Anthropic crosses its own safety line days before listing; OpenAI's first $10B quarter; United bars booking agents

🔥Today's one thing

🔍What is real, what is not

• AlphaProof Code — Google's AI lab says its agents mathematically proved the security of the running code behind Signal's messaging — the first time live code of a widely used encryption protocol has been verified this deeply — and a proof-checking program confirmed every step, so this claim genuinely holds up. Work that normally occupies expert teams for years took four weeks; provable software security is getting cheap, which changes what you can demand from vendors. Google's other launch of the week, its strongest model Gemini 4 Carbon, is on an introductory $2 per million input tokens until the offer doubles after November 30. Google DeepMind

• Northern Trust — says an onboarding agent it built with Microsoft engineers stationed inside the bank turns a data-gathering job that occupied analysts for about two and a half weeks into a three-day task — an 83% saving both companies announced Monday, with analysts still approving the agent's output at every step. That is the shape of proof to demand from your own teams: a named process, a before-and-after, a human checkpoint. The bank says a second agent now spares relationship managers two to three hours per client memo. CIO Dive

• Dario Amodei — In the 42-minute video Anthropic made for IPO investors, the chief executive said cancer, Alzheimer's, heart disease and mental illness would be essentially gone within ten years. Medical figures pushed back publicly and hard, and a biotech executive who has taken two companies public called it desperate talk for a listing pitch. This is the hype half of the week. Anthropic has bought biotech startups and opened a lab, but nothing in its record supports the timeline — treat it as salesmanship, not a plan. Business Insider

  “Curing cancer, Alzheimer's and mental illnesses can definitely benefit from AI. But that is impossible in just 10 years.”

📈Money and the market

• ANT — Anthropic began meeting investors for its Nasdaq listing, offering 825 million shares at $70 to $78 to bring in as much as $64.4 billion, below the $2 trillion valuation once discussed. Filings put annualized revenue near $81 billion as of October 31, against almost $900 billion of computing commitments. This listing is the live referendum on AI economics; how the shares trade will shape what AI vendors can charge you and what investors will fund next year. Trading could begin Thursday or Friday of next week, per the Wall Street Journal. Anthropic

• OpenAI — OpenAI's quarterly revenue passed $10 billion for the first time, Bloomberg reports, citing people it does not name — a jump of nearly half from the prior quarter's $6.7 billion, with losses narrowing. The company's annual run rate sits at about $50 billion. Rival Anthropic is further ahead on revenue and further behind on profitability, and the two count partner sales differently, so headline comparisons mislead — but both are still spending far more than they earn, which is what keeps pressure on the prices you pay. TechCrunch

• TSMC — Analysts again flagged the circular financing spreading through AI, in which suppliers bankroll the customers who buy their output, making true demand hard to gauge. If part of the money flowing into AI is vendors investing in one another, your own AI plans should assume less certainty about who is really paying full price. TSMC earned more than $23 billion in the July-to-September quarter, its sixth straight record. Ars Technica

• CoreWeave — CoreWeave is close to a borrowing of $4 billion or more, in loans carrying investment-grade ratings that the company says should cost 7% to 7.5% — roughly half the yield on its existing junk bonds — with the deal possible as early as next week, Bloomberg reports, citing unnamed sources. Cheaper money would ease the squeeze on a fast-growing company that has been losing money, and hints that lenders now trust AI computing contracts enough to lend against them on better terms. Bloomberg via Techmeme

⚠️Risk the board will raise

• United Airlines — said Wednesday it will bar third-party AI agents from booking or checking in passengers, on its own apps and through the travel agencies and fare sites that sell its tickets. The airline says customers who hand an assistant their login expose payment details and air miles, and that agent check-ins could break federal luggage rules. It is the first big consumer brand to formally shut agents out — Southwest is letting ChatGPT users browse its fares — and it hands any company selling online a ready-made argument for doing the same. CBS News

• Nvidia — Bloomberg, citing unnamed sources, reports that Nvidia's directors have hired an outside firm to identify candidates to replace chief executive Jensen Huang, and are weighing naming a temporary chief while the search runs. Nothing is confirmed and the company has not commented. For your board, this is the key-man question made visible: the world's most important AI supplier, a founder-led company with no public succession plan, is suddenly a live governance crisis — expect directors to ask how exposed your own critical vendors are. Bloomberg via Techmeme

🌍The big AI stories today

• Jensen Huang — the New York Times says he also had an affair with a Nvidia vice-president who reports to him; the Journal says his board is preparing to remove him; shares fell about 8% after hours Wednesday. Nvidia's governance mess is now an industry-wide supply risk. A House chairman says Huang will likely be subpoenaed if he refuses a closed-door briefing on national-security questions. TechCrunch

💡One takeaway