AI Industry Daily Briefing — September 26, 2026

The Executive Read

Friday put two different responses to the same anxiety side by side. In Washington, Bernie Sanders and Greg Casar introduced the most aggressive piece of AI legislation Congress has seen: a permanent ban on building artificial superintelligence, a pause on frontier development until a new cabinet-level Department of Artificial Intelligence writes safety rules, and, for violators, what Sanders called the “corporate death penalty” alongside prison terms of up to 20 years — the same penalty, he noted, as unlawfully building a nuclear weapon. It lands one day after 27 state attorneys general told Congress the industry can’t be trusted to police itself (Edition No. 25), and it will not pass a Republican-controlled Congress that has struggled to agree on far narrower AI rules. But it moves the outer edge of the debate, and it sets a marker for whatever does pass. The industry, for its part, spent the same 24 hours doing the opposite of pausing. Microsoft rebuilt Copilot from the ground up, merging its consumer chat product, its coding tool and a new autonomous “Autopilot” agent into a single subscription with usage-based billing for the agentic pieces — Satya Nadella called it “a new OS for work.” Akamai signed a seven-year, $11.6 billion contract to run Anthropic’s CPU workloads, with an option that could push the deal to $20 billion and a warrant handing Anthropic a stake of up to 5% of Akamai’s stock. And Elon Musk said SpaceXAI’s Colossus 2 data center could more than double its Nvidia chip count by year-end, from roughly 550,000 to as many as 1.2 million, arriving in 220,000-chip waves timed to how many fiber-optic cables fit into a single network switch. None of these three moves waited for Washington’s answer on how much oversight the industry needs; each was, in its own way, a bet that scale now is worth more than caution later. Congress wants to define the ceiling. The industry spent Friday building past whatever ceiling exists today.

---

Top AI Headlines

Microsoft merges Copilot Home, Code and Autopilot into a single product it calls “a new OS for work”

What happened. Microsoft announced its biggest Copilot redesign to date on September 25, according to the company’s own blog post. The new Copilot combines three modes: Home, which merges the existing Chat and Cowork experiences and now embeds Word, Excel and PowerPoint directly inside the app as “Office in Copilot”; Code, which lets any employee — not just developers — build small dashboards, widgets and automation workflows from natural-language descriptions, running on the same technology as GitHub Copilot in a sandboxed environment; and Autopilot, a persistent agent, previously called Scout, that has its own identity and memory and keeps working on recurring tasks inside Microsoft 365 without being re-prompted, surfacing inside Teams and Outlook. CEO Satya Nadella posted on X that Microsoft is “building Copilot as a new OS for work that spans every model, every form factor, and every task.” Pricing is splitting in two: a fixed-cost user subscription covers Chat, Office apps, Outlook and Teams, while the heavier agentic features — Cowork, Code and Autopilot — move to usage-based billing that is off by default, alongside new spending-policy and per-user credit-tracking tools Microsoft is calling “FinOps for AI.” Code and Home begin rolling out to Microsoft’s Frontier preview program in the coming weeks; Autopilot enters private preview by the end of September.

Why it matters. This is Microsoft absorbing its own consumer AI ambitions back into the enterprise business, six months after positioning Copilot Home as a standalone competitor to ChatGPT and Claude’s consumer apps — a tacit admission that the company’s advantage is distribution inside Word, Excel and Teams, not a standalone chatbot brand.

Business implication. Enterprise customers should expect their Copilot bills to become more variable as agentic features shift to usage-based pricing; the new FinOps tooling exists because Microsoft expects that variability to be large enough that finance teams will need to actively manage it, not because Microsoft is offering it as a convenience.

