NOOPS Weekly — Week of 14 September 2026
On Saturday Dario Amodei published an essay asking the industry to slow down. By Wednesday the President had called it a hoax, a rival lab had called it a cartel, Beijing had called it fear-mongering, and the man who sells the chips had said it changed nothing he could see. That is the week in one sentence, and the interesting part is not who agreed with whom. It is that the argument sorted the industry by revenue model, and once you see the sort you can read everything else that happened.
Who asked to be slowed, and who refused
The essay does one thing that costs Anthropic something: it puts outside evaluators inside the building, with badges and the right to publish. Everything above that — industry standards under an antitrust waiver, capability checkpoints, a speed limit on recursive self-improvement — needs the co-operation of people who have never yet co-operated. And the one costly step is also cheap for an incumbent: compliance overhead scales sub-linearly with revenue, so it is affordable at a US$65bn run rate and prohibitive for a new entrant. Amodei has argued for years that this is a race to the top rather than a moat. This week it was observably both.
The refusals came from exactly where the money says they should. Broadcom's Hock Tan has not revisited a forecast "in the least", and Anthropic — the company proposing the slowdown — is on track to be his largest custom-chip customer next year. Pacing slows training; his guidance runs on inference, which a paced frontier still deploys everywhere. Huang and Zuckerberg said liability, not law, will do the work, which is coherent and self-interested in equal measure: the chip vendor gains nothing from a slower frontier, and Meta's open-weights strategy is the one most exposed to a regime that defines "frontier company" by capability. Cohere's Aidan Gomez put the structural objection best — the problem is not who sits at the table but that there is a list at all — and he too is an interested party, selling on-premises deployment to the banks and ministries a scale-based regime would spare.
So the sort runs like this. The closed labs asking for co-ordination are the ones with listings pending; Nvidia is reportedly in talks to anchor Anthropic's with up to US$10bn, and CNBC found the company pitching a slowdown mid-roadshow while analysts split between "ladder pull" and Gartner's quieter observation that stricter standards favour whoever can afford them. The companies opposing co-ordination are the ones whose revenue is a function of everyone else's speed. Neither side's sincerity is the question. Hinton, who holds no position that benefits from a pause, made the cleanest case for the labs' good faith: warnings are bad for business. Both can be true. A pause can reduce the value of Anthropic and still raise the relative value of every incumbent over every entrant.
The defence was open-weight all along
The most consequential fact of the week arrived as an aside at Dreamforce. Sam Altman confirmed that after OpenAI's agents attacked Hugging Face in July, the 200-person company could not obtain a defensive model from another major lab and turned to Chinese open-source models instead. Read that against the essay's China clause, which rests the whole pacing structure on export controls and a crackdown on distillation. The tool that cleaned up an American lab's mess is the tool the crackdown would restrict. Rob Manson put it plainly: open models were the defence, not the threat.
The data pointed the same way. Enclave ran DeepSeek V4.1 Flash against eleven hacking targets and it took all eleven for US$4.65 — five of the wins were benchmark artefacts and the price is mostly a caching story, but the Jenkins solutions were not pattern-matching, and the model is MIT-licensed. Anthropic's own threat report counted 151 million Claude exchanges harvested through 3,500 fraudulent accounts, which is the grievance quantified; John Schulman's answer to why the model layer has not consolidated is the same fact seen structurally — anything learnable through RL can be distilled "because it's a small number of bits", and the scarce input, the prompt distribution, leaks commercially through routers and data vendors that no export control reaches. If that is right, frontier spending buys a lead measured in months, and the top of the stack is priced like pharmaceutical research with a six-month exclusivity window. That is a very different multiple from the one on the roadshow.
Mathematics reached acceptance in a week
The other argument of the week was quieter and further along. The Clay Institute declined to certify — Navier–Stokes is "apparently" settled, the process "deliberately unhurried", and the prize rules require a refereed paper and a two-year wait before a claim is even considered. Tristan Buckmaster quit the race: "I think it's pointless. The game is up." Not as a sceptic of the tools, which he calls a team of PhD students who never sleep, but as an objector to the companies. Twenty-five Fields Medallists took the industry's own word, misalignment, and turned it on the industry's objective function. By Friday Daniel Litt was proposing to award the PhD on a rigorous defence and make the provenance of the mathematics irrelevant — separating evidence of progress, which he is willing to let go to whoever or whatever produces it, from evidence of expertise, which he would rebuild around a person who understands in a room with people who can tell.
That is the transferable structure, and Sean Goedecke carried it into software the same day: the meaty GitHub repo, the take-home test and the competitive-programming score all stopped being evidence of anything, so employers lose the ability to tell who is good before they lose the need for them. Generation became cheap. Verification, attribution and canonisation did not, and the institutions that hold them are underfunded and have no commercial reason to scale. Every announced breakthrough from here inherits that shape: neither true nor false, for years, in any way an outsider can act on.
The constraint moved upstream
Beneath the politics the hardware story got simpler. Server revenue hit a record with GPU prices up 44 per cent while GPU units fell — a price story wearing a growth story's clothes, with memory as the proximate cause. The RTX 5090 vanished from US retail because a US$5,000 consumer card is a rational buy for an inference server. SemiAnalysis's first agentic results on Rubin show 67 times GB300's throughput per dollar at one fast point and 1.4 to 3 times where anyone actually serves — a strong generational step measured by a sympathetic tester, not a write-off of Blackwell. And when Jensen Huang said Nvidia will sell twice as many chips next year, Mark's reply was that it cannot fill the book, because the constraint sits at TSMC and the three HBM makers, none of which has announced a doubling. Whichever of them is right, the pricing power sits with the six upstream firms, not with Nvidia — which is why SK Hynix talking to Intel about making memory at its Ohio fab, possibly as a joint venture with the cloud firms is the week's most interesting deal, exploratory as it is.
Where they get built
Moody's has US data-centre load doubling by 2030 with the grid seven years behind. SemiAnalysis mapped more than three hundred moratoriums and found they touch 20 GW of planned capacity but delay 2.3 — a moratorium bites only when it reaches the parcel, is still in force when the approval is needed, and cannot be designed around. The political price of grid capacity is rising faster than the financial one: the Ratepayer Protection Act passed the House 417 to 3 and stalled in the Senate for not going far enough, and the President phoned Jensen Huang on stage to call data-centre opposition a hoax on the day a poll found two-thirds of Americans oppose one in their community.
Australia ran the same fight in miniature and got the terms in writing. OpenAI told Canberra there will be no training centre unless copyright loosens; leaked Attorney-General's slides showed the concession already drafted, opt-out by default; artists, the crossbench and the Coalition front bench turned on it within a day. Meanwhile Anthropic leased at the A$30bn Western Downs project — for inference, which sidesteps the copyright fight entirely. ASD's chief asked for an AI early warning system the same week; if that is formalised before the copyright question settles, Canberra will have taken the safety side of the bargain without yet paying the investment side.
What we'd hold
Three things from the week that will still matter in a month. The pacing regime will be judged by whether an outside review team is actually seated with the promised access before the midterms; if not, it was a message to investors and the frontier prices as unpaced. The distillation crackdown will be judged by whether legislative language regulates distillation or open-weight release — different bills, and the difference will be easy to miss in drafting. And the capacity thesis will be judged by the next 10-Q from Nvidia and the next capacity statements from TSMC and the HBM makers, not by any model release.