Anthropic's compute spending: who it pays, for which chips, on what terms

Anthropic has reportedly committed to spend at least $518 billion on AI infrastructure over roughly a decade, and about 80 percent of that is said to be non-cancellable or payable whatever the usage. The money flows mostly to Google, Amazon, Microsoft and Broadcom, with smaller deals alongside. These figures come from reported filings, not audited accounts.

This post sets out what is publicly known, what the company has said itself, where sources disagree, and why the contract structure matters as much as the headline number. Figures are as of 19 October 2026, and each one is labelled by how solid it is: company statement, reported by a news outlet, outside estimate, or simple arithmetic on a reported figure.

Key numbers at a glance

Every line below carries a source type and a date. Private company figures are unaudited.

• Total reported commitment: at least $518 billion over roughly a decade. Reported by Reuters from a confidential IPO prospectus, relayed by a cloud computing news site on 1 October 2026. Only the relayed account was available at the time of writing.
• Share that cannot be cancelled: about 80 percent. Same report, same date. The wording is "non-cancellable or payable regardless of usage".
• Google: about $111.1 billion. Reported by Reuters, as relayed on 1 October 2026.
• Amazon: about $110 billion. Same report. Anthropic's own 20 April 2026 post says "more than $100 billion over the next ten years".
• Microsoft: about $31.4 billion. Same report. Anthropic's 18 November 2025 post says a $30 billion Azure commitment.
• Broadcom: about $161.2 billion. Same report, described as largely non-cancellable equipment lease obligations.
• Amazon capacity: up to 5 gigawatts of new capacity, nearly 1 gigawatt expected online by the end of 2026. Company statement, 20 April 2026.
• Google TPU capacity: up to one million TPUs, well over a gigawatt in 2026. Company statement, 23 October 2025.
• Microsoft and Nvidia capacity: up to 1 gigawatt. Company statement, 18 November 2025.
• Colossus 1 in Memphis: more than 300 megawatts and more than 220,000 Nvidia GPUs. Reported May 2026, no price disclosed.
• 2025 compute and infrastructure spend: $7.33 billion. Reported by the New York Post from the filing, via a financial tracking page updated 15 October 2026. The relayed Reuters account says "over $7 billion".
• Share of 2025 sales routed through Amazon and Google cloud platforms: 47 percent. Reported from the draft prospectus.
• Run rate revenue: above $30 billion. Company statement, 6 April 2026, up from about $9 billion at the end of 2025.

The headline number and where it comes from

The $518 billion figure traces back to reporting by Reuters on a confidential IPO prospectus, as relayed by a cloud computing news site on 1 October 2026. A separate financial tracking page, updated on 15 October 2026, dates the Reuters story to 28 September 2026 and says the company made a confidential filing on 1 June 2026. Only secondary accounts of the Reuters story were available at the time of writing, so treat this as a second-hand account of a draft document. Two details travel with the number:

• The total is described as "at least" $518 billion.
• Around 80 percent of the commitments are non-cancellable or require payment regardless of usage.

That second point is the important one. A company that can walk away from a contract carries a very different risk from a company that owes the money even if demand falls. A non-cancellable commitment behaves like debt: it is a fixed cost that sits on the books while revenue moves up and down.

A rough sense of scale helps, though this is simple division and not a reported schedule. If $518 billion were spread evenly over ten years it would be about $52 billion a year. The filing reports compute and infrastructure spend of $7.33 billion in 2025 (via the New York Post), so the commitments only make sense if spending rises steeply. The company's reported quarterly revenue of $16.8 billion in the third quarter of 2026 (Financial Times, 14 October 2026, as listed on the tracking page) is why the plan may be affordable, but those revenue numbers are also unaudited. For the detail, see Anthropic's reported revenue, losses and margins.

Terms explained in plain language

Compute deals use vocabulary that is easy to misread. Here are the words you will meet in this post.

