Anthropic S-1: Cloud and Compute Commitments

Anthropic’s compute commitments are known only from confidential draft S-1s reviewed by reporters, not from a public EDGAR filing. As of checks through October 24, 2026, no public Anthropic S-1 was on EDGAR; the company confidentially submitted a draft on June 1, 2026. Two draft vintages have been reported, and the numbers do not match. Treat both as media extracts that can still change before a public filing.

The mechanism is take-or-pay capacity reservation, not ordinary cloud usage. Anthropic locked multi-year minimums with the same companies that invest in it, resell Claude, and compete with it, so unused capacity still has to be paid. That converts what used to be a variable cost into something closer to fixed debt, while the suppliers retain the ability to ration how much capacity they will actually deliver.

Two drafts, two totals

Reuters, which reviewed the confidential prospectus in late September 2026, reported that Anthropic expects to spend at least $518 billion over about a decade with six partners, and that about 80% is non-cancellable or payable regardless of usage. [1] [2] A later draft reviewed by The Wall Street Journal and reported on October 20–21, 2026, puts about $450 billion under cloud agreements over the next seven to ten years (most non-cancellable and payable “regardless of actual usage”), plus a more than $160 billion Broadcom obligation to lease Google chips, for infrastructure costs above $600 billion. Deals still under negotiation could add up to $150 billion, and the company has signaled potential compute spending of up to $750 billion. [3] [4] [5]

The hyperscaler minimums are stable across drafts. What moved is the Nvidia-capacity counterparty (xAI in the September account; Nscale and Lambda in the October account) and how the outer total is framed.

Counterparty	Amount (as reported)	Period	Minimum / cancellation	Silicon (as stated in the draft)	Source vintage
Google (cloud / infrastructure services)	At least $111.1 billion	April 2026 – July 2033	Payable “regardless of usage.” “If our actual spend falls short, we must pay Google the difference.”	Cloud services; TPU capacity is the separate Broadcom lease	Reuters, Sept. 29, 2026; WSJ draft rounds to ~$111 billion
Amazon / AWS	$110 billion	May 2026 – April 2036	Similar shortfall terms: pay the difference if spend falls short	Draft describes infrastructure/cloud service obligations. Trainium is not quoted from the S-1 text in the reporting; it comes from the April 2026 commercial announcement (up to 5 GW of Trainium generations)	Reuters; WSJ
Microsoft / Azure	$31.4 billion	November 2026 – May 2033	“Non-cancelable except in the event of Microsoft’s uncured material breach.” Payable regardless of actual usage	Draft describes cloud compute. Nov. 2025 announcement (not the S-1 extract) tied a $30 billion Azure purchase to Nvidia Grace Blackwell and Vera Rubin	Reuters; WSJ rounds to ~$31 billion
Broadcom (equipment leases)	About $161.2 billion	Largely locked; non-cancellable by either party except on default	Equipment lease obligations	Google-designed chips / TPUs. WSJ: “more than $160 billion” to lease Google chips via Broadcom	Reuters; WSJ
Broadcom (five-year TPU lease, disclosed in the same filing)	$125.2 billion	Five-year lease	Lease commitment; up to $42 billion convertible-note facility from Broadcom could finance about one-third	Explicitly “tensor processing unit (TPU) computing capacity”; next-gen TPU capacity from 2027	Reuters, Oct. 1, 2026
xAI (SpaceX-related)	Up to $84.5 billion	Through 2029	Largely cancellable on 90 days’ notice	Nvidia-based computing capacity	Reuters, Sept. 2026
Nscale and Lambda	Up to $86 billion	Through 2033 and 2034 respectively	Included in the later cloud-agreement stack	Rent Nvidia chips	WSJ draft, Oct. 2026
AMD	Capacity expected to exceed $20 billion; AMD to buy up to $5 billion of Anthropic stock	Not stated in the extracts	Equity-linked supply, not described as take-or-pay in the same language	Not named as a specific SKU in the Reuters extract. A separate report of the later draft says up to 2 GW of MI450, deliveries from 2027 — that chip detail is not in the Reuters prospectus quotes	Reuters; later secondary report

Sources for the table: Reuters via Business Day and InvestmentNews for the September extract; Reuters Oct. 1 for the TPU lease and $42 billion facility; WSJ via VKTR and Motley Fool for the October extract. [6] [2] [7] [3] [4] The $125.2 billion five-year TPU lease and the $161.2 billion Broadcom equipment-lease figure appear in the same filing but are not reconciled in the published extracts. Do not add them. They may overlap.

