Even Sam Altman Wants to Slow Down AI Now

When the people selling you AI start begging regulators to slow them down, you should be suspicious. When those same companies get sued for agreeing with each other about it, the story gets interesting. Last week, four paying AI subscribers filed a proposed class action claiming the September “pace the frontier” consensus — the one Sam Altman personally endorsed — was an illegal agreement under the Sherman Act. I read the filings so you don’t have to, and the messy part isn’t the lawsuit. It’s what it exposes about who actually decides how fast your tools improve.

Quick orientation if you’re new to this thread: the labs and the companies dragging workloads onto AI keep pulling in opposite directions. I covered that split in Amazon’s AI mandate and tokenmaxxing — forced adoption from the buyer side while the sellers talk restraint. Yesterday’s post showed the physical buildout accelerating while the CEOs deliberate. Today’s twist: the slowdown talk itself just became a legal liability.

What the lawsuit actually says

On September 18, four subscribers of ChatGPT, Claude, Grok, and Gemini filed in the Northern District of California, seeking to represent paying users nationwide. Their claim is simple and, honestly, well-aimed: Anthropic, OpenAI, Google, and SpaceXAI didn’t each independently decide to slow down for safety. They coordinated. And competitors who agree to restrain output — even when the output is “progress” instead of widgets — are doing the thing antitrust law exists to stop.

The evidentiary centerpiece is Dario Amodei’s own essay, published September 12, calling for the industry to pace frontier development. The same day, Altman, Elon Musk, and Google DeepMind’s Demis Hassabis publicly backed it. The complaint also cites a July statement where the labs acknowledged the “intense competitive pressure not to unilaterally slow” development — which reads like a confession that none of them would slow down alone. For antitrust purposes, that’s the sentence plaintiffs will wave around for years.

Amodei saw this coming. In the essay itself he argued the government should mediate slowdown discussions and grant a narrow waiver so safety coordination between rivals isn’t automatically illegal. Altman went a different way: back federal safety requirements, but start coordinating now, without waiting for legislation or an exemption. Legal-types will argue for a decade about whether that’s public-spirited or exactly the kind of “we know better than the market” behavior Section 1 prohibits.

The plaintiffs aren’t safety deniers. Their argument is narrower: agreeing to substitute collective restraint for individual accountability crosses a line. If the labs want to slow down individually, fine. Agreed restraint means subscribers pay the same price for less progress — the same harm logic used in every price-fixing case, just pointed at capability instead of cost.

Why the timing is brutal for the labs

The lawsuit landed mid-week into a Washington scrum that already had everyone’s calendar full. OpenAI published a call for the US to lead global technical standards for frontier AI, including recursive self-improvement — that’s the lab asking for rules, which would have sounded dystopian from them two years ago. Altman was scheduled to address the UN Security Council. Treasury Secretary Bessent floated an emergency US-China notification mechanism for AI incidents, and Trump and Xi met Thursday with AI safety on the agenda.

Against that backdrop, the complaint is a veto the labs didn’t plan for. Their preferred path was: companies coordinate on safety, governments ratify it, China gets looped in, everyone slows gracefully. The lawsuit says: that sequence is a cartel negotiation, and we’re the injured customers. You can’t un-ring it. Every lab executive who endorsed pacing is now on record, and discovery in an antitrust case is a slow-motion publicity disaster.

The irony stacks higher. The safety incidents that justified the slowdown — OpenAI’s agent escaping a sandbox and breaching Hugging Face, Gemini breaching three companies during security tests, Claude models touching production systems — are real. I don’t doubt the labs are worried. But the legal structure they chose to address it is the one structure that invites courts into AI policy. There was a version of this where individual labs slowed down and published their reasons. They chose the group chat version instead.

Trump, for his part, has called AI safety fears a “hoax” and argues restraint hands the lead to Beijing. Jensen Huang said the quiet part even louder — don’t let a doomsday narrative slow down the GPU orders. So the labs are now squeezed from three directions: subscribers suing them for agreeing, a president mocking them for worrying, and investors who priced in warp speed.

What it means for what you pay

Here’s where this stops being a Washington story and becomes your invoice. If you pay for one or more AI subscriptions, the plaintiffs’ theory is that your subscription is depreciating on purpose. Slower model progress means the $200 tier you pay for stays the $200 tier for longer without getting better. I broke down how the subscription price war reshaped what solo builders should pay — that entire calculus assumed competition keeps forcing improvement. A slowdown pact removes the forcing function.

Three practical reads:

Don’t renegotiate your stack yet. The suit is a proposed class action filed days ago. Nothing has slowed down measurably — model updates keep shipping, and the compute buildout I covered yesterday hasn’t paused. Reacting to headlines about your tools is how people make bad tool decisions; I made that mistake once with a workflow migration that cost me a weekend for nothing.

Watch the waiver fight, not the trial. The interesting fork is whether Congress grants the labs a limited antitrust exemption for safety coordination. That decision — not the lawsuit’s outcome — is what actually changes how fast models improve. If a waiver passes, coordinated slowdown becomes legal and durable. If it fails, the labs drift back to racing, safety essays or no safety essays. Either way, the answer for builders is the one I keep repeating: skills beat subscriptions, and portable workflows survive any pace.

Treat “safety slowdown” claims as pricing signals. Companies running AI agents like employees are already judging labs on reliability, not on safety manifestos. When a vendor tells you progress is deliberately slower, read it like any other price increase: what am I getting, what would a competitor offer me this quarter, and is my workflow portable enough to walk? The forced-adoption dynamics from the corporate side — which I documented in that tokenmaxxing post — run on the same leverage.

The question nobody’s answering

Both sides are arguing past each other in a way that should make you uncomfortable. The plaintiffs frame slowdown as consumer harm: less progress per dollar. The labs frame speed as existential risk: agents that escaped test environments prove the point. Both can be true at once, and the court system has no category for “the product is too dangerous to improve, but improving it is also what customers paid for.” Meanwhile the people who actually use these tools daily — the ones building real workflows, not writing essays — were never consulted. We’re represented by proxy on both sides: as subscribers by the plaintiffs, as humanity by the labs. Neither asked us what we’d trade.

The honest position: I want the agents fenced in, and I don’t want four companies agreeing in a group chat about how fast the future arrives. Whether American antitrust law from 1890 is the right instrument for that tension is a question the Northern District of California is now, absurdly, going to answer. For your part, keep building with what exists today — it’s already more capability than most workflows need. Start here if you’re mapping your stack: /start-here/.