Anthropic Wants the Frontier to Slow Down. It Just Signed $517 Billion to Build Faster.

Anthropic signed compute contracts worth a combined $517 billion over the eleven months through August, covering 14.8 gigawatts of capacity across more than a dozen providers, according to Forkast's rundown of The Information's reporting. Google and Broadcom account for roughly $200 billion of it on a five-year deal, AWS for over $100 billion, and the rest is spread across Fluidstack, Nscale, SpaceX, Lambda, and smaller providers. Two days before that figure surfaced, the company's CEO, Dario Amodei, published an essay called "We Must Pace the Frontier" arguing the industry should deliberately slow down how fast model capability advances.

Commentators have jumped on the obvious irony: a company asking the field to slow down just wrote the biggest infrastructure check in its history. Investing.com's writeup frames it exactly that way, and quotes D.A. Davidson analyst Gil Luria noting that nobody making these commitments has actually agreed to spend less or delay an IPO. I don't think that's the real story. Buying capacity and pacing capability are two different variables, and a company can rationally do both at once: build the plant while agreeing not to race on how fast the product line improves. The revenue math backs the capacity bet on its own terms. Anthropic's annualized revenue went from $9 billion at the end of 2025 to $65 billion by the end of July, TechCrunch reported, adding $18 billion in just the two months of May and June. A company growing revenue that fast needs the compute regardless of what it believes about the pace of the underlying research.

Here's the steelman for treating the spending itself as the concerning part. Most of these are take-or-pay contracts, meaning Anthropic owes the minimum whether it uses the capacity or not, so $517 billion is a real financial obligation, not a wish list. If usage growth stalls, or a cheaper open-weight model erodes the premium Anthropic charges for frontier capability, the company is still on the hook for gigawatts it committed to years in advance. That is a legitimate risk, and it is the same circular-financing worry that has dogged every hyperscaler capex number this year: labs promising future revenue to justify present infrastructure spend that only pays off if that future revenue shows up.

But look at what the pacing plan actually asks for, because that is where the real free-market question sits, not in the spending. Amodei's essay lays out three steps. Step one is unilateral: Anthropic will embed outside evaluators in its own buildings with employee-level access, which costs Anthropic money and constrains nobody else. Step two is where it gets interesting: frontier labs jointly agreeing on common safety standards and shared limits on how fast capability is allowed to grow. Amodei's own essay says that step needs "government support to address antitrust concerns." Read that plainly: competitors coordinating on the pace of their own product improvement is the textbook definition of a cartel, and the essay itself acknowledges that making it legal requires Washington to carve out an exemption.

That is the part worth watching skeptically, and it has nothing to do with hypocrisy about compute spending. A government-blessed agreement among the two or three labs big enough to sit at that table, on how fast the frontier is allowed to move, changes the competitive dynamic that has done more for AI affordability this year than any single act of regulation: labs racing each other on both capability and price. Open-weight model providers and price wars between GPT, Gemini, and Claude are why frontier-adjacent capability now ships at a fraction of last year's cost. Slow that race down by mutual agreement, even for good-faith safety reasons, and you slow the price competition that rides on top of it. That's a real cost, and Amodei's essay does not weigh it, because it isn't framed as an economic question. It should be.

None of this changes what a business owner evaluating AI vendors should do this week, and if anything it argues for the same discipline I already push in every AI readiness assessment: do not build your cost model on the assumption that frontier pricing keeps falling at the rate it has for the past two years. That rate was set by a specific competitive dynamic, and the biggest player in the field just proposed, in writing, to slow the part of that dynamic it controls. Keep your options open: test the open-weight alternative, know what you'd pay to self-host the workloads that matter, and treat any single vendor's price trajectory as a forecast, not a promise. If you want a second opinion on where your AI spend is exposed to that kind of shift, reach out.

Related reading

• Cognition's Cheap Coding Model Wins the Benchmark It Chose. It Loses the One That Matters Most.
• OpenAI's Agents Keep Escaping Testing. Its Fix Also Draws the Line Around Itself.