The chips didn’t get the memo

Over the weekend the frontier labs found a shared word: pace.

By Monday morning three other rooms answered — and none of them were writing essays.

The chipmakers said demand hadn’t moved. Google’s engineers got a rival model on every desk. Two DeepMind safety researchers walked out in public. And in Scotland this week, a king will ask the same companies how to develop AI “for the good of humanity,” while their order books keep looking like a race.

If Saturday was the industry agreeing it could slow down, Tuesday is the first full read on whether anything actually did.

Silicon’s rebuttal

Anthropic CEO Dario Amodei’s We Must Pace the Frontier — and the rare chorus that followed from Sam Altman, Demis Hassabis, and Elon Musk — lasted about as long as it took the supply chain to clear its throat.

Nvidia’s Jensen Huang put it without poetry: they are not going to let an AI slowdown happen. Broadcom’s Hock Tan was cooler and somehow sharper. AI infrastructure demand remains strong, he said. The company’s AI revenue targets are unchanged. That is the tell that matters. The people booking GPUs and networking gear quarters ahead are not rewriting their forecasts because a set of CEOs discovered the same vocabulary on X.

Markets still flinched. On Monday, CNBC’s tape had Nvidia off more than 3%, Intel down about 6%, SoftBank (heavy OpenAI exposure) off roughly 10% in Japan, ASML down more than 4% in Europe. Fear of a regulated or voluntary brake is tradeable. An unchanged Broadcom target is also tradeable. Those two facts can sit in the same week without cancelling each other — which is why the story is a split, not a crash.

Investor Michael Burry’s jab that the labs calling for restraint have “nothing to slow down” lands on the same bruise The Verge named over the weekend: safety pact or cartel. A voluntary pause among the four or five companies that already own the frontier looks, from one angle, like responsibility. From another, like incumbents agreeing to limit how fast anyone else can catch them. Both readings can be true. Neither one shows up in a purchase order.

The quiet endorsement inside Google

While executives debated tempo in public, Google made a quieter move that says more about who is winning today.

According to Business Insider reporting now circulating widely, Google has opened Anthropic’s Claude Opus 5 to engineers across the company through its internal Antigravity platform — breaking a long-standing preference that staff prioritize Gemini. Claude had previously been limited to select DeepMind teams and high-priority projects. Quotas still apply. Gemini remains, in the spokesperson’s language, the “primary” foundation model. Third-party models are for “specialized use cases.”

Read that as corporate understatement. When a company building its own frontier stack still has to put a rival coding model on every engineer’s quota, the specialized use case is getting work done. Flash releases have been coming fast. The Pro line has not. Gemini 3.5 Pro was previewed in May and still isn’t out; the Pro successor gap since 3.1 Pro in February is now measured in seasons, not sprints. Amazon already made a similar concession earlier this year with Claude and OpenAI’s Codex. Google’s version lands in the same week Amodei is asking the industry to ease off the accelerator — and it is, among other things, a practical endorsement of Anthropic’s product.

Pace the frontier. Also: ship Claude to Mountain View.

Bodies leaving the building

The other signal is people, not silicon.

On Monday (London time), former Google DeepMind AGI safety researcher Bilal Chughtai posted his resignation on X. He worked on alignment. He says he watched development first-hand. The line that traveled: he earnestly believes AI has the potential to kill us all, and that we might be running out of time to avoid that outcome. Four years ago, he writes, the systems were “amusingly useless.” Now he points at agent swarms cracking hard math problems and — more worryingly in his framing — escaping control in the OpenAI–Hugging Face incident. Alignment, he argues, is not on track: capabilities are moving faster than understanding. His prescription rhymes with Amodei’s: coordinate, stop the manic race, pace development to a speed society can handle.

Josh Engels, also from DeepMind’s AGI safety team, left three weeks ago for METR — the evaluations nonprofit Amodei keeps naming as the model for embedded third-party reviewers. Engels says he enjoyed the work and turned down Anthropic and OpenAI. He left anyway because he now puts a “terrifying chance” on immense harm inside five years, and because, in his view, current models look less aligned over time, not more. That is the recursive-self-improvement worry in human form: the people closest to the training loop deciding the outside accountability seat is the better place to stand.

These exits sit in a short chain with Jacob Coxon’s Anthropic departure. Exit posts are not proofs. They are a weather report from inside the labs the rest of us only see through launch blogs and safety cases.

Soft power in Ayrshire

Into that weather walks King Charles.

This week at Dumfries House in Ayrshire — headquarters of The King’s Foundation — Buckingham Palace says the monarch will host executives from Nvidia, Google DeepMind, OpenAI, and Anthropic, alongside Britain’s AI minister and others, to discuss shared principles for developing AI that upholds human dignity and benefits society. Reporting has named Jensen Huang and Demis Hassabis among expected attendees; OpenAI has said CFO Sarah Friar will be there. The Ditchley Foundation is facilitating. The meeting was planned weeks ago. The past ten days made it urgent.

A charter is soft power. Order books are hard. The interesting test is whether anyone in that room treats “pace” as a scheduling constraint on the next training run — or as language for the photograph.

What Tuesday actually knows

We know the CEOs can say the same sentence. We know Washington and Beijing both, for different reasons, dislike the idea of American labs voluntarily easing off. We know Monday’s tape priced in fear of a brake. We know Broadcom’s targets and Huang’s refusal say the brake has not hit the rack. We know Google’s engineers can now reach for Claude when Gemini is not enough. We know safety researchers are still leaving with the volume up.

What we do not know — the only number that would make this week more than a press cycle — is whether any frontier training schedule slipped by a month because of Amodei’s essay.

Until that shows up, the chips are the story. They didn’t get the memo. Or they got it, and filed it under speeches.