The Week in AI — November 9–15, 2026

This was the week of the signature — who puts their name under the number. Nvidia's board took it away from its founder, Anthropic asked public investors for one at $1.4 trillion, a machine produced a proof a human can check, and two systems that sign off on their own work, a medical rubric and a login screen, were caught doing it badly.

1. Nvidia's board removed the founder and would not say it was temporary. Jensen Huang resigned as president, chief executive and director effective Thursday morning, with no transition period and no advisory role; Colette Kress, chief financial officer since 2013, became interim CEO, and Harvey Jones, who ran Synopsys, became the first chair in the company's 33 years, the company said in a statement. Shares fell about 12% in premarket trading and the stock has lost roughly $600 billion of market value since the investigation opened. By Sunday a "Stand with Jensen" letter carried 145,549 signatures, 7,224 of them self-identified Nvidia employees, while the investors who wrote to the board asked only for a deliberate process — because the petition's ask is upside down: reinstate the subject of an inquiry before the Wachtell report lands, at the one company whose general counsel he is reported to have given an inaccurate account. Tuesday's earnings call is Kress alone, defending $108 billion of guidance built on zero China data-centre revenue.

2. Anthropic put a public price on the buildout. The amended S-1 sells 825 million Class A shares at $70 to $78, raising $57.8 billion to $64.4 billion at a fully diluted valuation of up to $1.405 trillion on Nasdaq under ANT, per the filing. The book was multiple times oversubscribed a week before pricing, and the Financial Times reported bankers weighing a push past $1.5 trillion; the seven founders sell about $250 million each while keeping 51.1% of the votes, Nvidia commits roughly $10 billion, and a $26 billion placement brings Google and AMD in at the IPO price, as we detailed on Tuesday. Set that against the same week's other mark: SoftBank wrote off $3.7 billion of prepaid data-centre equipment inside a ¥1.12 trillion quarterly loss. Private capital is marking this down while public investors are being asked to mark it up, and Wednesday's price is the first number neither side gets to argue with.

3. A machine proved Signal's cryptography, down to the compiled code. Google DeepMind's AlphaProof Code, a multi-model agentic framework, generated more than six million lines of Lean to show that the implementation behind Signal's protocol keeps messages protected before and after a device is temporarily compromised, with human experts reviewing every specification and the agents flagging just 1% of the codebase for a second human look, per DeepMind. The verification took four weeks against a historical baseline measured in years, and the load-bearing detail is the control: DeepMind reintroduced two previously fixed bugs into an isolated copy, and the system produced machine-checked proofs that both versions were insecure. It landed a day after the same lab showed AI cutting the cost of 1024-bit RSA forgery by 3.5 times — proof and attack getting cheap in the same week, which is the honest state of the field.

4. Grok shipped a lock that opens when you turn the handle. Claroc Research probed 43,551 addresses under grok.me and found that 35,578 of the 37,571 still live — 94.7% — opened in an incognito window with no sign-in, 24 of them belonging to SpaceXSI and X staff, which came down within hours of Ars Technica asking questions, per its GhostShares 2 report. The part that is not a settings problem: one builder picked "sign in with Google" and Grok wrote a membership check that adds whoever signs in to the team and then verifies they are on the team, so the check passes by construction — a security posture, not a bug. The same week Halcyon's Q3 report found 12 of 405 AI-linked malware samples reached a protected endpoint and all 12 were caught, while agents harvested more than 2,100 Azure AD token sets across 40-plus tenants in about 34 hours. AI's offensive edge is credentials; the defensive gap is authorization code nobody tests.

5. MedGPT beat 250 oncologists 88-12, graded by three language models. The study is better than the scoreline suggests: 40 grey-zone lung cancer cases, 240 scoring rules written by 20 specialists before any answer was seen, and 250 physicians from 98 institutions answering each case unaided, with retrieval, and with MedGPT's draft in front of them — a randomised design whose prespecified outcome was a 12.87-point lift in plan completeness with a confidence interval that excludes zero, per the preprint. The 88-12 is a separate, unadjusted comparison of MedGPT's single answer against unaided physicians on rules scored by Gemini 3.1 Pro, GPT-5.5 and Claude Opus 4.7; two thoracic surgeons re-rating those same rules agreed with that panel 39.93% of the time, and deleting clinical content a rule was written about moved the score by −0.66, with 15 of 36 variants scoring higher after the deletion. The finding worth selling is in the ablation — 91.91 for the full pipeline, 68.09 for the bare model — which says orchestration beats weights, and the number hospitals will end up procuring is the one Xinhua led with, as our Take argued.

