Anthropic Has Promised to Spend $142 Million a Day for Ten Years

The number is in its own draft prospectus. The question for investors is whether it is in theirs.

Richard G. Prouse, Principal Analyst, Confidential Risk Consulting (New Zealand)  ·  7 October 2026  ·  7 minute read
The briefing in 50 seconds.

Key points

  • Anthropic's draft IPO prospectus reportedly commits it to at least $518 billion of compute spending over ten years. About 80% is firm.
  • Revenue is growing fast enough to make that plausible. The costs are fixed all the same, and the revenue is not.
  • Funds that hold Amazon, Alphabet, Microsoft and Broadcom already own the other side of these contracts.
  • Nine facts an investor needs are not yet in the public record. Five questions test whether your own file would stand up.

In 2025, Anthropic took in about $12.6 million a day.

It has contracted to spend an average of $142 million a day on computing power. Every day. For the next decade.

That is not a critic's estimate. It is arithmetic on the company's own draft IPO prospectus, as reviewed by Reuters[1]: at least $518 billion of compute commitments over ten years. About 80% of it cannot be cancelled, or is payable whether the capacity is used or not. That firm portion alone averages about $113 million a day.

Column chart of Anthropic's average daily revenue in 2025, Q1 2026 and Q2 2026 against an average daily compute commitment of $142 million
Figure 1. Daily revenue in the second quarter of 2026 was ten times the 2025 average. The commitment line is the ten-year average of all reported commitments, including the cancellable portion; the actual payment schedule has not been published.

The maker of Claude is reportedly seeking a valuation above $2 trillion, in a listing expected in November[3].

This article does not argue that Anthropic will fail. It argues something narrower and less comfortable. Once a liability is disclosed, it stops being the company's secret and becomes the buyer's responsibility.

Where the money is owed

Six counterparties account for the $518 billion. Four of them hold contracts that cannot be walked away from.

Bar chart of Anthropic's reported compute commitments by counterparty: Broadcom-related leases $161.2 billion, Google at least $111.1 billion, Amazon $110 billion, xAI up to $84.5 billion, Microsoft $31.4 billion, AMD more than $20 billion
Figure 2. Only the xAI agreement is reported as largely cancellable, on 90 days' notice.

Set against that, the 2025 accounts are small. Revenue was $4.59 billion[9]. The operating loss was $8.06 billion, and compute alone cost $7.33 billion, about three times the year before[2]. The net loss of about $42 billion includes a $34 billion non-cash charge.

The commitments now signed are 113 times that year's revenue.

Three places the cost does not yet show

1. The bill arrives after the asset does

Accounting rules recognise infrastructure cost late, and legitimately so. Under US GAAP, an asset under construction sits on the balance sheet as construction in progress and is not depreciated until it is ready for use (ASC 360). A lease creates no balance-sheet liability until it commences (ASC 842). Until then it appears only in the notes.

Anthropic's draft reportedly lists about $161.2 billion of Broadcom-related equipment lease obligations. Some of the hardware is held outside the company altogether. According to the Financial Times, as summarised by The Decoder[6], a special-purpose vehicle funded mainly by Apollo and Blackstone bought about one gigawatt of Google chips for $35 billion and leases them to Anthropic. Broadcom backstops about $30 billion if lease payments stop.

The consequence is simple. A trailing income statement describes a smaller machine than the one already ordered. How much smaller cannot be verified until the prospectus is public. Two notes will answer it: leases not yet commenced, and construction in progress.

2. The investors, the suppliers and the sales channel are the same firms

The draft is direct about this. Amazon, Google and Microsoft each act as investor, customer, cloud provider, distributor and competitor[1].

Column chart showing the share of Anthropic's sales routed through Amazon and Google rising from 11% in 2023 to 32% in 2024 and 47% in 2025
Figure 3. Nearly half of 2025 sales reached customers through two companies that are also investors and suppliers[5].

