Anthropic has committed to spending $200bn with Google Cloud over five years, according to a report from The Information citing a person familiar with the deal. The figure would account for more than 40% of the cloud-revenue backlog Google disclosed to investors last week — making the AI lab Google Cloud's single largest contracted customer by a wide margin. Alphabet shares were up about 2% in extended trading after the report. Anthropic declined to comment; Google redirected queries to Anthropic.
The commitment underpins an April agreement between Anthropic, Google, and Google's chip partner Broadcom for multiple gigawatts of TPU capacity, expected to come online from 2027. It is the latest in a stack of compute deals Anthropic has lined up over the past year — a multi-year arrangement with CoreWeave last month, and a roughly one-gigawatt commitment to AWS's Trainium chips by year-end. Claude is now trained and served across Trainium, TPU, and Nvidia GPU silicon simultaneously.
Contracts involving Anthropic and OpenAI between them now account for more than half of the roughly $2tn in disclosed multi-year backlog across AWS, Microsoft Azure, and Google Cloud. Two model labs, three hyperscalers, more than $1tn in committed forward spend — the AI-cloud relationship has graduated from "fast-growing customer" to "the line that explains the chart".
The cycle
Three weeks before this $200bn outbound commitment, Alphabet announced an investment of up to $40bn going the other way: into Anthropic itself. The same shape holds on the AWS side, where Amazon is a multi-billion-dollar Anthropic investor and Anthropic is a multi-billion-dollar AWS customer. This is what we mean when we say AI capital is moving in circles. A hyperscaler invests in a model lab; the model lab signs a long-dated supply agreement with the hyperscaler; the hyperscaler discloses that agreement as committed backlog; the lab discloses the investment as cash on hand. The arithmetic each side reports is honest. The arithmetic of the system is harder to read.
It is worth distinguishing genuine demand from accounting choreography. The demand is real. Claude usage has grown enough that Anthropic genuinely needs gigawatts of additional compute, the chips and the power supplying them are physically constrained, and the build-out is the literal industrial expansion of the decade. But the financing structure layered on top of that real demand means the same dollar can show up as a contractual commitment on one balance sheet and as an equity stake on another. The economic activity is a strict subset of the headline numbers; the discount factor between the two is what investors and supervisors should be modelling. The Financial Stability Board, which warned this week about the AI build-out's exposure to private credit, was, in effect, asking the same question.
One detail worth holding on to from this particular deal: the spend is locked in for five years; the chips it is paying for do not start coming online until 2027. The cash and the silicon are on different schedules. So is the demand they are intended to serve.
