The Pulse
Anthropic Lines Up $517 Billion in Compute Deals
Anthropic has entered agreements covering at least 14.8 gigawatts of compute capacity since October, according to an analysis by The Information. The contracts could cost as much as $517 billion over roughly a decade, although the company h
AI.info Team ·
Anthropic’s infrastructure plans now stretch across enough computing capacity to power a major city, but the largest number attached to them is not a check the company has written. An analysis by The Information estimates that agreements signed since October cover at least 14.8 gigawatts of potential compute capacity and could cost as much as $517 billion over roughly the next decade.
The figure comes from adding the upper limits of publicly announced agreements and prior reporting on contracts with cloud providers, chip companies and data-center developers. Anthropic has never publicly disclosed a total for its compute commitments, and the company may not use every site or server covered by the agreements. The $517 billion figure therefore describes the maximum value of a sprawling contract portfolio, not money already spent or a confirmed capital budget.
That distinction matters as Anthropic prepares for a possible initial public offering. Investors will want to know how much capacity the company can actually access, how much it must pay for, when that capacity becomes available and whether demand for Claude can support the expense. The size of the commitments also shows how AI companies are turning future revenue expectations into long-term claims on chips, electricity and data-center space.
Anthropic’s 14.8-Gigawatt Compute Book
The Information’s analysis says Anthropic has entered agreements for at least 14.8 gigawatts of capacity since October, in addition to the 1 to 2 gigawatts it had secured earlier, according to a person familiar with the matter. The agreements span cloud capacity, direct chip purchases and leases for data centers that will house hardware supplied by companies including Nvidia, Google, Broadcom and AMD.
Amazon and Google represent the largest portion of the reported total. Together, they account for about 11 gigawatts of capacity that The Information estimates will cost more than $300 billion over approximately 10 years. Anthropic’s April agreement with Amazon alone covers up to 5 gigawatts of new capacity and more than $100 billion in spending on AWS technologies over the next decade.
Anthropic says the Amazon agreement includes Trainium2 through Trainium4 systems, along with Graviton processors and future generations of Amazon’s custom silicon. The company says it already uses more than one million Trainium2 chips to train and serve Claude through Project Rainier. Amazon is investing $5 billion in Anthropic immediately, with the possibility of an additional $20 billion later.
“Our users tell us Claude is increasingly essential to how they work, and we need to build the infrastructure to keep pace with rapidly growing demand,” Dario Amodei, Anthropic’s chief executive and co-founder, said in the company’s April announcement. “Our collaboration with Amazon will allow us to continue advancing AI research while delivering Claude to our customers, including the more than 100,000 building on AWS.”
Google, Microsoft and SpaceX Expand the Supplier List
Google and Broadcom signed an agreement with Anthropic in April covering multiple gigawatts of next-generation tensor processing units. Anthropic said the capacity is expected to begin coming online in 2027 and will support future Claude models. The arrangement expands an earlier relationship under which Anthropic gained access to Google TPUs and made Google Cloud one of the platforms where customers can use Claude.
Microsoft entered the picture with a commitment from Anthropic to buy at least $30 billion in Azure computing capacity, according to The Information’s account. The deal covers roughly 1 gigawatt of Azure servers running Nvidia chips. Microsoft has backed OpenAI for years, making Anthropic’s agreement with Azure another example of the company spreading its infrastructure demand across suppliers that also serve its closest competitors.
Anthropic also agreed to use computing capacity from SpaceX. Under the arrangement described by The Information, Anthropic pays $1.25 billion per month through May 2029 to rent AI servers at SpaceX’s Colossus facilities, which The Information values at up to $40 billion — below the arithmetic on the monthly figure, because the first months are discounted. The deal includes a 90-day cancellation provision for either party, a term that makes the headline value less certain than a fixed, noncancelable purchase commitment.
Smaller providers add another large layer. The Information says Anthropic signed $80 billion in cloud contracts with Lambda and U.K.-based Nscale for at least 460 megawatts of capacity over six years. The report also points to a 20-year lease with Riot Platforms in Texas, a 401-megawatt facility being developed by TeraWulf in Kentucky, and Anthropic’s $50 billion plan with Fluidstack to build data centers in Texas and New York.
