Anthropic, the company behind Claude, has signed computing agreements that could cost up to $517 billion over roughly the next decade, according to The Information's September reporting. The expansion puts a concrete price on a tension its chief executive, Dario Amodei, has described himself: reserve too little capacity and growth becomes harder; reserve too much and the bills can outrun demand.

The reported total describes a decade of commitments

The $517 billion figure is an outside tally, rather than an aggregate disclosed by Anthropic. It describes potential spending across agreements, not money already spent or a payment due today. Keeping the words “up to” attached matters: a maximum commitment is a different claim from a settled bill.

The physical expansion is striking on its own. Before October 2025, Anthropic had between one and two gigawatts of computing capacity under contract. It subsequently added at least 14.8 gigawatts, according to Data Center Dynamics' account of the reporting. That is additional capacity, not the company's total, and contracted capacity should not be confused with servers already operating.

A gigawatt measures power. In this context, it conveys the scale of the infrastructure reserved to run computers, rather than a direct measure of how capable Claude is. More reserved power does not, by itself, tell us how efficiently those computers will operate or how much revenue their work will generate.

Anthropic's own infrastructure announcement confirms relationships with multiple suppliers. The reported aggregate is therefore best understood as a portfolio of capacity arrangements, with the headline number compressing their different timelines into one figure.

The older budget covers a shorter period

There is an obvious comparison with Anthropic's earlier plans. In December 2025, it showed investors a server-leasing budget of about $180 billion through 2029, The Information reports.

The new upper estimate is close to three times that amount. But calling it a tripling of the same budget would obscure a crucial distinction: one figure runs through 2029, while the other stretches across roughly a decade. The comparison illustrates the scale of the commitments; it does not establish a threefold increase in annual spending or a like-for-like budget overrun.

For understanding the business, the missing detail is the schedule. How much becomes payable in each year, and how closely does that track capacity becoming usable? An aggregate cannot answer those questions.

Amodei's warning was about getting the timing wrong

In his February 2026 interview with Dwarkesh Patel, Amodei said competitors buying computing capacity “don't really understand the risks they're taking.” He described a conditional scenario in which revenue growth arriving a year later than expected could be enough to bankrupt a company that had committed too aggressively.

That was a warning about a possible mismatch, not a prediction that Anthropic would fail. It also explains why a fast-growing AI company can make enormous reservations while recognising the danger. The decision depends on demand several years ahead, while the consequences of a mistaken forecast may arrive on a contractual timetable.

Our reading is that the interesting question is how much room Anthropic has to be wrong. Knowing the risk does not remove it. Nor does a large commitment alone establish that management has misjudged demand.

Revenue growth strengthens the case, but is not cash available

There is substantial evidence behind the expansion. Anthropic's annualized revenue run rate rose from about $9 billion in late 2025 to $65 billion in July 2026, according to CNBC's reporting on the company's figures.

A run rate extends a recent pace of sales over a year. It is not a completed year's revenue, a profit figure or money sitting in the bank. The July observation supports the case that demand had grown sharply; it cannot guarantee the pace over the life of the agreements.

Dividing the $517 billion upper estimate evenly across ten years gives $51.7 billion a year, or roughly $52 billion. Against July's $65 billion run rate, that is a useful illustration of scale. It is not a cash-flow forecast: it assumes even payments, holds revenue at one historical pace and leaves other costs out of the comparison.

Why it matters: demand and payments can move separately

Reporting describes take-or-pay reservations, meaning payments can fall due even when the reserved capacity is not fully used. The exact terms of individual agreements matter; the headline total does not reveal all their flexibility.

The implication is a trade-off between securing room to grow and preserving room to slow down. If demand keeps rising quickly, reservations can support expansion. If growth disappoints, the same commitments can become a burden before the long-term opportunity disappears.

For customers, investors and anyone following the economics of AI, the useful measure is therefore how closely usable capacity, paying demand and payment obligations line up. The biggest number attracts attention; the timing determines how manageable it is.

Should Anthropic accept the risk of turning customers away to preserve financial flexibility, or commit ahead of demand and accept the cost of being early?