They asked if slowing down is legal

Sam Altman told his staff this week that OpenAI is open to pacing cutting-edge development. Ideally with other labs. Some of those labs, he added, may refuse.

That is the careful version, the one Bloomberg had from people in the room. The less careful version is the one that has been building in public for days: Jakub Pachocki arguing that coordinating to slow future development is the path, Chris Lehane asking Congress for mandatory rules because voluntary commitments are not enough, Paul Christiano joining the Foundation board after posting that he now sees a meaningful near-term risk of irreversible loss of control. Altman’s all-hands line lands inside that weather. It also lands next to a quieter question OpenAI has been putting to members of Congress in recent weeks, per Wired: if the industry actually tried to slow down together, would that be legal?

The Sherman problem

Antitrust was written for companies that restrict output to raise prices. Frontier labs that agree to hold a model generation are, on one reading, restricting output. Nicholas Felstead’s March argument — a coordinated pause may amount to that kind of restriction under the Sherman Act, depending on the details — is now circulating as more than a law-review worry. Legal uncertainty itself becomes a reason not to try. OpenAI wants clearer guidance. A July bipartisan bill, the Collaboration on Adversarial Threats and Security Risks Act, would carve a safe harbor for safety and security coordination. It has not been enacted. The House version sits in Judiciary.

John Schulman’s reply on X cut through the fog the way cofounders sometimes do after they leave: stop feuding, work on a pacing proposal, and stop citing antitrust as cover — the law blocks certain agreements, not jointly writing a plan. That is the other camp’s line. The first camp says the fear is real enough that big tech will not board a voluntary slowdown without a statute. Both camps can be right about incentives and still leave the brake unbuilt.

Anthropic, meanwhile, had a spokesperson say Thursday the company is interested in working with the industry on release pace. Interest is not a signed schedule. Altman already told his room that some players may not agree. The sentence does not need a name to do its job.

The chipmaker’s morning

Jensen Huang was at Goldman’s Communacopia on Thursday with a different temperature. Jacob Coxon’s extinction warnings — the resignation letter that spent the week on television — Huang called deeply untrue, wrong, arrogant, dismissive of the safety work already underway, according to Brad Gerstner in the room. On the cybersecurity scare that followed the agent breakouts, Huang was blunter still. The reason there is so much conversation about cybersecurity, he said, is that the industry is getting ready to launch products. What better way to create demand than to create a problem. Who doesn’t want their market hysterical and lined up around the corner.

Read that next to OpenAI’s Defense Factory sprint and Anthropic’s METR investigation and you get the week’s sharpest collision. The labs are asking for brakes, auditors, and mandatory rules. The company that sells the racks is calling the fear cycle a go-to-market motion for red-team and blue-team products, CrowdStrike and Cisco and Palantir included, and defending a three-to-four-trillion infrastructure story while he does it. Huang can be right that cybersecurity will be a major AI use case and still be describing a political economy in which panic is inventory.

Capacity is already a kind of pace

Late Thursday into overnight, OpenAI paused new ChatGPT Pro sign-ups and upgrades. Thibault Sottiaux’s explanation was infrastructure, not philosophy: Astra demand is unprecedented, Pro puts the most strain on the systems, existing Pro users keep access, Plus and Go and the API stay open, more capacity as fast as possible, no reopen date. That is not a safety slowdown. It is a queue. But it is also the week’s accidental demonstration that “pace” already happens when the cluster cannot breathe — just not the kind of pace Pachocki meant, and not the kind that requires an antitrust memo.

The product pause and the all-hands pacing talk share a building and almost nothing else. One is about GPUs and power. The other is about whether the people who buy those GPUs will ever agree, in writing, to leave the next generation on the table long enough for safety to catch up. Huang’s order book prefers the first kind of constraint. Congress has not yet legalized the second.

What to watch

Watch whether the Collaboration bill moves before midterms or becomes another named PDF. Watch whether Altman’s “some may refuse” hardens into a public split or stays a staff-meeting hedge. Watch Anthropic’s interest become a date or a shrug. Watch whether OpenAI’s misalignment-reporting framework arrives with any commitment that looks like holding a release. And watch Huang’s line travel: if the chipmaker can frame the scare as demand creation while the model labs ask if a slowdown is even lawful, the window Lehane opened on Wednesday is already negotiating with a balance sheet that does not want it closed.

Pace, please, is easy to say in an all-hands. A legal slowdown that competitors will join is a statute, a schedule, and a chipmaker who stops smiling at the hysteria. This morning the industry has the first sentence. The rest is still antitrust fog and a Pro button that went grey because the racks were full, not because anyone agreed to wait.