• Gigawatt. A unit of electrical power, a thousand megawatts. When a lab says "a gigawatt of capacity" it means data centres that can draw about that much power at full load. It is a size measure, not a price and not a count of chips. As a rough rule quoted by TrendForce in a related OpenAI context, industry estimates put 10 gigawatts at around 4 to 5 million high end GPUs, so one gigawatt is on the order of several hundred thousand. That is an industry estimate, and it varies by chip generation.
• Non-cancellable. The buyer must keep paying even if it does not use the capacity. Lawyers also say take or pay, meaning you pay for what you agreed to take whether or not you take it.
• Lease obligation. A promise to pay rent on equipment or buildings over a set period. The Broadcom line in the filing is described this way.
• Purchase commitment. A promise to buy cloud services, such as compute time, up to a stated value over a stated period.
• Run rate. A recent period multiplied up to a year. It is a signal, not earned revenue.
• TPU. Tensor Processing Unit, Google's own AI chip.
• Trainium. Amazon's own AI chip family, made with its Annapurna Labs subsidiary.
• GPU. Graphics Processing Unit, the type of chip Nvidia sells and the most widely used for AI.
• Training and inference. Training builds a model. Inference runs the finished model to answer requests.
• Equity investment. Money a supplier puts into the company in return for shares. It is not spending by the company and should not be netted into commitment totals.

A timeline of the announcements

The sequence matters, because later deals often build on earlier ones and reports sometimes count them twice.

• 23 October 2025: Google TPUs. Anthropic says it will use up to one million Google Cloud TPUs, in an expansion "worth tens of billions of dollars", bringing well over a gigawatt online in 2026.
• 12 November 2025: United States build-out. Anthropic announces $50 billion of investment in American computing infrastructure with the data centre developer Fluidstack, with sites in Texas and New York, custom built for Anthropic and coming online throughout 2026. The company says this means about 800 permanent jobs and 2,400 construction jobs.
• 18 November 2025: Microsoft and Nvidia. Anthropic commits to buy $30 billion of Azure compute and to contract up to one gigawatt more on Nvidia Grace Blackwell and Vera Rubin systems. Nvidia says it may invest up to $10 billion in Anthropic and Microsoft up to $5 billion.
• 12 February 2026: Series G. Anthropic raises $30 billion at a $380 billion post-money valuation, led by GIC and Coatue, and reports run rate revenue of $14 billion.
• 6 April 2026: Google and Broadcom. Multiple gigawatts of next generation TPU capacity, coming online from 2027.
• 20 April 2026: Amazon. More than $100 billion over ten years and up to 5 gigawatts. Amazon invests $5 billion now and up to $20 billion more, on top of the $8 billion it invested earlier.
• May 2026: Colossus 1. Reported access to the Memphis data centre. A run rate above $47 billion is reported for the same month.
• 1 June 2026: confidential IPO filing. Per the financial tracking page.
• 26 August 2026: Nscale. TechCrunch, via the tracking page, reports a compute deal of about $45 billion over six years.
• 28 September and 1 October 2026: the reported prospectus figures. Reuters reporting, relayed by a cloud computing news site.

Put side by side, the pattern is clear: three big announcements in five months of late 2025, then a second round in April 2026 that was larger and longer, then the filing figures in autumn 2026.

The provider by provider picture

The same news site lists the commitments by provider, citing the Reuters report and Anthropic's own announcements. In plain terms:

• Google: about $111.1 billion. This sits on top of an October 2025 deal for up to one million TPUs and, in Anthropic's words, "well over a gigawatt of capacity" coming online in 2026. An April 2026 expansion with Google and Broadcom covers multiple gigawatts of next-generation TPUs starting in 2027. The news site says the April deal was later described as 5 gigawatts. Anthropic's own post does not state that figure, so treat it as reported.
• Amazon: about $110 billion. The April 2026 expansion is described as more than $100 billion over ten years, with up to 5 gigawatts of additional capacity and nearly 1 gigawatt expected by the end of 2026. The chips are Amazon's own Trainium2 and Trainium3, and Anthropic's post says the commitment covers Graviton and Trainium2 through Trainium4, with the option to buy later generations. Significant Trainium2 capacity was due in the second quarter of 2026 and scaled Trainium3 capacity later in 2026.
• Microsoft: about $31.4 billion. This lines up with a $30 billion Azure compute commitment agreed in November 2025, with up to 1 gigawatt of additional capacity on Nvidia Grace Blackwell and Vera Rubin systems. The report does not reconcile the small gap between the two figures.
• Broadcom: about $161.2 billion. The report describes these as largely non-cancellable equipment lease obligations. It does not say what is being leased or over what period.
• SpaceX, Colossus 1. A separate deal for the Memphis data centre, reported at more than 300 megawatts and more than 220,000 Nvidia GPUs.

Add the four large lines together and you reach roughly $414 billion. The remainder of the $518 billion has not been itemised publicly, and other reported deals (a roughly $45 billion arrangement with Nscale over six years, and a $50 billion United States build-out commitment announced in November 2025) do not map neatly onto any single line. Be careful about adding numbers from different sources: some are purchase commitments, some are lease obligations, some are build programs, and some are investments in the company rather than spending by it.