Google + Amazon + Microsoft alone are about $252.5 billion of minimum infrastructure-service spend over seven to ten years, which Reuters summarized as roughly $250 billion of required minimum spending. [1]

Contractual and balance-sheet treatment

The draft does not, in the published extracts, walk through ASC 842 lease classification line by line. What it does state is the legal character of the obligations and two balance-sheet snapshots:

• Year-end 2025: $54.6 billion of non-cancellable hosting and computing commitments. [8]
• Early 2026: total long-term commitments exceeded $417 billion, covering 3.5 gigawatts of dedicated computing capacity. [8]
• Google and Amazon are true minimum-purchase / shortfall contracts: unused capacity is still owed. That is take-or-pay, not a usage cap. [6]
• Microsoft is non-cancellable except for Microsoft’s own uncured material breach — Anthropic cannot walk away because demand disappoints. [6]
• Broadcom leases are non-cancellable by either party except on default. Certain payment or performance defaults can accelerate a substantial portion of the lease obligations and simultaneously restrict use of the up-to-$42 billion financing facility that was meant to help pay them. [7]
• Anthropic deposited cash into a restricted account for Broadcom’s benefit in April 2026 and may have to add more. That cash is already constrained on the balance sheet even if the underlying lease liability is not fully recognized yet. [7]
• The $42 billion Broadcom facility is a convertible note that could finance about one-third of the $125.2 billion five-year TPU lease. Anthropic said it does not expect any notes to be sold before the IPO. Conversion would turn a supplier into a shareholder on the same capacity it is financing. [7]

Standard GAAP implication, not a quote from the filing: non-cancellable purchase commitments and leases that have not yet commenced are usually disclosed in the commitments footnote and recognized as liabilities only when the recognition criteria are met. The $54.6 billion year-end 2025 figure is the recognized-disclosure snapshot; the $417 billion / $518 billion / $600 billion figures are forward contractual exposure, most of which had not yet hit the liability side of the balance sheet. Cash ended 2025 at $20.28 billion against an operating loss above $8 billion and compute-and-infrastructure spend of $7.33 billion that year. [2] The mismatch is the point of the risk factors: revenue is not contracted for the life of the compute bills.

Separate from the S-1 extracts, structured financings (SPVs buying TPUs and leasing them back, with Broadcom residual-value support) have been reported as a way to keep hardware off the corporate balance sheet. Those structures are not established by the prospectus quotes above; do not treat them as S-1 accounting policy.

Chip dependence: what the draft actually names

The draft is explicit on two silicon families and silent, in the published extracts, on a third.

• TPU. The filing names a $125.2 billion five-year lease of tensor processing unit computing capacity, with Google and Broadcom as the design/supply pair, and next-generation TPU capacity beginning in 2027. The WSJ draft describes the Broadcom piece as leasing Google chips. This is the cleanest chip-level disclosure in the prospectus. [7] [4]
• Nvidia GPU. The September draft discloses up to $84.5 billion of Nvidia-based capacity from xAI through 2029, mostly cancellable on 90 days’ notice — the only large commitment with a real exit. The October draft instead (or additionally) shows up to $86 billion to rent Nvidia chips from Nscale and Lambda through 2033 and 2034. [2] [3] Whether the xAI figure was replaced, reclassified, or sits alongside the neocloud deals is not resolved in the public reporting.
• Trainium. Not quoted from the S-1 text in Reuters or WSJ extracts. The $110 billion Amazon line is described as long-term infrastructure service obligations, payable regardless of usage. Dependence on Trainium2/3/4 and “up to 5 GW” is from Anthropic and Amazon’s April 2026 commercial announcements, not from a sentence reporters pulled out of the prospectus. Infer that a large share of the AWS minimum is custom Amazon silicon, but do not attribute a Trainium SKU schedule to the S-1 until the public filing says so.
• AMD. Capacity expected to exceed $20 billion, plus an equity commitment of up to $5 billion. A specific MI450 / 2 GW claim appears in secondary coverage of later materials, not in the Reuters prospectus quotes. [2]

