$517B and 14.8 GW in 11 Months: The Full Anatomy of Anthropic's Compute Bet

Eleven months ago, Anthropic was a well-funded frontier lab whose biggest infrastructure liability was a single cloud partnership. Today, according to a September 6 analysis by The Information’s Valida Pau, the company has quietly assembled one of the largest private compute portfolios in the history of the industry: agreements covering at least 14.8 gigawatts of capacity and totaling up to $517 billion over the next decade — signed in just eleven months, almost all of it since last October.

The scale deserves a moment to sink in. As recently as early 2025, an entire hyperscale data center campus was measured in hundreds of megawatts. Anthropic’s contracted portfolio is now the equivalent of roughly fifteen of those campuses. If it were a power plant fleet, it would rank among the larger electricity producers in a mid-sized European country. And unlike a utility, none of these commitments appear as traditional debt on Anthropic’s balance sheet — they are take-or-pay capacity reservations, contractual promises to buy compute that the company has not yet paid for and, in most cases, has not yet used.

The deal-by-deal anatomy

The Information’s tally stitches together a string of announcements that individually made headlines but had not been summed:

• AWS / Amazon — $100 billion over ten years, securing up to 5 GW of Trainium3-based capacity for training and serving Claude.
• Google and Broadcom — a $200 billion, five-year agreement for multiple gigawatts of next-generation TPU capacity, expected to come online starting in 2027.
• Fluidstack — $50 billion in data center capacity, part of the $50B U.S. infrastructure push announced in November 2025.
• Nscale — $45 billion, including the $45B compute deal signed August 26 after Microsoft and Google reportedly passed on it; Nscale’s contracted revenue backlog jumped from $51B to roughly $103B on the strength of it.
• SpaceX — $45 billion, roughly $1.25 billion per month, anchored on the Colossus 1 data center and its follow-on capacity.
• Lambda — $35 billion in neocloud capacity, including approximately $45 billion over six years on 460 MW of NVIDIA Vera Rubin GPUs under a separate long-term arrangement.

Add them up and the portfolio spans three hyperscalers, two neoclouds, one rocket company, and at least four chip architectures: Amazon’s Trainium3, Google’s TPU (co-designed with Broadcom), and NVIDIA’s Vera Rubin generation across the independent providers.

Why “take-or-pay” is the load-bearing phrase

The analysis frames every one of these as take-or-pay obligations — Anthropic pays for the capacity whether or not it consumes it. That structure is what lets a still-private company sign half a trillion dollars of commitments without raising half a trillion dollars: the payments are spread over five to ten years, and the vendors — who themselves finance data center construction against these contracts — carry the construction risk.

But the same structure concentrates enormous demand risk on Anthropic. If Claude’s revenue growth slows, the company still owes the money. The Information notes the commitments amount to a bet worth roughly 30% of Anthropic’s valuation — a ratio that would be unremarkable for a capital-intensive utility and is extraordinary for a software company. The implicit wager is that demand for frontier intelligence grows faster than any historical compute curve, and that Claude’s enterprise traction converts contracted watts into contracted revenue before the bills come due.

There is a second-order effect worth watching: vendors are now financing the AI buildout against Anthropic’s paper. Nscale’s pre-IPO pitch to investors — a $3.5B raise including $2B from NVIDIA — leans heavily on its $103B backlog, of which Anthropic is the anchor tenant. Amazon, Google, and SpaceX are all, in effect, underwriting Anthropic’s growth with their own capital. The industry’s circular financing is well documented at the OpenAI-NVIDIA-Oracle layer; the September 6 tally shows the same architecture has arrived at Anthropic’s layer of the stack.

The context: an industry-wide arms race in gigawatts

Anthropic’s 14.8 GW does not exist in a vacuum. OpenAI’s compute commitments have been estimated well beyond that figure across its own multi-vendor portfolio, and Google and Microsoft are building internal fleets measured in gigawatts for their in-house models. What distinguishes the Anthropic portfolio is that virtually all of it was assembled in eleven months, by a company that does not own its own chips, and that must therefore rent capacity from anyone who will build it — hyperscaler, neocloud, or rocket company.

That rental strategy has a strategic cost. Every provider in the stack above Anthropic also competes with it or negotiates with it from a position of strength. The AWS deal makes Amazon one of Anthropic’s largest suppliers and its boardroom an awkward place; the Google deal gives a direct competitor’s parent company both revenue and visibility into Anthropic’s capacity planning. The SpaceX arrangement, meanwhile, ties Claude’s training runs to a data center operator whose own flagship customer is xAI — a fact that reportedly shaped the Colossus capacity discussions all summer.

What it means for the market

Three takeaways stand out from the tally.

First, the frontier is now gated by contract, not by chips. The binding constraint on the next generation of Claude models is not whether NVIDIA can fab GPUs — it is whether Anthropic’s counterparties deliver power, buildings, and interconnects on schedule. Several of the 14.8 GW are 2027-2028 deliveries, which makes Anthropic’s model roadmap hostage to construction timelines in Texas, upstate New York, and wherever Colossus’s successors land.

Second, private labs’ true leverage is far larger than their balance sheets. A valuation of a few hundred billion dollars against over half a trillion in purchase obligations means the market is pricing Anthropic’s equity as a call option on AI demand. bulls see a moat; skeptics see a utility with a startup’s cost of capital.

Third, the compute market has matured into a traded good. Gigawatt-year contracts, equity stakes as deal sweeteners, revenue backlogs as IPO collateral — the apparatus of commodity finance has fully arrived in AI infrastructure. Swiss Re’s same-week estimate that AI data centers will generate $200B in insurance premiums by 2030 is the risk-market catching up to what the tally makes explicit: this is now a half-trillion-dollar asset class with a single lab’s product cycle at its center.

The bottom line

The Information’s accounting reframes a year of scattered mega-announcements into a single, coherent picture: Anthropic has spent eleven months converting its balance sheet lightness into contractual heaviness, locking 14.8 GW and up to $517 billion of compute before a single one of those watts has trained a shipping model. Whether that proves to be the shrewdest capacity trade in corporate history or the largest unhedged demand bet ever made by a private company now depends on one variable — whether the world’s appetite for Claude grows as fast as the company has promised its suppliers it will.