Anthropic’s 318-page IPO prospectus, which the company made public today, makes one thing clear: Compute is existential to Anthropic’s business, and not securing enough of it is a significant risk factor.
“Our ability to develop, train, improve, and deploy frontier models depends on our access to enormous and increasing quantities of specialized compute,” the company wrote in the prospectus, which is also called an S-1, using the name of the form type the U.S. Securities and Exchange Commission requires companies to file ahead of an initial public stock offering. At the same time, “the market for compute is intensely competitive” and Anthropic competes with “some of the world’s largest and most well-capitalized technology companies,” the document said.
The company’s future computing costs are potentially massive and could swamp Anthropic’s ability to pay for them. In total, the company has committed to spending between $542 billion and $567 billion on future compute and infrastructure, the S-1 reveals. On Dec. 31, 2025, Anthropic signed a $417 billion compute deal with various providers. The deals start between 2026 and 2027, and lasts for one to 20 years. Anthropic also signed up for more capacity, starting between 2028 and 2031, totaling $125 billion to $150 billion.
Overall, compute is Anthropic’s number one operating expense, costing it $5.6 billion in the second quarter of 2026. But for now, the company generates a healthy amount of revenue from that spending. The company made $11.6 billion in revenue during the same quarter. That means it’s in the black as far as its revenue-to-compute ratio, a metric the S-1 says the company is focused on continuing to improve. It made a little over two dollars for every dollar it spent on compute.
Anthropic did not immediately respond to requests to comment for this story.
Anthropic Direct
Anthropic is securing compute from three main cloud service suppliers: Amazon, Google, and Microsoft. The S-1 also reveals deals with Elon Musk-owned SpaceX. From May to October 2026, Anthropic entered into a series of deals with SpaceX to access up to $120.5 billion in compute through 2029. (That compute is powered by Nvidia GPUs SpaceX owns).
The company has also built a “proprietary orchestration layer” that optimizes which workloads it runs on different chips across the different cloud providers, including AWS Trainium, Google TPUs, Nvidia GPUs, and AMD GPUs. This intelligent router of sorts took Anthropic “years of significant investment to build,” the company says. It helps Anthropic keep costs down, and reduces dependence on any one supplier.
The problem is, Anthropic’s relationship with Amazon, Google, and Microsoft is “multifaceted and complex,” the S-1 says. These suppliers are also its competitors, who need compute for their own AI products, so they could “prioritize their own compute needs.” The companies may also “gain insight” into how Anthropic’s business runs, and its compute requirements, that “could inform their competitive strategies.” Adding even more wrinkles to these circular relationships, Microsoft, Amazon, and Google have all made “significant” investments in Anthropic, and may continue to do so in the future. The kicker: they’re also customers of Anthropic’s products, using its models.
To reduce dependence on its competitors, Anthropic is pursuing a new compute strategy called Anthropic Direct, which means essentially standing up its own data centers, the filing says. Anthropic expects its Direct initiative to “represent a growing share” of its “compute investments over time,” yielding “long-term economic advantages.”
Starting around Sept. 30, 2026, the company began entering into purchase commitments and lease agreements for Anthropic Direct. Its third-party partners will secure the real estate, power, and facilities, and manage the data centers on an ongoing basis. Anthropic will lease the compute it needs from those partners. Its rent payments over the five-year lease are $125.2 billion, excluding interest.
Anthropic’s largest infrastructure partner in 2026 is Broadcom, at $42 billion in potential financing. Google follows at $20 billion, and Amazon at $15 billion. Broadcom’s stock is up 3% today on the news.
The downside of this approach is that Anthropic needs to pay out these costs in advance, without getting a quick return on investment. As a result, the company said in its S-1 that it expects to “incur significant cash outflows” that could “adversely affect our liquidity, financial condition, and results of operations.”
Powering the data centers is another major concern. Anthropic expects “substantial energy demands,” according to the S-1. To that end, it has signed power generation and infrastructure deals, although it did not specify the type of deals or power source. The biggest risk is connecting the power to the data center, which could be delayed due to permitting, legal challenges, supply chain disruptions, and more, the company said.
Multiple other shortages could affect Anthropic’s success, including the supply of chips and data center equipment. But Anthropic—and all AI companies—will need to navigate these challenges, or face the possibility of failure.
“If we are unable to successfully execute on our data center infrastructure initiatives, experience significant delays or cost overruns, or find that the infrastructure we lease, deploy, or build does not meet our evolving needs, we may not realize the expected benefits of these investments, which could adversely affect our business, financial condition, and results of operation,” the filing says.
The post Anthropic plans to spend up to $567 billion on compute. Its IPO prospectus reveals plans to set up its own data centers appeared first on Fortune.
