Anthropic is asking the public to buy 825 million shares at $70 to $78 each, and it has separately agreed to sell $26 billion of non-voting stock at the same price to AMD, Broadcom, Google, Micron, Microsoft, Nvidia and TSMC. Several of those buyers are also its suppliers and partners. That is the thread through today’s issue: the AI industry is writing down, in filings, statutes and product rules, exactly who controls what.
Today’s Headlines
Anthropic Goes Public
- Anthropic Sets IPO Price Range of $70 to $78 a Share. The updated filing offers 825 million Class A shares to raise $57.75 billion to $64.35 billion, and an implied $1.404 trillion valuation at the top of the range, on a fully diluted basis. That is well above the $965 billion from May but under the $2 trillion reportedly targeted a couple of months ago. Run-rate revenue hit $81 billion by the end of October, about $10 billion more than at the end of September. Shares are expected to trade near November 19 under the ticker ANT, with up to 20 percent set aside for retail investors.
- All 7 Anthropic Founders Will Each Cash Out About $250M. Each of the seven will sell exactly 3,378,378 shares, which is just under $250 million at the $74 midpoint, between about $236 million and $264 million across the range. They will each hold about 3.2 percent of Class A afterward, roughly 22 percent together, and keep control through supervoting Class F stock. Board member Reed Hastings has signaled he wants to buy $300 million of shares through his donor-advised fund, an early hint of demand.
Agents at Work
- Anthropic Launches Claude Meetings. Claude joins a video call as a visible participant, announces itself in the chat, drafts the plan being discussed, maps a timeline on a whiteboard and answers questions using connected work apps, such as last year’s figures. It will not join calls with external guests unless an admin authorizes it, and it leaves automatically if an outside guest arrives. It starts in beta on Team and Enterprise plans with Google Meet only. Whiteboards and Sheets launched alongside it, in a race with OpenAI’s recent workplace tools and Meta’s Muse.
- Adobe Brings AI Agents to Photoshop, Premiere and Firefly. At MAX in Miami, Adobe released an MCP server so outside chatbots like ChatGPT, Claude and Gemini can use its features, and said its own agent will become proactive and always on. Firefly gets a private-beta prototyping tool called Builder, aimed at Figma’s territory, plus an on-brand chart tool.
- Kora Takes Aim at Windows. Kirt McMaster, who ran Cyanogen, is building a free, Linux-based desktop operating system with models and agents integrated, due in 2027. About 50 engineers are using Codex and Claude, and he says that made development at least five times faster, mostly in the last six months. Windows held about 77 percent of desktop share in October, so he is targeting heavy AI users rather than everyone.
- The smaller agent stories. Meta’s Muse is now on Apple Watch and Wear OS, a month after its iPad app. Spotify’s AI DJ, used by over 94 million Premium listeners, adds nine voices and a lower-chatter mode. Otter adds 27 languages, for 31 total, with 13 live now and a coming feature that follows speakers as they switch languages mid-call. Ring launched Drone Cam, a “flying indoor security camera,” and a patrol robot for businesses.
Platforms and Commerce
- OpenAI’s Business Hub and Business Agents. The beta lets companies verify a profile, edit product information and see profile views, mentions and website clicks, a feature TechCrunch compares to Google Analytics. The Verge adds that eligible US e-commerce and online service businesses can run a verified agent inside ChatGPT, kept separate from your main chat. OpenAI had retired a commission-based checkout effort in March. I only read the opening of the Verge piece.
- OpenAI Acquires Odessia. This looks like an acqui-hire: founder Francis Davidson, formerly of Sonder, joins OpenAI, and the year-old travel agent, which raised $6 million from Sequoia, has already been shut down. OpenAI has announced no travel product, though it says 1.2 billion people use ChatGPT weekly.
- Oura Buys Counsel Health for $625 Million. The AI-native primary care service has been inside the Oura app since June. The deal follows Oura indefinitely postponing an IPO of up to $2.2 billion.
- Google’s Two New TPUs. TPU 8tx, built with AMD, handles confidential training and inference with protection in the hardware, while TPU 8i Max combines two inference chips for double the memory, compute and bandwidth. Both extend the eighth generation introduced in April.
Rules, Risks and Proof
- New York’s RAISE Act Takes Effect. The portal opened Tuesday. Developers with more than $500 million in annual revenue must register by January 1, 2027, publish safety frameworks, report critical incidents within 72 hours and issue transparency reports, or face fines of up to $1 million, then $3 million. A new office called DIGIT enforces it, and a $55 million research center at Binghamton University will evaluate advanced systems.
- PSI and the Aaronson-Ambainis Conjecture. David Turturean used early access to Fable 5.2 to solve an 18-year-old quantum computing problem. Scott Aaronson says he does not fully understand the proof but believes it because it was formalized in Lean, and that he is puzzled it was missing from OpenAI’s recent batch of solved problems. Turturean said it might be the last paper he ever writes.
