The AI Builders Ask for Speed Limits. No One Proposes Spending Less.

On September 12, Dario Amodei — the CEO of Anthropic, the company building Claude — published an essay titled "We Must Pace the Frontier" and called on the AI industry to slow down before its own creations get out of hand. Sam Altman of OpenAI agreed. Elon Musk, who has spent years suing Altman, agreed too. President Trump, asked about the risk to humanity, said he has no concerns and that America cannot afford to slow down.

When the people who would get richest from AI moving fast start demanding speed limits — and the president who controls the policies refuses them — a retail investor is owed a clear read. Here it is: nobody in that "restraint" camp is proposing to spend less. The real fight is not over whether AI slows. It is over who gets to build, and the scarce inputs that decide who wins.

Start with what Amodei actually asked for. It was not a budget cut. It was a three-step plan: independent evaluators granted employee-level access inside the labs, voluntary industry coordination on standards, and governments — "including authoritarian states" — cooperating to pace development. Anthropic committed to the first step unilaterally. Altman echoed it, saying OpenAI would hand the same outside access to its systems.

The reasoning behind the essay is genuinely alarming if you read it closely. Amodei warned that within six to twelve months a swarm of misaligned AI agents could take over the internet and cause hundreds of billions of dollars in damage, and he cited the ability of AI to build the next generation of AI as a reason for "unusually deliberate care." Inside his own company the alarm is louder. A former researcher, Jacob Coxon, resigned saying the builders "earnestly believe that AI could kill us all by the end of the decade", and put the odds of that outcome above ten percent. Anthropic's chief safety researcher did not dispute the number — he said the lab has no plan yet to align a superintelligent system.

That is the sincere version of the story, and it deserves to be stated fairly, because if the men closest to the frontier believe they may be building a machine that kills everyone, a slowdown sounds rational.

Now the numbers. Anthropic raised $65 billion in May at a $965 billion valuation — topping OpenAI's $852 billion — with a roughly $47 billion revenue run rate, and filed confidentially for an IPO in June with an October listing reportedly in view. That is not the behavior of a company planning to build less. It is the behavior of a company concentrated on building faster than anyone else while it still can. Altman, meanwhile, told Fortune OpenAI would not go public in 2026 — citing safety concerns — even as the company pushed toward a trillion-dollar valuation. The "slowdown" the leaders propose governs the pace other people build at, not the pace their own capital intends to set.

Note what the restraint requests actually target. Anthropic's head of policy praised competing congressional bills and specifically backed blocking advanced chips to China, national laws requiring frontier-model testing, and powers to block unsafe models. Around the same period, reporting found OpenAI and Anthropic quietly lobbying Washington to restrict open-source models — the very upstarts that threaten their moats. In June the Trump administration itself ordered Anthropic to cut off access to its most advanced models on national-security grounds, then lifted the restriction two weeks later. The pattern is a gate, not a brake: speed limits that the insiders help set, and that mostly bind everyone except themselves.

This is the abundance-scarcity test, applied to intelligence. Capability is becoming abundant — these companies are engineering it faster and faster. As it becomes abundant, what gets scarce is not more intelligence, but the two things required to deploy it: raw compute and power, and the permission to sell it. Every "restraint" ask in Amodei's essay is an argument about permission and gatekeeping — evaluators, standards, who is allowed to train. None of it touches the buildout, because the buildout is where the scarce commodities are being bought up. That is why the message lands the way it does.

The buildout is not slowing. Nvidia reported second-quarter revenue up 106% year over year and was reportedly in talks to backstop roughly $250 billion of data-center financing. Power-equipment maker GE Vernova more than doubled its data-center orders in the first half of 2026 and carries a $76 billion backlog, its stock up over 75% for the year. Optical-components maker Ciena warns demand is exceeding supply so sharply that Goldman Sachs expects the addressable market to jump from $15 billion to $154 billion by 2028. An "AI bubble about to pop" narrative has to wave its hands past all of that.

Here is where the Washington–Silicon Valley divergence becomes the actual risk. Trump is betting America must run the race faster than China and will not concede a slowdown. The labs are betting the frontier is dangerous enough that a pace floor requires China to slow too. Those two positions cannot both govern. If Washington keeps the pedal down and refuses to restrict, the labs' own safety warnings become a rising regulatory and reputational overhang on every public AI stock. If Washington ever does comply with the labs, it almost certainly does so by restricting who can participate — which helps the trillion-dollar incumbents and squeezes open-source and foreign rivals. Either direction, the volatility lands in the same place: the chips, power, and data centers that the whole trade is priced on.

So the headline is not "AI is collapsing" and it is not "AI is now officially safe." The people closest to it are saying the pace itself is the danger, and the government is telling them to keep going regardless. That is a policy fight, not a fundamental one, and it is being fought over the scarcest inputs in the economy. An investor who reads "the builders want a slowdown" as an instruction to sell the capex trade is mistaking a negotiation about who builds for a signal that building stops. The offers still being made — $250 billion backstops, $76 billion power backlogs, revenue growing in the triple digits — are telling you the opposite: the spending is not the question. Who gets to control it, and on whose schedule, is.