Anthropic’s S-1 shows it has committed to spending between $542 billion and $567 billion on future compute and infrastructure. For context, that’s nearly half the $965 billion valuation Anthropic carried into its last funding round in May was based on at the time — and it’s more than a hundred times the roughly $4.6 billion Anthropic actually made in revenue in all of 2025. No AI lab has put a bigger number on the table for how much it expects to spend just keeping Claude trained, running, and scaling.
Where the number comes from
The $542-567 billion figure is a forward-looking total across all of Anthropic’s compute and infrastructure obligations, but the single biggest building block inside it is a $417 billion commitment the company locked in on December 31, 2025. According to the filing, that commitment covers “aggregate future minimum payment obligations” tied to new leases and amended agreements for compute and data center capacity, primarily with its existing cloud and chip partners. The deals span terms from under a year to as long as 20 years, with payments commencing between 2026 and 2027.
Inside that $417 billion, the biggest single pieces go to the companies Anthropic already leans on most heavily: at least $110 billion to Amazon through April 2036, and at least $114.5 billion to Google through July 2033, both described in the filing as owed “regardless of actual usage.” Broadcom, which supplies custom chip and networking hardware, accounts for roughly $161 billion more. Microsoft, brought in more recently as a cloud and compute partner, is on the hook for a smaller but still substantial $44.9 billion.
Anthropic frames this as the cost of running what it calls a “multi-chip compute architecture” spanning AWS Trainium, Google TPUs, NVIDIA GPUs, and AMD GPUs, treating compute as “a single, fungible pool” sourced from multiple partners rather than betting on any one chipmaker or cloud provider. It’s the same diversification logic that led the company to add AMD as a fourth major chip supplier earlier this year, and to lock in roughly one million Google TPUs worth tens of billions of dollars the year before that.
Why the number is so large
Part of the answer is simply that Anthropic’s growth has outpaced its own forecasts, repeatedly. CEO Dario Amodei has said the company planned for 10x growth and instead saw 80x annualized growth in a single quarter, a compute shortfall severe enough that it forced rate limits, session caps, and an emergency compute-sharing deal with xAI. The S-1 itself admits as much, telling investors the company has “significantly underestimated its compute needs in the past” — which helps explain why the forward commitments are now so large: Anthropic is trying not to get caught short again.
The other half of the story is that compute spending, while enormous in absolute terms, has actually been getting more efficient relative to revenue. Anthropic is now generating 5.2x more revenue per dollar of compute spend than it was at the start of 2025, with the ratio improving from $0.40 of revenue per compute dollar in Q1 2025 to $2.08 by Q2 2026. That’s the number Anthropic wants investors to focus on — not the sheer size of the $567 billion commitment, but the trajectory suggesting the commitment will look smaller relative to revenue with each passing quarter, assuming growth holds up.
Anthropic Direct: building its own data centers
The most notable new disclosure in the S-1 isn’t a dollar figure at all — it’s a strategy shift. Anthropic says it’s pursuing something it calls “Anthropic Direct,” which amounts to standing up its own data centers rather than renting all its capacity through AWS, Google Cloud, and other partners. The filing says Anthropic Direct is meant to give the company a more “flexible use of capital” by separating its commitments for data center capacity from its commitments for the compute hardware deployed inside those buildings, and that the company expects the initiative to “represent a growing share” of its compute investments over time, yielding “long-term economic advantages.”
The mechanics work like this: Anthropic’s third-party partners secure the real estate, power, and facilities, and handle ongoing data center management, while Anthropic leases the compute capacity it needs from them. As of mid-2026, the company had already begun entering purchase commitments and lease agreements under this model, with rent payments on one five-year lease alone totaling $125.2 billion, excluding interest. Separate reporting has pointed to Anthropic signing more than a dozen non-binding letters of intent for direct U.S. data center leases totaling over a gigawatt of capacity, alongside new leases for data center capacity in Australia and Canada as part of a broader push to reduce how much of its compute infrastructure sits concentrated in the United States.
The strategic logic is straightforward: every dollar Anthropic pays AWS or Google Cloud for capacity comes with a markup, and Amazon and Google are also two of Anthropic’s largest shareholders, which makes the dependency look less like a simple vendor relationship and more like something investors are already flagging as a conflict of interest. Owning more of its own infrastructure, even indirectly through long-term leases, gives Anthropic a path to lower unit costs over time and less exposure to any single supplier’s pricing or capacity decisions — the same motivation that’s driven it to diversify across four different chip architectures rather than standardize on one.
What it means heading into the IPO
A $567 billion compute bet, layered on top of an $80.7 billion net loss for the first half of 2026, is the kind of number that makes the AI infrastructure arms race tangible in a way revenue growth charts don’t. It’s also, by Anthropic’s own account, a bet it’s making more efficiently with each passing quarter, and one it’s trying to de-risk by building owned infrastructure capacity rather than staying fully dependent on hyperscaler rental agreements. Whether that bet pays off before the committed spending comes due is likely to be one of the central questions investors weigh once ANT starts trading.