-of-control incidents such as the Hugging Face attack.” Newsom said in a statement: “We’re not waiting to act — we’re going to speed up our work on substantial and responsible AI oversight before it’s too late.”

Why it matters. This is the most concrete state-level move toward a binding kill-switch requirement anywhere in the country, and it arrives while the Senate’s own frontier-AI bill remains unpublished and contested (see Regulation, below) — California is moving on its own two-month clock rather than waiting for federal negotiators who, per Commerce Committee Chair Ted Cruz, are still not sure they can agree on bill text this month.

Business implication. Labs operating in California — which is to say nearly every frontier lab — now face a live, dated process that could produce binding verification and shutdown requirements within roughly two months, well ahead of any comparable federal timeline; compliance teams should treat this as the nearer-term regulatory deadline, not the Senate bill.

Sources: Governor of California, official executive order announcement · CNBC · NBC News

---

Unsealed filings quote Microsoft director calling AI scraping “the largest theft of labor in human history”

What happened. Previously redacted filings in The New York Times’ copyright lawsuit against Microsoft and OpenAI, filed in the Southern District of New York, were unsealed and reported September 17. They quote Microsoft director of applied science Brent Hecht, in a January 2023 internal memo, describing AI content scraping as “an astonishing theft of unprecedented proportions” and, separately, “the largest theft of labor in human history.” A later Hecht presentation, from January 2024, warned that declining traffic to news publishers like the Times created a “doom loop” that would eventually “hurt the performance of our models and the entire web at the same time,” and that “it is highly unusual that an end-product threatens the economic foundations of its essential suppliers.” The filings also cite Microsoft’s own data showing Copilot’s “answer engine” cut click-through rates to the Times’ domain by as much as 93% compared with traditional Bing search results, and quote OpenAI’s head of ChatGPT, Nick Turley, describing AI products as an “existential threat” to publishers that is “largely substitutive” and getting more so. Microsoft has said Hecht’s comments represent “one employee’s individual perspective” that “does not represent the company’s views.”

Why it matters. This is the clearest internal admission yet, from inside one of the two defendants, that echoes exactly what the plaintiff is arguing in court — that AI products substitute for, rather than merely reference, the journalism they were trained on — surfacing not as a leak but as evidence in active federal litigation over whether that training was lawful.

Business implication. Any enterprise licensing content for AI training or relying on a fair-use defense for scraped data should read Microsoft’s internal characterization as a preview of how a jury might eventually be asked to view comparable behavior at other companies; the filings don’t decide the legal question, but they narrow the gap between what companies argue in public and what their own employees wrote in private.

Sources: Docket, The New York Times Company v. Microsoft Corporation, No. 1:23-cv-11195 (S.D.N.Y.), via CourtListener · TechCrunch, original reporting · Tom’s Hardware

---

Crusoe raises $3.9 billion Series F at a $30.9 billion valuation

What happened. Crusoe, which builds AI data centers and GPU cloud capacity, announced the initial close of a $3.9 billion Series F on Thursday, September 17, at a $30.9 billion post-money valuation, up from roughly $10 billion a year earlier. The oversubscribed round was co-led by Atreides Management, Mubadala Capital and Valor Equity Partners, with participation from more than two dozen other investors including Founders Fund, GIC, Nvidia, the Qatar Investment Authority, Radical Ventures, TPG, Fidelity, T. Rowe Price and Tiger Global. Crusoe said the capital will fund its AI factories, including large-scale data-center campuses and its modular Crusoe Spark units, plus its Crusoe Cloud service. The company disclosed metrics alongside the round: more than $140 billion in total contracted value across its platform, more than 6 gigawatts of gross contracted capacity against 1 gigawatt currently operational, 20-fold year-over-year growth in Crusoe Cloud bookings, and more than $100 million in contracted annual recurring revenue from managed inference. Crusoe CEO Chase Lochmiller said, “Getting there means controlling the infrastructure from electrons to tokens.”

Why it matters. Crusoe began in 2018 turning otherwise-wasted natural gas from oil wells into electricity for cryptocurrency mining; its jump to a $30.9 billion valuation, on the strength of a self-reported $140 billion contracted-value pipeline, illustrates how far capital has moved from betting on AI models themselves toward betting on the physical capacity — power, land, cooling — needed to run them.

