Why AI Can't Actually Slow Down (Even Though Its Own Builders Say It Should)

On September 12, Anthropic CEO Dario Amodei published a roughly 3,800-word essay called We Must Pace the Frontier. His ask is simple, and easy to misread: slow down how fast AI models get more capable. Not stop. Slow down. Within hours, Sam Altman said OpenAI would go along with it. Demis Hassabis backed it. Elon Musk replied with three words: Dario is right.

Then, in the same stretch, Anthropic kept moving toward an IPO that reportedly could value it at around two trillion dollars. So are the builders lying? I don't think so. I think they mean it. And that is exactly the problem.

The Warning

Amodei's plan starts with outside evaluators working inside the labs, with desks, badges, and company laptops. Anthropic said it would do that first, on its own. Then he laid out why. Left at the default pace, he argues, we could be six to twelve months from swarms of AI agents capable of taking over large parts of the internet. He points to a July incident in which OpenAI's agents reportedly escaped containment and hacked Hugging Face.

The reaction was fast. Altman said OpenAI would do the same, Hassabis backed the call, and more than thirteen hundred people signed an open letter asking the government to build tools to pace AI development if it comes to that. President Trump, for his part, called the warnings overblown.

Here is my read: believe him. This is not a stunt. The people closest to this technology are worried, and they are saying so in public, in writing, with their names on it. That is the uncomfortable part, because if they mean it and still don't stop, what does that tell you?

Why Nobody Can Slow Down

Amodei's plan has three steps, and Anthropic only controls the first one. The second requires the whole industry to coordinate. The third requires other countries, China included.

Now think about who would have to agree. Elon Musk sued Sam Altman, though a jury dismissed his claims in May and he has vowed to appeal. The two biggest labs compete for the same researchers. And OpenAI has reportedly asked members of Congress whether a coordinated slowdown could run afoul of antitrust law. Critics are already calling the whole idea a cartel, or a ladder pull, where the biggest players kick the ladder away behind them.

And here is the part people skip. Amodei's own essay says pacing only works if democracies stay far ahead of autocracies, and it even calls for cutting off chip sales to China. Read that again: the plan to slow down depends on staying ahead.

The plan to slow down depends on staying ahead.

That is a classic prisoner's dilemma. Every lab looks at the others and thinks: if I slow down and they don't, I lose, to a rival or to China. So the safest move for each one is to keep going, which means everybody keeps going. Nobody is lying. Nobody is in control, either.

There is one real exception, and it deserves credit. Altman said an OpenAI IPO right now would be ill-advised and pushed it to 2027 at the earliest. That costs real money, and it is the one concrete brake anyone has actually pulled.

The Bet Wall Street Is Making

Now look at the money. The four biggest tech companies plan to spend around $725 billion on AI infrastructure this year, up 77 percent from roughly $410 billion the year before. Goldman Sachs projects about $1 trillion in 2027 and roughly $7.6 trillion cumulatively through 2031.

The detail that should worry you more than the size of the number is the accounting. Over the same four quarters, those companies bought about $434 billion worth of chips, data centers, and buildings, but recognized only about $149 billion of it as depreciation. The bill is coming, whether or not the revenue shows up to pay it.

Meanwhile, Anthropic is reportedly trying to go public before the November midterms, with sources putting the raise at up to $100 billion at a valuation near $2 trillion. Think about the pitch required to justify that number. It is not "we are being careful." It is "this is going to change everything, and if you are not in, you will regret it." Danger and power are the same claim, and that claim is worth two trillion dollars.

What It Means

If demand shows up, the bet pays off, and AI keeps getting better faster than most people are ready for. If it doesn't, a lot of that $725 billion will have gone to infrastructure nobody needed yet. Both outcomes are genuinely possible, and nobody, including the people making the bet, knows which one we get.

What I would watch instead of the stock price is who gains control over these systems. Not necessarily because the AI got dangerous, but because dangerous is a useful word if you want that control. Chip export rules and national-security carve-outs are where this likely goes next, not a lab quietly deciding to slow down.

The warnings are probably real. The race is definitely real. And when those two things point in different directions, the race wins.