Sources: Microsoft, official blog post · Satya Nadella, official statement on X · GeekWire

---

Akamai signs a seven-year, $11.6 billion cloud deal with Anthropic — with an option to grow to $20 billion

What happened. Akamai Technologies disclosed in a September 25 SEC filing that it has signed a seven-year, $11.6 billion agreement under which Anthropic will run its growing CPU computing workloads on Akamai’s distributed cloud infrastructure — the largest contract in Akamai’s history. The deal carries an option for Anthropic to expand its commitment by another $9 billion, which would bring the total to roughly $20 billion. As part of the agreement, Akamai issued Anthropic a warrant to purchase non-voting convertible preferred stock equal to up to 7.7 million shares of Akamai common stock, or about 5% of shares outstanding, at an exercise price of $111.33 per share; roughly 2% vests once Anthropic makes its first payment, with the rest tied to further spending in $3 billion increments. Akamai said the contract will drive about $5.5 billion in associated capital expenditures and that it expects to raise its 2026 capital spending by roughly $1.7 billion to secure supply-chain components, including memory, with no change to 2026 revenue guidance. CEO Tom Leighton said in the filing that “Anthropic is advancing the AI revolution and we are thrilled they chose Akamai’s capabilities for building and operating AI infrastructure at scale.” Akamai shares rose about 15% in after-hours and Friday trading.

Why it matters. This is a large AI infrastructure commitment built around CPUs rather than the Nvidia GPUs that dominate most AI capex headlines, and it gives Anthropic a direct equity stake in one of its own infrastructure vendors — a structure that ties the vendor’s stock price to the customer’s continued spending, similar to arrangements OpenAI has struck with chip and cloud partners.

Business implication. Akamai’s stock move shows investors will reward a company for landing a large AI compute contract even outside the GPU narrative; other CDN and edge-infrastructure providers now have a template — and a competitive threat — for pitching AI labs on CPU-workload capacity that Nvidia-focused clouds aren’t built to handle as efficiently.

Sources: Akamai Technologies, SEC Form 8-K exhibit · TechCrunch · The Motley Fool

---

Sanders and Casar introduce a bill to permanently ban artificial superintelligence and create a federal AI department

What happened. Senator Bernie Sanders (I-Vt.) and Representative Greg Casar (D-Texas) introduced the Ban Artificial Superintelligence Act on September 23, according to their joint press release. The bill would permanently prohibit developing or deploying “artificial superintelligence,” which it defines as an AI system that exceeds human cognitive performance across most domains or has sufficient capability to disempower humanity, including by overthrowing the federal government. It would also temporarily pause development of the most advanced frontier systems until a new cabinet-level Department of Artificial Intelligence, advised by outside experts, establishes binding safety rules covering testing, monitoring and the removal of dangerous capabilities from deployed models. Violations would carry what the sponsors called a “corporate death penalty” — dissolution of the offending company — plus up to 20 years in prison for individuals, a penalty Sanders compared directly to the one for unlawfully building nuclear weapons. The bill directs the new department to work with the State Department on international agreements to prevent superintelligent AI from being developed anywhere. Sanders said “the future of humanity cannot be left in the hands of a handful of Big Tech oligarchs”; Casar said “cutting-edge AI technology is less regulated than the average food truck.” Roll Call and other outlets covering Capitol Hill note the bill faces long odds in a Republican-controlled Congress that has not coalesced around even more modest AI legislation.

Why it matters. This is the first federal bill to propose an outright ban on a category of AI system rather than disclosure, testing or liability rules, and it arrives the same week Google, OpenAI and Anthropic are reportedly assembling a self-funded standards body of their own (Edition No. 25) — the gap between the two proposals is now about as wide as US AI policy debate gets.

Business implication. The bill is very unlikely to become law this Congress, but it shifts where the center of the debate sits; frontier labs’ own self-regulatory pitch now has to be measured not just against state attorneys general demanding federal testing, but against a sitting senator’s proposal to make what some labs describe as their long-term goal a felony.