Why the $517 Billion Total Is Not a Spending Forecast
The calculation uses the upper limits of agreements whose structures differ widely. Some are cloud contracts, some are leases, some involve optional future capacity, and some depend on facilities receiving power, permits and hardware. The Information notes that Anthropic could end up using less than the agreed capacity, while some planned projects may arrive later than expected or never become operational.
Data-center construction introduces several points of uncertainty. Developers need land near substations and transmission lines, approvals from local governments and access to equipment that is already in short supply. Even after a contract is signed, a project can face delays in grid interconnection, construction or delivery of accelerators.
Anthropic’s SpaceX agreement illustrates another issue: a large nominal value can coexist with an exit clause. A payment schedule running to May 2029 produces a value of up to $40 billion, but either party can cancel with 90 days’ notice. The total is useful for showing the scale of the relationship, but it should not be read as guaranteed revenue for SpaceX or guaranteed expense for Anthropic.
The company’s previous planning figure provides another reference point. Anthropic told investors in December 2025 that it expected to spend about $180 billion on server rentals through 2029. The Information’s new tally is nearly three times that amount, although the two figures cover different periods and may include different categories of infrastructure. One describes a prior spending expectation; the other adds the maximum values of newer agreements that extend into the 2030s and beyond.
Claude’s Growth Is Driving the Buildout
Anthropic’s infrastructure push follows a sharp rise in usage of Claude, Claude Code and Cowork. The company said in April that its run-rate revenue had passed $30 billion, up from approximately $9 billion at the end of 2025. In May, when Anthropic announced a $65 billion funding round at a $965 billion post-money valuation, it said run-rate revenue had crossed $47 billion.
The Information later put Anthropic’s annualized revenue above $65 billion, compared with more than $40 billion for OpenAI as of July. Both companies are privately held and do not publish the same level of audited financial information as public companies, so the figures are operating estimates rather than standardized revenue disclosures.
Demand does not automatically translate into profitable compute consumption. Training a new model creates a concentrated burst of infrastructure use, while serving millions of users requires capacity that must remain available during peak periods. Anthropic’s April announcement acknowledged that consumer growth had strained reliability and performance for free, Pro, Max and Team users, particularly during busy hours.
The supplier mix also reflects a technical strategy. Anthropic uses Amazon Trainium, Google TPUs and Nvidia systems rather than relying on a single accelerator family. It has announced a 2-gigawatt AMD deployment expected to begin in 2027, giving the company another hardware option as it tries to secure enough capacity for training and inference.
IPO Investors Will Examine the Contracts
Anthropic’s possible IPO gives the compute total an immediate financial audience. Investors are likely to examine whether the company has committed to fixed payments, minimum-volume purchases or capacity options, and whether its revenue growth can cover those obligations without forcing the company to raise additional capital at unfavorable terms.
The company’s agreements also affect its operating flexibility. Long leases can secure sites and electricity before competitors do, but they can leave Anthropic paying for capacity that becomes unnecessary if model efficiency improves or customer demand slows. Cloud contracts offer faster access to hardware, while direct data-center leases can provide more control over deployment but expose Anthropic to construction and power-delivery risks.
OpenAI’s own infrastructure plans provide the comparison investors will make. The Information says OpenAI told investors it expected to spend about $750 billion on compute through 2030, up from a previous forecast of $665 billion, and has discussed a 30-gigawatt target by that year. Anthropic’s 14.8 gigawatts of newly identified capacity would be smaller, but its agreements frequently run beyond 2030, making simple comparisons difficult.
Anthropic has not confirmed the $517 billion aggregate, and no public filing currently provides an audited schedule of its compute commitments. For now, the most defensible description is narrower: since October, the company has assembled agreements covering at least 14.8 gigawatts, with an upper-limit value estimated by The Information at $517 billion. The number measures the scale of Anthropic’s infrastructure bet, while the contracts’ terms will determine how much of that bet becomes an actual bill.