Business implication. The gap between Crusoe’s 1 gigawatt of operating capacity and its 6-plus gigawatts of contracted capacity is the number to watch: contracted capacity is a forward promise, not built infrastructure, and enterprises signing multiyear compute deals with any infrastructure provider should ask what share of that provider’s pipeline is actually energized today versus promised for later years.

Sources: Crusoe, official announcement · TechCrunch

---

Huawei unveils new AI chip cluster, accelerates roadmap aimed at reducing reliance on Nvidia

What happened. Huawei used its annual Huawei Connect conference in Shanghai on Thursday, September 17, to unveil the Atlas 960E SuperPoD, which the company’s own announcement describes as the industry’s first AI computing cluster built on “near-packaged optics” — optical components placed physically closer to the compute chips to cut data-transfer losses. Huawei says the system scales to 4,096 NPUs (its term for AI accelerator chips), delivers 8 exaflops of computing power at FP8 precision, and is designed to train and run models with up to 10 trillion parameters. Huawei’s Hi-ONE optical technology, the company says, replaces roughly 48,000 conventional optical modules with 5,500 units, cutting power draw by more than 550 kilowatts and pushing system availability to 99.8%. Deputy Chairman David Wang said Huawei is “committed to building powerful AI infrastructure and going open source and open system to build out computing ecosystems.” Separate trade reporting, not confirmed in Huawei’s own materials, said Huawei is also moving up the launch of its next Ascend chip generation, the 960DT, to the first quarter of 2027, three quarters ahead of its original schedule.

Why it matters. This is China’s most concrete public answer yet to the export controls that have limited its access to Nvidia’s most advanced chips — a domestically engineered path to bigger AI clusters built around optical interconnects rather than the chip-to-chip communication standards Nvidia and its partners control.

Business implication. Enterprises and governments building “sovereign AI” capacity outside the US chip supply chain — the same argument Cohere and Aleph Alpha made in their merger this week (Edition No. 18) — now have a second data point, alongside Nvidia’s continued growth guidance, for how seriously to weigh a non-Nvidia compute stack; independent verification of Huawei’s performance claims remains scarce outside the company’s own materials.

Sources: Huawei, official announcement · The Next Web · Global Times

---

Regulation and Policy Watch

The Senate’s bipartisan frontier-AI bill is still unpublished, and its own chief sponsor isn’t promising a vote. Commerce Committee Chair Ted Cruz told reporters this week, “I would love to mark it up this month, if we can get bipartisan agreement. We’re making progress” — a more hedged statement than a firm commitment. Senator Amy Klobuchar, one of the negotiators, described the bill’s current focus as model approvals, government-expert review, and requirements to alert authorities of imminent harm; Senator Maria Cantwell, not among the lead negotiators, continues to frame the proposal as one narrow piece of AI oversight rather than comprehensive regulation, and is separately pushing for federal-laboratory scientists, not company self-testing, to check the riskiest models. Majority Leader John Thune said whether the bill gets a committee vote “remains to be seen.” The bill still has no public text, no bill number, and — per reporting compiled this week — no resolution on the objections a coalition of AI safety organizations raised to the current draft (Edition No. 18). California’s Newsom order (above) is now the clearer near-term deadline on frontier-AI oversight in the United States.

Sources: Semafor

---

Emerging Startup Radar

EVAS Intelligence, a Chinese startup building RISC-V-based AI computing chips, raised nearly RMB 2 billion (roughly $280 million) in a round that closed this week, pushing its post-money valuation to close to RMB 15 billion (about $2.1 billion), according to multiple Chinese financial outlets. More than 20 investors participated, including Huatai Innovation, Eastern Bell Capital, SMIC Capital and Tongfu Microelectronics. Founded in 2022, EVAS combines a TPU-like chip architecture with the open-source RISC-V instruction set; its first Epoch-series processors are already in volume production, and the company says proceeds will fund mass production, its next flagship chip, and international expansion. The figures come from PANews, ChainCatcher and Gasgoo’s Chinese-language coverage rather than a company release in English; treat the exact yuan-to-dollar conversion as approximate.