The commercial logic Anthropic has stated outside the S-1 — match workloads to Trainium, TPU, or Nvidia GPU — is the operating reason these contracts exist side by side. The S-1’s version of that point is blunter: future demand for advanced AI “will be limited principally by the availability of compute,” so the company signed the minimums to reserve scarce supply. [1]

Related-party and circular disclosures

Amazon, Google, and Microsoft each occupy several seats at once: investor, cloud supplier, distributor, customer, and competitor. The filing says those overlapping roles create incentives that “may not be fully aligned” with Anthropic’s, and that reliance on a limited number of partners “creates complex dynamics that could give rise to conflicts of interest and adversely affect our access to compute.” [8] [2]

Reported commercial entanglement from the same prospectus:

• 47% of 2025 revenue — about $2.16 billion of nearly $4.6 billion — was routed through Amazon and Google cloud marketplaces. Distribution fees were about $351 million, roughly 16 cents per dollar of those sales, booked in sales, marketing, and partnerships expense. The share sold through those two rose from 11% in 2023 to 32% in 2024 to nearly half in 2025. [8]
• Those platforms collected 60% of the $909 million in customer bills outstanding at the end of 2025, up from 42% in 2024. Anthropic contracts with the end customer, but cash collection runs through the same firms that also decide how much compute it gets. [8]
• Cloud partners gain visibility into Anthropic’s pricing and commercial terms, which the filing says could influence compute allocation and how aggressively they sell Claude. [8]
• Broadcom is both the hardware supplier on the TPU lease and the potential lender of up to $42 billion against it. The prospectus flags that dual role as creating “potential conflicts of interest” that might affect access to compute, and warns that Broadcom’s pricing and hardware decisions could limit how much infrastructure Anthropic can procure. [7]
• AMD is both a capacity supplier (expected above $20 billion) and a committed buyer of up to $5 billion of Anthropic stock. [2]

Equity stakes themselves (Google reported around 14% and contractually capped, Amazon held largely in convertibles and non-voting preferred) come from court filings and Amazon’s own 10-Q, not from a cap-table table in the published S-1 extracts. The prospectus does state that Amazon and Google are large investors while also being the critical suppliers.

Disclosed risks tied to the commitments

These are the risk statements reporters extracted from the draft, not a paraphrase of generic AI-IPO boilerplate:

1. Shortfall payments. If actual spend falls short of the Google commitment, Anthropic must pay the difference. Similar terms apply to Amazon. The Microsoft contract is non-cancellable except for Microsoft’s uncured material breach. About 80% of the $518 billion stack is non-cancellable or payable regardless of usage. [6] [1]
2. Compute can still be cut even after the minimums are signed. “If the compute we have access to from third parties is curtailed, repriced, or terminated … our business, financial condition, and results of operations could be adversely affected.” [2] Paying for capacity and receiving it are different risks.
3. Partners are rationing. The later draft says Anthropic sought more capacity than Amazon, Google, and Microsoft were “willing to provide,” and that those partners increasingly view it as a competitor. That is why it is leasing its own data centers and hiring third-party operators, some with “limited operating history,” whose sites may be less reliable than the hyperscalers’. [3] [9]
4. Conflicts and information leakage. Multi-role partners have incentives that “may not be fully aligned.” Visibility into Anthropic’s pricing can shape both compute allocation and how hard the platforms push Claude versus their own models. [8] [2]
5. Revenue is cancellable; compute is not. Two unnamed customers each generated 12% of 2025 revenue. Many of the largest customers have no long-term contracts and can reduce or stop spending. The October draft says about one-fifth of first-half 2026 revenue came from two customers, and a “significant portion” is expected to come from a limited number of customers for the foreseeable future. [8] [3]
6. Collection-pipe risk. Disputes or delays in the Amazon/Google billing pipeline can hurt cash flow even though Anthropic contracts directly with the customer. [8]
7. Broadcom acceleration. Payment or performance defaults can make a substantial portion of lease obligations immediately due while cutting off the $42 billion facility intended to fund them. Restricted cash already sits for Broadcom’s benefit. [7]
8. Only one large Nvidia path has an exit. The xAI agreements, up to $84.5 billion through 2029, are largely cancellable on 90 days’ notice. Everything material with Google, Amazon, Microsoft, and Broadcom is not. [2]
9. Scale versus cash. End-2025 cash was $20.28 billion. First-half 2026 compute and infrastructure spend was nearly $9 billion, more than all of 2025’s $7.33 billion. The contractual stack is an order of magnitude larger than either figure. [2] [3]

