Anthropic’s IPO prospectus shows that the company behind Claude has become an infrastructure customer of enormous scale: with Broadcom, Google, Amazon, Microsoft, and several GPU providers, it has racked up compute commitments that, by various estimates, range from roughly $518 billion to $637 billion. The filing, registered with the SEC on October 26, 2026, also describes a hybrid model in which public cloud coexists with dedicated data centers leased and operated by third parties.
Anthropic’s infrastructure commitments in 30 seconds
- Anthropic filed its S-1 on October 26, 2026, and plans to list on the Nasdaq under the ticker ANT.
- As of June 30, it had $265.6 billion in non-cancelable purchase commitments and $321.4 billion in leases that had not yet begun.
- Its largest contracts are with Broadcom (TPUs), Google, Amazon, and Nvidia GPUs accessed through third parties.
- It combines public cloud with dedicated data centers leased from third parties, and acknowledges it previously underestimated its compute needs.
Portal Financiero detailed the technical architecture, and Noticias.ai explained how Anthropic splits its AI workloads across four chip families: Trainium, TPUs, Nvidia GPUs, and AMD GPUs, tied together by its own orchestration layer. For the cloud market, the most relevant part of the filing is who Anthropic pays, how much, and under what terms.
Who gets the contract
According to The Motley Fool’s summary, the named commitments are as follows (approximate dollar figures):
- Broadcom: $161.2 billion for five years of TPU capacity, through equipment leases. Portal Financiero cites a different figure, around $125.2 billion in projected interest-free lease payments for a March 2026 agreement, so the amount depends on the contract and the cutoff date used.
- SpaceXAI: up to $120.5 billion for Nvidia GPUs through 2029, with termination clauses once the initial terms expire.
- Google: $114.5 billion between 2026 and 2033.
- Amazon: $110 billion between 2026 and 2036.
- Microsoft: $44.9 billion.
- Nscale and Lambda: $44.6 billion and $41.3 billion, respectively, for access to Nvidia GPUs, payable only once the capacity is delivered.
Reuters, cited by Yahoo Finance, puts the total at around $518 billion and says roughly 80% of that is binding, non-cancelable obligations. Portal Financiero puts it at between $542 billion and $567 billion based on published analyses, while The Motley Fool puts potential commitments as high as $637 billion through 2036, including contingent agreements. The differences come down to different cutoff dates and criteria, and to the fact that the filing itself leaves some totals blank.
Public cloud, leased capacity, and owned equipment
The model Anthropic describes is neither that of a pure cloud customer nor that of an operator that owns its own data centers. It will keep using cloud providers, but wants to add capacity in dedicated facilities leased and operated by third parties, under agreements that can include land, buildings, and power. In some cases, it buys the equipment directly. According to Decoding Discontinuity, the leases will hit the balance sheet as facilities come online between 2026 and 2028, and Google, Amazon, and Broadcom are offering up to $77 billion in potential financing, none of which had been drawn as of June 30.
For Amazon and Google, which also develop their own models, Anthropic is a major customer of both their in-house chips and their cloud capacity. For Anthropic, the model means taking on fixed payments for years to come: according to Reuters, most of the obligations are binding, though some contracts, such as those with Nscale and Lambda, are only paid once the capacity is delivered.
The filing also adds context to the debate over local AI. Cloudnews.tech has already looked at the case for paying per token versus buying a 24GB GPU; Anthropic’s numbers show the scale on the other side of that argument: hundreds of billions of dollars committed to keep up the capacity needed to serve frontier models.
Frequently asked questions
How much has Anthropic committed to infrastructure?
Depending on the cutoff used, somewhere between roughly $518 billion and $637 billion; as of June 30, 2026, it totaled $587 billion in purchases and leases that had not yet begun.
Which providers does Anthropic have contracts with?
The largest are Broadcom, SpaceXAI, Google, Amazon, and Microsoft, plus Nscale and Lambda for Nvidia GPUs.
Is Anthropic building its own data centers?
It combines public cloud with dedicated centers leased and operated by third parties, and in some cases buys the equipment directly.
Why do the figures vary so much?
Because the different analyses use different cutoff dates, some include contingent agreements and others don’t, and the filing itself leaves some totals blank.
Sources:
- Portal Financiero: Anthropic prepares AI infrastructure with TPUs, GPUs, and its own data centers
- Noticias.ai: Anthropic splits its AI across four chip families
- SEC: Anthropic PBC’s S-1 registration statement
- The Motley Fool: 5 Key Numbers From Anthropic’s S-1 Filing
- Decoding Discontinuity: Anthropic S-1, compute commitments
- Yahoo Finance: Anthropic’s S-1 Is Here. The $518 Billion Commitment Is the Real Story