One coincidence is worth flagging without over reading it. The four large lines sum to about 80 percent of $518 billion, which is also the reported non-cancellable share. No published source says the two are connected, and the non-cancellable share is described as covering commitments in general.

What Anthropic says in its own words

Several company announcements are the firmest ground available.

In October 2025, Anthropic said it planned to use up to one million Google TPUs in an expansion "worth tens of billions of dollars," expected to bring well over a gigawatt online in 2026. The same page said the company serves more than 300,000 business customers and uses three chip platforms: Google TPUs, Amazon Trainium and Nvidia GPUs. It called Amazon its primary training partner and cloud provider. Google Cloud's chief executive Thomas Kurian was quoted saying Anthropic's choice reflected strong price-performance and efficiency.

In November 2025, the Microsoft and Nvidia announcement made Azure a third major cloud route and added Nvidia hardware at gigawatt scale. The same post said Amazon remains Anthropic's primary cloud provider and training partner.

In April 2026, Anthropic announced multiple gigawatts of next-generation TPU capacity with Google and Broadcom, expected to start coming online in 2027, with the vast majority sited in the United States. The page said run-rate revenue had passed $30 billion, up from about $9 billion at the end of 2025, and that the number of business customers spending over $1 million a year had passed 1,000, roughly doubling in under two months from over 500 at the February funding announcement. The company's chief financial officer, Krishna Rao, called it the most significant compute commitment to date.

Two weeks later the Amazon post added that Anthropic uses over one million Trainium2 chips to train and serve Claude, that over 100,000 customers run Claude on Amazon Bedrock, and that inference expansion in Asia and Europe is included.

The Colossus 1 lease: what is known and what is not

In May 2026 a deal was reported under which Anthropic gets access to the Colossus 1 data centre in Memphis, which the reports describe as more than 220,000 Nvidia GPUs and roughly 300 megawatts. The annual value is not disclosed by the parties. Two outside estimates differ: one projects $5 to $6 billion a year and one, from an analyst firm cited by Fortune, projects $3 to $4 billion. A monthly price, a contract length and a termination notice period all circulate in different reports, but none of them has been confirmed by the parties or by a primary document. If you see a precise monthly figure quoted as fact, ask where it comes from.

The deal is also unusual in kind. Most of the large commitments are with cloud providers who build and operate capacity for many customers. This one is access to a single existing site. Capacity that already exists can be used quickly, which is its appeal when new gigawatt sites take years to finish.

Why the chips matter as much as the dollars

The three chip families are not interchangeable. Google TPUs, Amazon Trainium and Nvidia GPUs each have different software stacks, different supply chains and different price curves. Anthropic's stated approach is to match workloads to the chips best suited to them, which also reduces dependence on any single supplier. According to the relayed Reuters report, Anthropic works with Annapurna Labs on future Trainium generations and with Nvidia on joint design, and Amazon's chief executive said Trainium at scale could cut annual capital spending by tens of billions of dollars.

Amazon describes its Project Rainier cluster, spread across several United States data centres including a site in St Joseph County, Indiana, as nearly half a million Trainium2 chips, used for both training and inference, with Claude expected to run on more than one million Trainium2 chips by the end of the year. Amazon says the cluster delivers more than five times the compute Anthropic used to train its previous models, and that the Trainium2 infrastructure is 70 percent larger than any other AI computing platform in AWS history. An UltraServer combines four Trainium2 servers of 16 chips each, so 64 chips per unit. Amazon does not state megawatts, acreage or site count for Rainier.

For more on the chip side, see how AI labs split work between custom chips and GPUs, the comparison of Trainium, Maia and Jalapeño and of Google's TPU 8t and 8i.

Who sells to whom: concentration on both sides

The draft prospectus, as relayed, lists dependence on large technology partners as a business risk. Two figures show why. About 47 percent of 2025 sales are reported to have passed through the Amazon and Google cloud platforms, and nearly a quarter of 2025 revenue is reported to have come from two customers. Both figures come from the draft prospectus via the tracking page and the relayed Reuters story.

This is a two way dependence. Anthropic needs these providers for capacity. The providers are also its distribution channel and, in Amazon's case, a large shareholder. The relayed report adds Microsoft figures for context: Microsoft Cloud revenue of $59.3 billion in its fiscal fourth quarter of 2026, up 27 percent, and commercial remaining performance obligations of $678 billion, with nearly 90 percent of full-year Microsoft Cloud revenue coming from customers outside frontier model companies. Remaining performance obligations are contracted revenue not yet recognised, so a large figure shows how much future business cloud providers have already signed.