What this means if you are on the other side of the table

A competitor or a new model lab cannot “add a cloud account” and match this position. The scarce asset is not an API key; it is multi-year, gigawatt-scale reservations on three incompatible stacks (Google TPU via Broadcom, AWS capacity that commercially is Trainium-heavy, and Nvidia GPUs via Azure and neoclouds), mostly take-or-pay, mostly with parties that also own a piece of the buyer. The flexible slice — 90-day cancellable Nvidia capacity via xAI — is the exception that shows how unusual the rest of the book is.

For a supplier, Anthropic’s draft is contracted demand, not a forecast: Google, Amazon, and Microsoft have minimums they collect even if Claude usage misses, and Broadcom has non-cancellable equipment leases plus an option to convert financing into equity. The offsetting risk, which the filing itself flags, is that those same suppliers are competitors who have already declined to sell Anthropic all the capacity it asked for. The contract guarantees payment. It does not guarantee that the next increment of power and chips goes to Claude rather than to Gemini, an Amazon model, or Copilot.

For anyone underwriting the IPO, the load-bearing question is not the $80.7 billion first-half 2026 net loss — about $72 billion of that was non-cash accounting on convertible instruments. [3] It is whether revenue, which customers can cut, grows fast enough to absorb minimums that Anthropic cannot cut, while the landlords of that compute remain free to ration the next tranche because they also sell a rival model.

---

Recent Findings Supplement (October 2026)

Anthropic’s draft IPO prospectus (S-1), reviewed by The Wall Street Journal in mid-to-late October 2026, details over $600 billion in long-term AI compute and infrastructure commitments (with potential to reach ~$750 billion including negotiations), the vast majority non-cancelable or payable regardless of actual usage.[1][2]

These stem from expanded April 2026 partnerships and reflect a multi-vendor strategy across AWS Trainium, Google TPUs, Nvidia GPUs (via third parties), and Broadcom-leased Google chips. The filing highlights heavy reliance on a small number of hyperscaler partners—who are also investors—creating disclosed conflicts, rationing risks, and fixed-cost exposure that could pressure profitability despite rapid revenue growth ($16.3 billion in H1 2026).[1][3]

This represents the most recent and detailed public insight into Anthropic’s obligations (as of October 26, 2026); a full public S-1 had not yet been filed.

Cloud and Infrastructure Commitments Breakdown

Anthropic’s draft filing quantifies massive take-or-pay style obligations spanning 7–10 years, with specific dollar amounts, counterparties, and chip dependencies reported from the document. These convert variable compute costs into near-fixed liabilities, recorded as long-term contractual commitments (with minimum purchase obligations implied by “payable regardless of usage” language).[4]

Table of Key Commitments (from draft S-1 as reported):