There is a second-order point for portfolio managers. Anthropic's obligations are those companies' contracted revenue. A fund that already holds Amazon, Alphabet, Microsoft and Broadcom, and then buys Anthropic, has bought the same risk twice.

3. The costs are fixed. The revenue is not.

On the Google contract the filing says:

"If our actual spend falls short, we must pay Google the difference."[1]

On the other side of the ledger, nearly a quarter of 2025 revenue came from two customers[4]. Many of the largest hold no long-term contracts and may cut or stop spending.

A decade of payments that cannot be cancelled is being funded by income that can.

The case for the other side

It is a serious case, and any honest account has to state it.

Revenue was $4.73 billion in the first quarter of 2026 and $11.5 billion in the second[4]. The company was on track for a second straight quarter of adjusted operating profit. Annualised revenue reached $65 billion by July[4].

Anthropic's own rationale is that AI development will be "limited principally by the availability of compute"[5]. On that view the commitments are not a burden. They are the moat.

Both things can be true. The growth is real and the obligation is fixed. In the first quarter, compute still cost 71 cents of every dollar of revenue[4]. A $2 trillion price is a wager that revenue keeps compounding faster than the payments fall due, for ten years, from customers who are free to leave.

What a risk model says about your file

In January 2026 I published a working paper on SSRN, Bridging the Documentation Void[8]. It was written about security controls. Its argument is that a control with no evidentiary trail is "effectively legally non-existent".

Lr = (T × V) − (Ce × Cd)

Residual liability (Lr) is threat times vulnerability, less the effectiveness of the control (Ce) multiplied by its Documentation Coefficient (Cd), a score from 0.0 to 1.0. If that coefficient is near zero, the control counts for nothing, however good it was in practice.

Now apply it to an investment decision. This application is mine and is made here for the first time; it is not in the paper. The control is your due diligence. The documentation is what your file shows you knew on the day you bought.

Nine documentation voids in the public record

Start with what cannot yet be checked. The prospectus is still confidential, so everything known comes from press reporting on a draft. On that record, nine facts an investor needs in order to price these commitments are missing or incomplete.

What an investor needsStatus in public reportingWhere to look in the prospectus
1Year-by-year payment schedule for the $518 billionOnly totals and end dates reportedContractual obligations table
2Leases signed but not yet commencedNot reportedLeases note
3Construction in progressNot reportedProperty and equipment note
4Accounting treatment of the chip-leasing vehicleReported from the financing side onlyLeases and variable-interest notes
5The two customers behind nearly a quarter of 2025 revenueUnnamedConcentration note
6Length of the largest customer contractsDescribed only as often not long-termRisk factors
7Cancellation terms of the AMD commitmentNot reportedCommitments note
8Size of the raise and use of proceedsPress estimates onlyUse of proceeds
9Lock-up terms and expiry dateNot publicUnderwriting section
Exhibit 4. The void register, as at 7 October 2026. Status reflects the press coverage cited in the sources.

None of this is an allegation of concealment. A confidential draft is not required to be public, and the final prospectus may close every one of these gaps. Until it does, each is a number that a buyer is supplying from assumption. When the filing appears, this register is the checklist: any item still open at that point is a void in the company's own documentation.

Five questions for your own file

The second set of voids is the investor's. Before 28 September, a buyer could say the commitments were unknown. That defence is gone. Five questions decide what replaces it.

  1. Is the commitment schedule in your investment memo?The $518 billion, the share that is firm, the counterparties.
  2. Does your margin model use contracted capacity?Not the cost base of last year.
  3. Have you counted the exposure you already hold?Amazon, Alphabet, Microsoft and Broadcom.
  4. Is a maximum acceptable loss and an exit rule recorded?Including the lock-up expiry date, once it is known.
  5. Was all of it written before the purchase?Not after the fall.

Each "no" is a documentation void. The paper's term for a control that exists in practice but not on paper is an Unchecked PFI Gap.