How the money gets paid for

Anthropic's reported 2025 operating loss was above $8 billion and its net loss $42 billion, of which about $34 billion was a non-cash accounting charge (Reuters, via the tracking page). Cash and short term investments at the end of 2025 were reported at $20.28 billion. Commitments of this size must be funded from revenue growth, new equity, supplier investments and, where leases are involved, financing arranged by others. The guides to how AI data centers are financed and to what an AI data center costs per gigawatt explain the building blocks.

What is firm and what is uncertain

• Firm: the company announcements. The October 2025, November 2025, April 2026 Google and Broadcom, and April 2026 Amazon posts are Anthropic's own words. Their headline figures (one million TPUs, $30 billion Azure, more than $100 billion over ten years) are firm as statements of intent.
• Reasonably firm: the provider lines. They come from Reuters on a prospectus, but only relayed accounts were available at the time of writing.
• Uncertain: the unitemised remainder. About $104 billion of the $518 billion has not been broken out publicly.
• Uncertain: what Broadcom is leasing and for how long.
• Uncertain: Colossus 1 price, term and exit terms.
• Uncertain: the 5 gigawatt Google figure, which the news site reports but Anthropic's post does not state.
• Changing: everything about timing. Capacity "coming online" in 2026 and 2027 depends on construction, power and chip supply.
• Unaudited: all revenue, loss and cash figures, which come from investor materials and leaks.

What to watch next

• A public prospectus. The tracking page says no public S-1 was on the SEC filing system as of 15 October 2026. Bloomberg is reported to have said formal marketing could start the week of 9 November and trading before 26 November, but timing has slipped before. Audited notes on commitments would settle most open questions.
• Whether the non-cancellable share changes. Newer deals may be written differently from older ones.
• Capacity actually online. Nearly 1 gigawatt of Amazon capacity by the end of 2026 is a checkable milestone.
• Chip generations. Trainium3 scale-up, Google's next TPUs from 2027 and Nvidia Vera Rubin systems are the next steps.
• Power and permitting. The guides to how long an AI data center takes to build and who pays for data centre power show where delays tend to arise.

How to read these numbers without being misled

Five habits help:

• Separate the layers. Capacity (gigawatts), contracted spend (dollars over years), investment in the company, and cash actually paid are four different things.
• Check the tense. "Will come online in 2027" is a plan. "Came online" is a fact.
• Look for the source. A draft prospectus relayed by a news site is weaker than a filed one, and a blog quoting a blog is weaker still.
• Watch the denominator. A decade-long commitment sounds smaller per year, but only if revenue grows to carry it.
• Expect revisions. Several of these figures have already shifted between sources. For a general method, read how to read an AI lab's finances.

What this means for security and procurement teams

If your organisation builds on a frontier model, the vendor's fixed cost base is part of your supplier risk picture. Large non-cancellable commitments mean a vendor has strong reasons to keep selling capacity and weaker room to retreat if its market shrinks, which tends to favour continuity of service but can also push pricing and product changes. The same model is offered through three clouds (Amazon Bedrock, Google Vertex AI and Microsoft Foundry, per Anthropic), so your own contract and data location terms depend on which route you use. For a practical way to see what software and AI dependencies your own estate relies on, Safeguard offers a free scan at https://gold.safeguard.sh/scan.

Quick check

Sources

• Anthropic's $518 billion AI infrastructure plan spans major cloud providers, 1 October 2026.
• Expanding our use of Google Cloud TPUs and services, Anthropic, 23 October 2025.
• Anthropic invests $50 billion in American AI infrastructure, Anthropic, 12 November 2025.
• Microsoft, Nvidia and Anthropic announce strategic partnerships, Anthropic, 18 November 2025.
• Anthropic raises $30 billion in Series G funding, Anthropic, 12 February 2026.
• Anthropic expands Google and Broadcom compute deal, Anthropic, 6 April 2026.
• Anthropic and Amazon compute expansion, Anthropic, 20 April 2026 (updated 21 April).
• AWS Project Rainier, Amazon.
• xAI leases Colossus 1 supercomputer to Anthropic, secondary report.
• Is Anthropic profitable in 2026, Value Add VC, updated 15 October 2026, for the Nscale figure, filing dates and cash figures.
• OpenAI ramps up global compute power with Nvidia, AMD and Broadcom, TrendForce, 15 October 2025, for the GPUs per gigawatt rule of thumb.

Last reviewed: 19 October 2026.