Anthropic's $517B compute number is take-or-pay

Someone finally added it all up. The Information reported that Anthropic committed to roughly $517 billion of compute covering 14.8 gigawatts in the eleven months through August 2026 — a figure Anthropic has never published itself, assembled by aggregating the deals it announced one at a time. If you buy Claude tokens, the number is less interesting than the contract type sitting under it.

What actually happened

This is reporting, not a disclosure. The Information built the total from announced agreements, and coverage of the report lists the pieces: AWS at roughly $100B over ten years for about 5 GW of Trainium3, Google and Broadcom at about $200B over five years for another 5 GW, plus Fluidstack, Nscale, SpaceX and Lambda. Anthropic has not confirmed the total on the record.

Three things about that number matter more than its size.

It is an upper bound, not a bill. These are take-or-pay capacity reservations and long-duration leases, phased across years and in some cases past 2030. None of it sits on the balance sheet today. Contracted maximum is not cash spent.

It is a 2.9x jump from the last figure. The prior number that reached investors was around $180 billion in server rentals through 2029. The escalation happened in months.

The revenue side moved too. The Financial Times, citing investor briefings, reported a run rate climbing from roughly $9 billion to $65 billion in seven months, a second consecutive quarter of adjusted operating profit, and inference gross margins above 80% before distribution costs.

The timing is the awkward part. The aggregate landed the same week Dario Amodei published an essay asking frontier labs to slow the pace of capability gains. Both can be true — you can want slower model releases and still reserve every gigawatt you can get — but the spending is the part with a signature on it.

Why your vendor's compute contracts matter for your business

Take-or-pay means the capacity gets sold. A reservation you have to pay for whether you use it or not creates pressure to fill it. Historically that is good news for buyers: more capacity chasing demand tends to push token prices down, not up. The counter-case is Google, which doubled Gemini 3.8 Flash pricing effective January 1. Cheap tokens are a strategy, not a law.

80% inference margins are your negotiating floor. If margin before distribution is that high, there is room in the price. Enterprise commitments are negotiable. Ask.

Capacity that lands in 2027 does not serve your traffic in 2026. Most of this is construction and lease schedules. When a rate limit bites next quarter, the gigawatts in the press release are not the ones serving you. Build for the swap — keep prompts, evals, and logs in your own systems so a second provider is a routing change. We have written about that when Ramp's data showed how fast teams move between labs.

Sources: Investing.com on The Information's report, Forkast.