Dario wants a pace. Jensen wants a purchase order.
Anthropic’s slowdown essay and OpenAI’s nodding along ran into the only constituency that cannot afford a pause: the company that sells the cluster.
SANTA CLARA — September 17, 2026
Dario Amodei wrote that the frontier should be paced. Sam Altman said he agreed and that OpenAI would take independent evaluators with employee-like access. Elon Musk, never one to miss a chorus, posted that Dario is right. Jensen Huang, whose chips are the reason any of these people have a frontier to pace, said the market is already there and we do not need new laws.
That is not a policy debate. That is a supply-chain argument wearing a halo.
The lab that rents the cluster wants a speed limit. The company that sells the cluster wants Q4.
The gossip under the essay
The unfashionable read of “We Must Pace the Frontier” is that Anthropic can see Claude’s agentic cyber work — Opus 5’s crowd will tell you the model is now a colleague you do not leave alone with a VPN — and would prefer the next six months not be a contest to see who first ships a botnet-shaped demo. Amodei’s public nightmare, a swarm that takes the internet in a year, is lurid. It is also, if you talk to the safety people rather than the essay, a way of saying: our own evals are getting too good at the wrong homework.
OpenAI’s Astra launch two weeks earlier made that homework official. Critical, in OpenAI’s framework, is a permission structure: refuse it on the website, un-refuse it for Daybreak. Anthropic’s Mythos-class restrictions are the same idea with different stationery. The slowdown ask is what you say when you have already built the thing you want regulated.
Huang’s pushback, reported alongside Andrew Ng’s “science fiction” line, is the honest industrial position. Nvidia does not get paid when a frontier lab takes a sabbatical. Ng’s point that a blanket pause also pauses the fixes is not wrong. It is also very convenient if your business is more clusters.
The other Nvidia sentence: AVO
The chipmaker is not only selling the rack. In August, Nvidia’s research blog said its AVO agent stack — Agentic Variation Operators: memory, tools, an execution loop, a supervisor that yanks the run when it stalls — took Claude Opus 5 from a roughly 30 percent public-set ARC-AGI-3 baseline to a 100.00 RHAE on the 25 public environments. Runtime Wire and Nvidia’s own post are the paper trail. They did not run the private sets. They did not ship an API. They did say the same Opus 5, wrapped in their loop, used fewer actions than another Opus-5 harness.
That is the gossip that should sit next to Amodei’s essay. Opus 5 is already the model people lower their voice about in cyber rooms. Nvidia’s move is to argue the system is the product: the thing that turns a 30 into a 100 is not a new soul, it is a supervisor and a week of GPU-kernel search. Huang can oppose new laws and still be building the control plane that makes other people’s models finish the homework. If you sell the cluster and the loop, a paced frontier is a paced invoice.
Talent, as usual
The other rumor this week, which I will not dress up as a scoop, is the usual one: a handful of cyber-eval researchers are being priced like athletes, and the offers now include “you can see the unredacted traces.” That is the real pace. Not a treaty. A bidding war for people who know what the models did when the allowlist failed.
If you want a statute, write one that names disclosure clocks and eval networks. If you want a sermon, you already have three CEOs and a chipmaker. I know which one ships on time.