Sources: Sen. Bernie Sanders, official press release · Rep. Greg Casar, official press release · NBC News

---

Musk says SpaceXAI’s Colossus 2 could more than double to roughly 1.2 million Nvidia chips by year-end

What happened. Elon Musk said on X on September 25 that Colossus 2, the Memphis data center run by SpaceXAI (the merged SpaceX-xAI company that began trading as SPCX in June), currently runs 110,000 Nvidia GB200 chips and 440,000 GB300 chips — about 550,000 total. He said another 220,000 GB300s will be “fully operational next week,” a further 220,000 in November, and, “if we get lucky,” another 220,000 by late December, which would bring the total to roughly 1.2 million chips if every tranche arrives on schedule. Musk said the recurring 110,000-chip increments reflect “the number of fiber optic cables that can be plugged into a central switch,” a networking constraint rather than a chip-supply one. The post came in reply to a question about Colossus 2’s current scale and was not accompanied by a formal company announcement or a capital-spending figure.

Why it matters. This is one of the largest single-cluster GPU buildouts anyone has stated a number for, and it is a direct, if informal, data point on Nvidia demand at a moment when other buyers — Oracle among them — are hitting real-world limits on power and permitting rather than chip supply (Edition No. 25); Musk’s own account of the timeline is the only public source for these figures and has not been independently verified.

Business implication. A cluster approaching 1.2 million high-end Nvidia chips implies power and cooling commitments on a scale most utilities have not previously had to plan for in one location; investors and Memphis-area residents tracking Colossus’s air-quality and grid impact (Edition No. 1) have a much larger number to weigh against those effects if the buildout proceeds as described.

Sources: Elon Musk, official post on X · Bloomberg, via Yahoo Finance · Benzinga

---

Model and Product Updates

Two Anthropic physicists used Claude to compute a nine-loop scattering amplitude in N=4 super Yang-Mills theory, a step beyond the previous eight-loop record, Anthropic disclosed September 25 in a guest research post. N=4 super Yang-Mills is a simplified “toy model” theoretical physicists use to test calculation techniques that later apply to real particle-physics predictions; a scattering amplitude describes the probability that particles collide and produce a given outcome, and each additional “loop” represents a more precise, and exponentially harder, level of calculation. Researchers Liam Fitzpatrick and Siddharth Mishra-Sharma gave Claude the problem and let it work largely unsupervised using two independent methods, which produced matching results; the full computation cost between $1,000 and $2,000 and ran on 96 processors for about a week. The challenge was set by Matt von Hippel, a former theoretical physicist who posed it publicly on August 7. Von Hippel, who reviewed the result, was careful to note that Claude used established calculation methods rather than inventing new ones: “Claude used known methods, with a bit more compute than people had tried to use before.” (Anthropic, official research post)

Google’s Gemini 3.8 Live with Live Avatar became generally available inside Gemini Enterprise on September 24, according to Google’s own blog post. The feature pairs Google’s existing real-time voice model with a low-latency video layer that generates an animated avatar — a real person’s likeness or a designed character — whose mouth movements and facial expressions stay synced to speech at 24 frames per second across 97 languages, intended for customer-service and interactive-kiosk deployments. Enterprises can choose from a library of pre-built avatars; creating a custom one requires a separate verification and allowlisting process. Google says all generated audio and video carries an imperceptible SynthID watermark so the output can be identified as AI-generated. (Google, official blog post)

---

Regulation and Policy Watch

Trump and Xi Jinping concluded Xi’s three-day Washington state visit September 25 with an extension of their trade truce to January 10, 2027, and an agreement to hold a follow-up summit in Shenzhen in November tied to the APEC gathering, where the two leaders said they would continue talks on AI risks and benefits. According to CNBC’s and Bloomberg’s coverage of the summit’s close, US Trade Representative Jamieson Greer said limited additional trade understandings on agricultural exports and medical devices would be detailed the following Monday, and the two sides discussed but did not formalize a channel for sharing alerts about AI-related security incidents, a step first floated Wednesday (Edition No. 24). Advanced semiconductor export controls were not reopened. Both leaders also confirmed they would attend each other’s remaining 2026 multilateral summits — APEC in Shenzhen in November and the G20 in Miami in December. The Sanders-Casar bill (above) and the ongoing dispute between the industry’s self-regulatory push and 27 state attorneys general (Edition No. 25) remain the more consequential near-term US policy threads; Friday’s summit produced a date for further AI talks, not a framework. (CNBC · Bloomberg)