• Google Cloud: ~$111–111.1 billion (April 2026–July 2033, ~7 years); TPU capacity; payable regardless of usage.[3]
• AWS: ~$110 billion (May 2026–April 2036, ~10 years); up to 5 GW Trainium2/3 (and future) capacity; payable regardless of usage.[4]
• Microsoft Azure: ~$31–31.4 billion (November 2026–May 2033, ~7 years); GPU/cloud capacity; payable regardless of usage.[3]
• Broadcom (Google chip leases/TPUs): >$160–161.2 billion; equipment leases tied to Google TPUs.[5]
• Nscale + Lambda (Nvidia GPUs): Up to ~$84.5–86 billion (through 2033/2034); third-party Nvidia-powered capacity.[2]
• Other/Additional: AMD expected >$20 billion (MI450 and capacity); xAI up to $84.5 billion (Nvidia capacity, largely cancelable on 90 days’ notice); further negotiated deals could add ~$150 billion.[6]

Total disclosed: ~$450 billion in cloud agreements over 7–10 years (bulk non-cancelable); overall infrastructure >$600 billion; potential spend up to $750 billion.[1]

Balance-sheet/contractual treatment: These appear as long-term non-cancelable contractual obligations or commitments (take-or-pay minimums), payable even if unused. They function as fixed costs rather than variable expenses, with related-party elements tied to equity investments from the same hyperscalers.[1]

Chip Dependence and Multi-Vendor Strategy

The filing underscores dependence on specific accelerators for training and inference of Claude models, with diversification across incompatible stacks to mitigate single-vendor risk:
- AWS Trainium (Trainium2/3): Primary/anchor for large-scale training and inference (Project Rainier cluster); native integration and cost advantages claimed.
- Google TPU (various generations, including Trillium/Ironwood): Significant capacity for compatible workloads; accessed directly and via Broadcom leases.
- Nvidia GPUs: Rented via third parties (Nscale, Lambda, others) for CUDA-dependent or flexible workloads.
- Broadcom custom/Google chips: Leased for additional TPU-scale capacity.

This approach allows workload matching (e.g., TPU for certain JAX/XLA models, Trainium for Neuron-optimized) but creates complexity in software portability and exposes the company to supplier-specific shortages or pricing.[7]

Related-Party Disclosures and Conflicts

Hyperscalers are both major suppliers and investors:
- Amazon has invested ~$13 billion+ (with potential for more tied to milestones) and supplies Trainium/AWS capacity.
- Google has committed up to $40 billion in investment (phased) alongside TPU/cloud deals.
- Microsoft has made investments while providing Azure capacity.

The draft explicitly flags risks from these dynamics, including potential conflicts of interest where partners may prioritize their own models (e.g., Gemini, competing services) or ration capacity to Anthropic as a perceived competitor.[8]

Disclosed Risks Tied to Commitments

The prospectus highlights several compute-related risk factors (reported from the draft):
- Supply rationing and insufficiency: Cloud partners have declined to provide all requested capacity; Anthropic is pursuing third-party data centers and own builds to compensate.[9]
- Fixed-cost burden and profitability delay: Hundreds of billions in non-cancelable obligations could outpace revenue growth, delaying breakeven or requiring ongoing capital raises.
- Supplier concentration and conflicts: Heavy reliance on few partners (who are also investors/competitors) creates access, pricing, and strategic risks.
- Chip and tech dependence: Dependence on specific accelerators (Trainium, TPU, GPU) exposes the company to hardware shortages, performance shortfalls, or migration costs.
- Customer concentration: A small number of customers drive a significant revenue portion, amplifying overall business risk alongside compute costs.
- Additional: Potential for AI model risks (e.g., unintended behaviors) and broader infrastructure scaling challenges.[10]

For competitors or new entrants: These commitments lock in capacity at hyperscaler scale but create a high bar—replicating Anthropic’s multi-vendor leverage requires similar capital access and relationships. The disclosed rationing and conflicts signal that even well-funded players face supply constraints, favoring those with proprietary chips, own data centers, or more flexible/cancelable terms. The take-or-pay structure turns compute into a leveraged bet on sustained demand growth.[9]

All figures and details derive from October 2026 reporting on the confidential draft; actual public S-1 language may vary slightly upon filing. No earlier post-April 2026 updates provide comparable granularity.