Sources: PANews · Gasgoo

---

AI Infrastructure and Market Signals

Amazon took a warrant stake in generator maker Generac as part of an $8 billion supply agreement for data-center backup power, disclosed in a Generac SEC filing September 16. Under the deal, Generac issued Amazon a warrant for up to 1,693,745 shares at a $200.9266 strike price, exercisable through September 2033; 307,954 shares vested immediately, with the rest vesting as Amazon’s purchases of backup generators for its data centers progress toward $8 billion in aggregate. The companies’ supply agreement calls for $2.4 billion of initial generator deliveries in 2027 and 2028. Generac shares jumped sharply in after-hours trading on the news.

Gartner raised its 2026 worldwide AI spending forecast to $2.7 trillion, up 49.5% year-over-year, in a report published September 16 — its third upward revision of the year, after forecasts of $2.5 trillion in January and $2.52–2.59 trillion in May. Gartner distinguished VP analyst John-David Lovelock said “the buildout of AI data center capacity is the largest infrastructure project humanity has ever undertaken.” The figure is Gartner’s own forecast methodology, not an audited industry total.

India launched Phase 2 of its Semiconductor Mission at SEMICON India 2026 in New Delhi on September 17, committing $13.5 billion over 12 years, up from an $8 billion first phase, with Prime Minister Narendra Modi citing 12 chip projects already approved. Applied Materials committed $5 billion over 10 years, and Lam Research pledged roughly ₹10,000 crore (about $1.2 billion) toward what organizers describe as India’s first semiconductor fab under the program.

Sources: Bloomberg · Gartner, official press release · Business Today

---

Public Investment Watchlist

Informational only. Nothing here is a recommendation to buy or sell.

US stocks closed mixed Friday, September 18, capping a volatile week: the S&P 500 rose 12.74 points, or 0.2%, to 7,650.50, the Nasdaq Composite gained 104.25 points, or 0.4%, to 26,552.55, while the Dow Jones Industrial Average fell 95.40 points, or 0.2%, to 51,682.64, according to the Associated Press’s market wrap. The 10-year Treasury yield climbed to 5%, from 4.94% late Thursday, after briefly topping that level earlier in the week for the first time since 2023; Brent crude settled at $103.87, down 0.9%. The AP attributed most of Friday’s stock weakness to pressure from the bond market rather than company-specific news, compounded by “triple witching” — the quarterly expiration of stock, index-option and futures contracts that forces traders to roll over or establish new positions. Individual movers included Nucor, down 6.3%, General Motors, down 5.1%, and Qualcomm, down 5.8%, against gains for Coinbase Global, up 11.7%, and Robinhood Markets, up 9.1% — none tied to specific AI news. Generac was the AI-infrastructure-adjacent outlier, jumping sharply after Tuesday’s disclosure of its Amazon supply deal (see Infrastructure, above), and Crusoe’s Series F, as a private company, produced no public-market move.

Sources: Associated Press, via Press Democrat

---

Watchlist

1. Whether Anthropic’s embedded-evaluation arrangement with Accenture produces a public finding — a flagged incident, a safety gap, a disagreement over access — that tests whether the program catches something Anthropic’s own review process would have missed.
2. What California’s two-month expert review, ordered under Newsom’s September 18 executive order, actually recommends, and whether it produces binding kill-switch and onsite-verification requirements before the Senate’s federal bill has even reached committee markup.
3. Whether the unsealed Microsoft filings in the New York Times suit affect the pending summary-judgment fight, and whether OpenAI or Microsoft respond beyond calling Brent Hecht’s comments an individual employee’s view.
4. Whether the Senate’s frontier-AI bill gets a committee markup this month, per Ted Cruz’s stated hope, or slips further as Senator Cantwell continues pressing for federal-laboratory testing over company self-testing.
5. The still-undecided jury verdict in Andersen v. Stability AI, more than a week into trial testimony with no outcome reported as of this writing.
6. The Third Circuit’s still-pending ruling in Thomson Reuters v. ROSS Intelligence, more than three months after June 11 oral argument on whether AI training on copyrighted material is fair use.
7. Whether Crusoe’s contracted-capacity pipeline — 6-plus gigawatts contracted against 1 gigawatt operational — converts to energized capacity on the timeline its investors are underwriting.