---

Emerging Startup Radar

OpenEvidence, an AI clinical-search platform used by physicians, raised $250 million at a $15 billion valuation, up from the $12 billion it reached in a January 2026 round, according to Axios’s reporting and corroborating coverage. The round was led by Andreessen Horowitz and Byers Capital, with participation from undisclosed hospital-system partners; Nelson Advisors described it as an “alternative financing round” rather than a standard equity raise. The company, whose total funding now approaches $950 million over roughly 19 months, is using the round to fund a pivot into oncology drug development — CEO Daniel Nadler told Forbes the company aims to put its first drug candidate into clinical trials before year-end, targeting rare cancers that large pharmaceutical companies “are not doing, and maybe can’t do,” with three to five additional candidates to follow. OpenEvidence also announced an expanded partnership with Memorial Sloan Kettering Cancer Center to integrate the hospital’s OncoKB precision-oncology database into its platform. (Axios · Dealroom)

---

AI Infrastructure and Market Signals

The 10-year Treasury yield closed at roughly 5.18% Friday, September 25, its highest level of the week and its highest since 2007, even as equities rallied — a reminder that Friday’s two largest AI infrastructure commitments, Akamai’s $11.6 billion Anthropic deal and Musk’s plan to more than double Colossus 2’s chip count, were both struck against a borrowing environment that keeps getting more expensive, not less. Akamai itself said the Anthropic contract will require roughly $1.7 billion in additional 2026 capital spending to secure memory and other supply-chain components, a smaller-scale version of the same financing pressure that pushed Oracle to invoke force majeure on its New Mexico Stargate site a day earlier (Edition No. 25). None of Friday’s infrastructure announcements came with financing details showing how they will be funded if credit conditions tighten further.

---

Public Investment Watchlist

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

US stocks rose Friday, September 25, closing out a volatile week with gains: the Dow Jones Industrial Average rose 478.64 points, or 0.93%, to 51,828.62, snapping a three-week losing streak; the S&P 500 gained 0.51% to 7,743.41; and the Nasdaq Composite rose about 0.5% to 27,068.72. Akamai shares surged roughly 15% on its Anthropic cloud deal, and Microsoft rose about 4% following its Copilot overhaul announcement. The gains came despite the 10-year Treasury yield climbing to roughly 5.18%, its highest close since 2007, as falling oil prices — on optimism about a possible reopening of the Strait of Hormuz — helped offset rate pressure. All three major indexes finished the week higher. None of Friday’s moves guarantee anything about next week’s session, which opens with markets still digesting both the Trump-Xi summit’s conclusion and Friday’s AI infrastructure announcements.

Sources: Yahoo Finance · The Motley Fool · TheStreet

---

Watchlist

1. Whether the Ban Artificial Superintelligence Act gets a hearing or any Republican co-sponsors, given that Congress has not coalesced around even narrower AI legislation, and how frontier labs respond publicly to a bill that would criminalize what some describe as their long-term research goal.
2. Whether Microsoft’s usage-based billing for Cowork, Code and Autopilot produces real cost increases for enterprise customers once the features leave preview, and how that pricing compares to Anthropic’s and OpenAI’s own enterprise agent pricing.
3. Whether Akamai’s $11.6 billion Anthropic contract expands toward its $20 billion ceiling, and whether other CDN or edge-infrastructure providers announce comparable CPU-workload deals with frontier labs.
4. Whether SpaceXAI’s Colossus 2 chip additions arrive on the schedule Musk described — 220,000 GB300s “next week,” another 220,000 in November, and a final tranche “if we get lucky” in December — given that Musk’s public timelines for AI infrastructure have previously slipped.
5. What emerges from the announced Trump-Xi AI dialogue at November’s APEC summit in Shenzhen, the second AI-specific venue floated this week after Wednesday’s discussed incident-notification channel, and whether either produces anything beyond a further meeting date.
6. The still-undecided jury verdict in Andersen v. Stability AI, closing in on three weeks since testimony began September 8, with no outcome reported as of this writing.
7. 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.