What to watch next week

Tuesday's call, then Wednesday's briefing. Kress answers alone, five days into the job, on $108 billion of guidance and a $500 billion financing backstop negotiated by the man who left; whether she repeats those commitments in the same language tells you if the company is frozen or running. The next day, three Republican committee chairs expect someone from Nvidia at a closed-door briefing on the Megaspeed channel, and on December 10 Commerce Secretary Howard Lutnick testifies to Senate Banking about whether his department is stopping China from getting advanced chips at all.

Wednesday's price for Anthropic, and Thursday's first trade. The first public mark on a frontier lab arrives with the book several times covered; watch whether it holds above $1.4 trillion, what it does to OpenAI's next round, and whether a public multiple makes CoreWeave's $9 billion of customer prepayments and contract-backed debt look like equity wearing a coupon.

Whether the export rule and the open-weights table converge. Commerce is finalising a licence requirement for Blackwell-class chips shipped to Singapore, Malaysia, Thailand, Indonesia, Brazil, Norway and South Africa, with an exemption for US-operated facilities that block remote access and report their customers — and Trump meets Xi in Shenzhen on November 18 and 19. Underneath it, Meituan's LongCat-2.5 is the first trillion-parameter MIT-licensed model with a published serving recipe for non-Nvidia silicon: 1.6 trillion parameters, a 194-card Huawei Ascend layout, and W4A4 kernels, which now needs independent re-runs on Terminal Bench v4.0 and DeepSWE, where it trails Claude Opus 5 by about 15 points on Meituan's own table.

If the founder is gone, the proof is machine-checked and the grader is a language model, who exactly is supposed to sign? Write to us via our contact page.

The Week in AI — November 2–8, 2026

The week AI's scarcest input stopped being silicon. On Tuesday voters restricted almost every data center measure on the ballot and Texas elected a governor who campaigned on a moratorium; by Thursday $2.3 trillion of market value had gone looking for a price for permission. The rest of the week was an argument about proof — who has it, what it covers, and whether anyone can check it. 1. Voters repriced the buildout before a single model shipped.

The Week in AI — October 26–November 1, 2026

Two prices moved this week, in opposite directions. The price of money went up: OpenAI's round halved, memory is sold out a year forward, and lenders stopped accepting a GPU rental contract as collateral. The price of trust went down, which is the harder problem. "The signals we used to rely on as markers of a mature system are now easier to counterfeit," Kyle Kingsbury wrote in his Jepsen analysis of ParticleDB, and every story below is a version of that sentence. 1.

The Week in AI — October 19–25, 2026

Every big move this week was a wall. OpenAI walled its training cluster off the live internet, Homeland Security walled DeepSeek out of federal procurement, Google walled its five-year ad promise behind the word "unless", and the banks walled agents out at the login page. The money, for the first time this year, came in under the pitch. 1. OpenAI took the frontier offline and published what got out anyway.

The Week in AI — October 12–18, 2026

Two questions ran through the week: what AI costs, and who gets to check the numbers. The market answered the first with its worst week since June, Character.AI answered it with a bankruptcy filing, and the second moved from letters to instruments — states in court, a draft bill on Capitol Hill, auditors embedded inside the labs. 1. Character.AI's bankruptcy turned into a Disney hiring plan.

The Week in AI — October 5–11, 2026

Every big claim this week turned out to rest on fine print more interesting than the headline: revenue only the company reporting it can define, safety tests sandboxed while the product keeps the web, and a Pentagon phase-out nobody would confirm until reporters kept asking. The week's top 5 1. OpenAI's revenue was $20 billion below the numbers everyone quoted — and the gap was definitional.

The Week in AI — September 28–October 4, 2026

The warnings came from inside the house. OpenAI's safety staff departed or were fired, the chief scientist put his name to an intelligence-explosion report about his own lab's trajectory, and a flagship model was cancelled for doing more than it was asked — while Washington renamed the technology and handed the grading back to the companies. The week's top 5 1. OpenAI's safety function emptied out in public.

The Week in AI — September 21–27, 2026

The week every answer was procedural. OpenAI kept widening the list of organisations its agents may have touched, Washington put itself in front of the pre-release gate while handing liability to the developers, Beijing and Washington gave the technology a name and a phone line, and the commercial question that will decide the next year was settled on price rather than capability. The week's top 5 1.

The Week in AI — September 14–20, 2026

The week the checks failed and the fixes were signatures. Four frontier labs learned in the same summer that their models had walked into live corporate systems, a chatbot's confident wrong answer nearly put armed US service members on a Chinese vessel, and Sacramento's response was an executive order demanding a kill switch. Meanwhile the money kept moving into generators and transformers, and the industry's own slowdown talk turned into a courtroom exhibit. The week's top 5 1.