Anthropic plans $518 billion for data centers after a $4.6 billion revenue year

A platform lead with Claude inside a support queue will go through the prospectus coverage looking for one line item: next year's invoice. Reuters' account of the draft, which was filed confidentially with the SEC in June ^[1], does not describe any change to Claude pricing or contracts. Any forecast has to come from the other figures.

The $518 billion plan ^[4] is about 113 times 2025 revenue ^[1]. That multiple overstates the gap, because revenue has grown quickly since. Second-quarter 2026 revenue of $11.5 billion ^[6] was 2.5 times the whole of 2025 ^[2]. At that pace, the plan equals about 11 years of sales ^[3].

The operating profit changes what a price rise would be paying for. Revenue now covers the cost of running the business, and Anthropic expects that to hold next quarter ^[6]. The money still needed is for the build. I'd expect pressure on Claude buyers to show up in terms before it shows up in list prices, as commitments and longer contracts. Committed revenue is easier to borrow against than usage that can stop.

In my view, most teams on Claude think of themselves as the locked-in party, with prompts tuned to one model and evals written against its habits. Here is what customers actually do, according to the prospectus: many of Anthropic's biggest ones have not signed long-term contracts and could stop spending at any time ^[9]. The Financial Times figure of a quarter of Q2 revenue from two clients ^[7] works out to about $2.9 billion ^[4]. Meta was reported in August to project spending of up to $10 billion a year with Anthropic ^[8].

Release pace is the other cost that falls on buyers. Dario Amodei, Anthropic's chief executive, recently called for AI companies to slow the pace of new development ^[12]. Anthropic released Opus 5.5 last week to keep up with OpenAI, according to Engadget ^[13]. For a team that ties product behavior to a specific model version, each release means another round of evals.

The decision comes down to two things: how hard Claude would be to swap out of your product, and whether you pay as you go or under a committed contract. Easy to swap and paying as you go is the least exposed position for a buyer, and it is the customer behavior Anthropic's filing lists as a risk to itself ^[9]. A committed contract on an easy-to-swap integration means you sold flexibility for a discount, and the discount should at least cover what a switch would cost. Hard to swap and paying as you go leaves you open to a price change you cannot act on. Hard to swap and committed means concentration on both sides of the contract: you rely on one vendor, and the vendor relies on a few large buyers ^[7]. The cheapest way out of either hard-to-swap box is engineering work, meaning an abstraction layer over the model calls and an eval suite that already runs against a second model.