There is one date to add. Insiders in an IPO are usually barred from selling for a period after listing. A study of 1,948 lock-up agreements found a three-day abnormal return of minus 1.5% at expiry and a permanent 40% rise in trading volume, with larger effects for venture-backed firms[7]. Anthropic's lock-up terms are not yet public. When they are, the expiry date belongs in the file.

The bottom line

Anthropic's growth is real, and so is the bill. My assessment is that the commitment risk is material, disclosed in outline, and not yet tested by the market: about $414 billion of fixed obligations rests on revenue that customers can withdraw, and nine facts needed to price it are not yet in the public record.

That does not make the listing a bad investment. It makes an undocumented one indefensible. If you cannot close the nine voids from the final prospectus, and answer the five questions in writing before you buy, the largest risk in this IPO is in your own file.

Richard G. Prouse, Principal Analyst, Confidential Risk Consulting

Know your exposure before the market tests it

Confidential Risk Consulting measures the risk that has not been written down, in investment decisions, operations and security. If you need to know what you are exposed to, and whether your position would stand up to scrutiny, we will quantify it.

The five questions above are available as a one-page worksheet for investment committees. Email richard@confidentialriskconsulting.com with the subject line "Documentation Test".

About the author. Richard G. Prouse is Principal Analyst at Confidential Risk Consulting (New Zealand) and the author of the Prouse Residual Liability Model, published as an SSRN working paper (DOI 10.2139/ssrn.6175099).

Notice. This article is general information and commentary. It is not financial advice or a recommendation to buy or sell any financial product. Figures are taken from press reporting on a confidential draft prospectus, as available on 7 October 2026, and may change in the public filing. Disclosure. The author holds no position, long or short, in any company named.

Sources

  1. InvestmentNews, "Anthropic's landmark IPO filing shows 12-fold revenue jump, $518B compute bill", citing Reuters. https://investmentnews.com/equities/anthropics-landmark-ipo-filing-shows-12-fold-revenue-jump-518b-compute-bill/268391
  2. The Next Web, "Anthropic lost $42bn in 2025 as revenue grew 12-fold, Reuters reports". https://thenextweb.com/news/anthropic-ipo-prospectus-42bn-loss-518bn-compute
  3. SiliconANGLE, "Leaked Anthropic IPO filing reveals $8B operating loss, rapid revenue growth", 29 September 2026. https://siliconangle.com/2026/09/29/leaked-anthropic-ipo-filing-reveals-8b-operating-loss-rapid-revenue-growth/
  4. PYMNTS, "Anthropic's IPO Filing Puts a $518 Billion Price Tag on AI Ambition", 29 September 2026. https://www.pymnts.com/news/artificial-intelligence/2026/anthropic-prospectus-shows-what-2-trillion-dollar-ai-company-costs-run/
  5. TrendForce, "Anthropic Eyes $2T IPO Valuation, $518B AI Buildout Spans Amazon, Google, Broadcom, and More", 30 September 2026, citing Reuters. https://www.trendforce.com/news/?p=64682
  6. The Decoder, "Google moves billions in Anthropic chip risk off its balance sheet", citing the Financial Times. https://the-decoder.com/google-moves-billions-in-anthropic-chip-risk-off-its-balance-sheet/
  7. Field, L. C. and Hanka, G. (2001), "The Expiration of IPO Share Lockups", Journal of Finance 56(2), 471-500. https://ideas.repec.org/a/bla/jfinan/v56y2001i2p471-500.html
  8. Prouse, R. G. (2026), "Bridging the Documentation Void: A Quantitative Approach to Residual Liability (Lr) and Fiduciary Defensibility", SSRN working paper, DOI 10.2139/ssrn.6175099. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6175099
  9. Sovereign Magazine, "Anthropic Commits to $518 Billion in Computing Power Ahead of Its IPO". https://www.sovereignmagazine.com/article/anthropic-ipo-518-billion-compute-commitments