- Hugging Face on Open Source Security. Irene Solaiman links October’s OpenCode attack on API keys, a summer breach of Hugging Face by an OpenAI agent swarm after evaluation data, and Fireworks’ disclosure that attackers reached some customers’ model weights, datasets and billing information. Her worry is that defending these platforms means closing the permissiveness that makes them special.
- The Huang coverage. TechCrunch’s timeline says Megaspeed reportedly bought $4.8 billion of Nvidia chips in two years, and that The Information reported a related Chinese company facilitated $7.5 billion of purchases, including banned chips. Bloomberg reported Nvidia’s head lawyer intervened in a 2025 government probe of Megaspeed. The Verge opinion piece argues the chips that power the boom make their maker an obvious espionage target; I read only its opening.
- ChatGPT Plus for Veterans. OpenAI says more than 20,000 veterans and transitioning service members have claimed a free year of ChatGPT Plus. OpenAI blocked automated reading, so this is from its feed summary only.
The Throughline
The Anthropic filing is a lesson in how control gets engineered. The founders sell about $250 million each, yet stay in charge through Class F stock, the Long-Term Benefit Trust picks a board majority, and the people who buy Class A get limited influence. Meanwhile $26 billion of non-voting Class N stock goes to seven companies that supply chips, cloud capacity and distribution. Money and dependence are tangled together: Nvidia is both a supplier and a $10 billion investor, and the same is true of Google. A valuation of $1.404 trillion is therefore not only a verdict on Claude. It is a bet on a whole supply chain that has chosen to sit inside the company it supplies.
The second thread is that every product launch today comes with a rule about who may be present. Claude Meetings introduces itself in the chat and refuses to stay once an outside guest joins. OpenAI’s Business Agents are verified, and kept apart from your private conversation. New York’s law demands registration and 72-hour incident reports. Even Hugging Face’s Solaiman is arguing about who gets to contribute. The detail worth noticing is how much of the safety work is now access control rather than model behavior: who can join, who is verified, who can submit a pull request, whose weights were reachable. Fireworks, OpenCode and the Hugging Face breach all fit that pattern, since in each the weak point was credentials or data sitting in one place.
Third, agents are becoming the default interface, and every company wants to be the one the agent calls. Adobe shipped an MCP so Claude, ChatGPT and Gemini can drive Photoshop. OpenAI wants businesses to build agents that live in ChatGPT. Meta put Muse on your wrist, Kora wants agents in the operating system itself, and Spotify made its DJ personalities selectable. Kora’s McMaster also says AI coding tools made his team at least five times faster, which is the quiet reason new platforms like his are suddenly feasible with 50 people.
Finally, the PSI result shows what verification looks like when the machine is ahead of the humans. Aaronson trusts a proof he cannot fully follow because Lean checked it, and a human team explained it. That is a different kind of trust from the one in the Claude Meetings pitch. Machine-checked proof of a math claim is rigorous, while a meeting summary is a judgment call. The industry is mixing the two without always saying which one it is offering.
The Bigger Picture
Zoom out and the sector is being reclassified from a collection of startups into infrastructure. A company on the verge of a $1.4 trillion listing, a state regulator with a registration portal and a fine schedule, and a chip supply chain now entangled with national security all point the same direction. Infrastructure gets disclosed, audited and governed, and its owners are asked to report incidents on a clock. The RAISE Act’s 72 hours is a modest rule, but it applies the same logic that governs utilities and banks, and it arrives in a state with a $55 million institution built to review the filings.
The honest caveat is how much of today’s evidence is early or partial. The IPO numbers are an estimated range, not a priced offering, and the filing itself shows how far valuations have moved from the $2 trillion figure of a few months ago. The PSI proof is one researcher’s result, checked formally but not yet widely digested. The hacking stories, from OpenCode to Fireworks, describe incidents still being investigated. What looks durable is the direction: concentration of capital in a few labs and suppliers, more formal oversight, and agents moving into every interface at once. The open question is whether the openness that Solaiman defends survives contact with all three.
What to Watch
- Where Anthropic actually prices. The range is unusually broad, and the founders’ payout swings by about $28 million each between $70 and $78. Pricing above the range, as TechCrunch notes is common in hot deals, would say more about demand than any analyst note. First trades are expected around November 19.
- Who registers in New York first. The January 1, 2027 deadline is seven weeks away, and the state declined to give specifics on enforcement against companies that do not register. Watch whether OpenAI and Anthropic publish their safety frameworks early and whether other states copy the 72-hour rule.
- Whether Claude Meetings leaves beta and leaves Google Meet. It supports only Google Meet for now, and the guest-exclusion rule will decide whether companies actually trust it in sales and client calls. Also watch whether Microsoft and Google answer with their own in-call agents.
- The next AI-found proof. Aaronson’s question about why the conjecture was missing from OpenAI’s earlier batch suggests labs will race to claim results. Check whether other open problems fall to specialized harnesses like PSI’s.