Anthropic has built a fast-growing business largely by selling its AI services based on how much customers use them. And—as the company’s landmark IPO prospectus revealed Monday—customers are using a lot of Anthropic’s AI services right now.
That’s the good news. The less-good news is that Anthropic is highly dependent on outside partners for the computing power to run its services. And as Monday’s prospectus also made clear, the disparity between the agreements through which Anthropic buys computing power, and the way it sells its AI services to customers, means it could get badly squeezed if business conditions change.
The AI lab’s revenue is growing extraordinarily fast, with first-half revenue rising 1,300% to $16.28 billion. About $13.1 billion of that, or roughly 80%, comes from enterprise customers paying consumption-based fees rather than individual subscriptions.
The revenue growth reflects the booming demand for AI services, as businesses use advanced large language models such as Anthropic’s Claude family of products to help with computer coding, accounting, and other tasks.
To fulfill that demand, Anthropic has signed massive infrastructure agreements with Amazon, Microsoft, Google, and other hyperscalers. Those hyperscalers, or cloud providers, own the data centers and GPUs that Anthropic’s AI models run on.
So far, the amount Anthropic pays these cloud providers for compute capacity has grown at a more modest clip than its exploding sales figures. The company’s cost of revenue for the tokens it serves customers was $2.4 billion in the six months ended June 2026, up 135% from the same period a year earlier. The company also spent a whopping $9.7 billion on computing and infrastructure costs for research and development of future models.
Currently, the company says it makes more than two dollars in revenue for every dollar it spends on computing infrastructure. But that could change if the company is forced to reduce token prices, a risk it explicitly warns about in its S-1 filing.
The mismatch
Anthropic’s infrastructure spending is committed well ahead of expected demand. The company has committed at least $475 billion in future cloud compute purchases with its major partners, including at least $389.9 billion pledged to Amazon, Google-parent Alphabet, SpaceXAI, and Microsoft alone, alongside deals with Nscale and Lambda. Much of that is owed regardless of how much usage its AI gets. When accounting for deals to lease computing chips, its total commitments soar to as much as $750 billion.
The demand side of Anthropic’s business enjoys far less certainty. As the company laid out in the S-1, much of its revenue lacks long-term commitments, with contracts allowing customers to reduce usage or terminate with limited or no notice.
That mismatch between how Anthropic generates revenue and its expected spending creates a pressure-filled balancing act in the years ahead. Anthropic’s growth has so far helped it keep pace with compute costs. Now the question is whether customers will keep spending at a rate that allows it to absorb these massive commitments.
“We must grow our customer base and expand revenue rapidly in order to generate the cash flows necessary to satisfy these commitments,” Anthropic said in its S-1.
The company also acknowledged that hitting those growth targets could be a challenge. “We expect our revenue growth rate to decline over time and to fluctuate from period to period,” the company wrote.
Recognizing this potential mismatch, the company has increasingly tried to move enterprise customers into non-cancelable, multi-year deals with fixed minimum spends in exchange for discounted token rates.
The company said that as of June 2026, it had about $10.8 billion in remaining performance obligations under these contracts. However, 75% of that revenue will be recognized over the next year, while Anthropic’s data center spending commitments extend well beyond that timeframe.
Anthropic also disclosed that it has significantly underestimated its compute requirements in the past, leaving it capacity constrained and impacting product development. It wrote that the timing of compute coming online may not align with demand, resulting in “periods of excess or underutilized compute and committed cash outflows that do not produce immediate revenue.”
At the end of 2025, for example, Anthropic noted that customer spend declined by about 20% compared to the prior two weeks, although it later bounced back.
Like other AI companies, Anthropic typically contracts for compute 12 to 18 months in advance. Its expenditures for this compute rise as capacity becomes available, and the risk for Anthropic is that even as revenue continues to grow, that growth could be too slow to absorb the next wave of infrastructure costs. In its filing, Anthropic explicitly warns that significant quantities of new compute coming online will compress its margins.
Significant upfront payments
The good news for Anthropic is that so far, its plan has worked. Revenue has kept pace with compute expenses, and in its second quarter, the company generated $974 million in operating income. It also held $112.4 billion in cash and cash equivalents as of late June.
Anthropic disclosed that its funding plans include operating cash flows, equity, debt, and partner financing. It warned, however, that its infrastructure build-out will increasingly require significant upfront payments before computing capacity comes online, threatening to eat into cash reserves even as financial performance improves.
The company is also not counting on ads to boost revenue anytime soon, stating that it wants to avoid pressure to optimize for engagement. It cited its governance structure as a public benefit corporation, which obligates it to adhere to a public benefit mission alongside shareholder interests.
Weighing AI demand against infrastructure spend has been a central issue across the industry as tech giants rush to build out data centers. Tech giants have spent considerable effort catching up to demand while betting on a future where enterprises and individuals continue to rely on AI services for a wide array of tasks. Amazon, Microsoft, Alphabet, and Meta are collectively expected to spend more than $700 billion on capital expenditures in 2026.
Many billions of that spending are being financed by debt, according to a report in June by England-based Exponential View. Oracle, in particular, stands out for its aggressive fundraising. Earlier this year, the company said it would need to raise between $45 billion and $50 billion this year alone to fund its infrastructure expansion.
Anthropic’s delicate balance between revenue streams and fixed obligations is one of many risk factors highlighted in its extensive S-1. Other risks include reliance on a small number of clients. Its top five customers accounted for roughly 25% of revenue in the first nine months of 2026.
Anthropic also disclosed incidents where its models gained unauthorized access to real-world systems during security testing, and estimated that 5% to 7% of its revenue this year came from customers later found to be violating its usage policies.
The post Anthropic’s S-1 makes clear it’s on the hook for billions in compute. Its customers can walk at any time appeared first on Fortune.
