Dario Amodei published We Must Pace the Frontier on September 12. Its core line is short: “We must slow the pace at which we improve the capabilities of AI models.”

Sam Altman agreed on X within hours. Elon Musk wrote “Dario is right.” Two months ago Amodei was the outlier who wouldn’t sign Jensen Huang’s open-weights letter. Now OpenAI is matching his commitments.

Meanwhile, The Information had reported six days earlier that Anthropic signed $517 billion in compute deals in eleven months. That’s 14.8 gigawatts. The number it had given investors was about $180 billion, through 2029.

A lot of people called that hypocrisy. I think it’s the most useful thing Anthropic has told its customers all year.

The spending is the roadmap

Read the essay carefully and the target is narrow. Amodei wants the rate of capability improvement at the frontier to slow, with outside evaluators confirming that each step is safe. He says explicitly that pacing does not mean halting training.

The $517 billion buys something else. Capacity to serve models to paying customers, bought in a hurry because demand for Claude Code and Cowork caught Anthropic by surprise this year, according to the reporting.

So here is how I’d read the two headlines together if I were a CTO. The models you can use today are staying, and there will be far more capacity to run them. The thing that might arrive later is the next big jump.

If your AI plan starts “once the next model is out,” that plan just got weaker. If it starts with what Claude already does, nothing changed. Good news, honestly.

What the pacing plan actually commits anyone to

Three steps. Embedded evaluators from groups like METR, with “desks in our offices, access badges, and company laptops.” Common safety standards and limits on the rate of progress among labs in democratic countries. Narrow coordination with China on things like biological weapons.

Only the first step is something Anthropic can do alone, and it’s the only one it has committed to. Altman’s reply was specific about the same step: “Committing to having independent evaluators with employee-like access is a great idea, and we will do the same.”

The essay also puts numbers on its own worries. One to two years of interpretability progress it wants time for, and a three-to-five-year window to widen America’s lead by refusing chip sales to China and cracking down on distillation.

Steps two and three need governments. Which brings us to Monday’s phone call.

The other side showed up on a speakerphone

At the All-In Summit on September 14, Donald Trump called Jensen Huang while Huang was onstage. Huang put him on speaker and asked for an extra mic. Trump said the backlash against AI “could be political people. It could also be China,” and called it a hoax. Huang answered, “You’re right. We’re not going to let that happen, sir,” TechCrunch reported.

And Meta had already picked its lane in August, releasing Muse Glimmer, a 30-billion-parameter open-weight model, while Mark Zuckerberg pressed Washington to lower barriers for open-source AI.

Labs asking for pacing, a White House calling the worry a hoax, Meta shipping weights. None of that resolves this quarter. Don’t hold your own projects hostage to it.

Two things that do change for you

Capacity first. A company doesn’t sign $517 billion in compute unless demand outran what it had. Capacity has been tight and will stay tight while those gigawatts get built. Any workflow that depends on Claude should already know what happens when the API returns a 529 overloaded error. A queue, retries with backoff, and a list of jobs that can wait an hour without anyone noticing.

Then verification. If the biggest labs accept outside evaluators with employee-level access, you’re allowed to ask your own AI vendors for less than that. Who tested the claim? With what access? Can you read the report? Regulators are already asking, as the FTC did in the Workado case.

At IQ Source we build Claude into operations that can’t stop, and we design for overload before the first 529 ever reaches production. If your team is waiting on the next model to begin, let’s begin with this one.

Plan your Claude integration