Anthropic's $800B Compute Commitments: What the IPO Filing Shows and What It Does Not

Anthropic's IPO filing has produced a pile of headline numbers for its compute spending, and they do not agree. Depending on the outlet, Anthropic has committed $417 billion, about $518 billion, more than $600 billion, or more than $800 billion. The latest version to circulate, "over $800B in future compute commitments as it warns growth will slow," was spreading on October 31, 2026.

This post lines the figures up, says what each one counts, and marks what we could and could not verify. For the underlying filing, start with our Anthropic S-1 explainer and the SEC filing itself.

TL;DR: the numbers side by side

The pattern is clear: the bigger the number, the broader the definition, and the less the outlet shows its arithmetic.

Why the totals differ

A compute commitment is not one kind of obligation. The filing, as summarized by the outlets we could open, mixes at least four.

• Equipment leases. Coverage says the filing lists $161.2 billion in leases with Broadcom, the largest single arrangement. Baxtel separately describes an "Anthropic Direct" structure in which partners secure land, power and facilities while Anthropic leases the compute, with rent on one five-year lease totaling $125.2 billion before interest.
• Cloud contracts with a minimum spend. Google, Amazon and Microsoft cloud agreements typically state a floor, not a full forecast. GeekWire's reading, as relayed in search summaries, puts the direct Google commitment at $114.5 billion through July 2033, plus an estimated $125 billion to $150 billion of additional TPU capacity from 2028 to 2031. That second tranche is an estimate, not a stated total.
• Capacity deals with other AI companies. SpaceX's AI arm is reported to supply Nvidia GPU capacity. GeekWire and Baxtel report up to $120.5 billion through 2029; a Stocktwits item cites $84.5 billion. The two have not been reconciled.
• Deals not yet signed. The WSJ-based "more than $600 billion" figure includes talks, which a purchase-commitment note in a financial statement would not.

Add all four and you can reach $800 billion. Count only legally binding purchase obligations and you land near the low end.

What we checked, and what we could not

We have already told readers once that the filing copy we read leaves several numbers blank, including the price range, share count and commitment totals, and that we could not reproduce an aggregator's $265.6 billion "minimum cloud commitments" figure. That is in our S-1 explainer. The same discipline applies here.

Reported consistently by the outlets we could read (we could not open the filing ourselves): Broadcom is described as the largest single counterparty, at about $161 billion of leases in the Finimize and Markman Capital Insight summaries; SpaceX is a significant new supplier, with up to $120.5 billion through 2029 per Baxtel; the filing warns that revenue is mostly consumption-based and not under long-term contract while compute commitments run for years.

Not confirmed:

• The $800 billion total as a single line in the filing. Finimize's own list (Broadcom $161B, Google $150B, SpaceXAI $120.5B) sums to about $431.5 billion, and the article does not itemize the rest.
• Whether the Broadcom, Google, Amazon and SpaceX figures overlap with the $417 billion total. Baxtel says it does not state this.
• Whether a later S-1/A amendment changed any number. We found no coverage of an amendment, and the figures in circulation seem to come from the original filing and secondary analysis.

If you need a number for a model or a slide, quote it with its definition: "about $X billion of purchase obligations and leases per the S-1, as summarized by Y."

What else the filing coverage says

Finimize and other summaries report that revenue rose from $4.6 billion last year to more than $33 billion in the first nine months of 2026, which squares with the roughly $16.3 billion first-half figure in our own reading plus the $16.8 billion Q3 number covered in our Q3 revenue fact-check. The same Finimize piece says compute cost per revenue dollar fell from over $2.50 in early 2025 to under 50 cents by mid-2026, that the company reportedly posted a roughly $1 billion quarterly operating profit, and that nearly half of first-half customer sales came through three big-tech cloud marketplaces, some of which are developing rival models. Those are secondary summaries; verify them in the filing's MD&A before you repeat them.

The slowdown line is the most quoted. Cryptobriefing and Finimize both frame the filing as saying growth is cooling. The better reading is the filing's own risk language: Anthropic does not expect revenue to keep growing at comparable rates, which is standard IPO caution, but it matters when set next to multi-year fixed commitments.

How the number grew in two months

The headline total has moved with each report, which is itself a reason for caution. In early September, The Information reported about $517 billion in commitments over the 11 months through August, covering 14.8 gigawatts of capacity; we covered that in our 517B explainer. Around the October 26 to 27 filing, outlets quoted $417 billion, then roughly $518 billion, then more than $600 billion. By October 31 the figure circulating on X was $800 billion-plus.

Some of that growth is real, because new deals were signed between May and October, including the SpaceX capacity agreements that coverage dates to that window. Some is definitional. A reader who sees only the latest number cannot tell which, and that is why we ask every outlet for its parts list.

What this means if you choose a model provider

If you build on Claude, the useful questions are operational, not financial. Does the provider have enough capacity at your peak hours? Does it publish status and rate-limit guidance you can plan around? Is there a second provider you can fail over to? A very large commitment backlog is neither a green flag nor a red flag on its own. It tells you the lab expects demand to grow into its capacity, and it tells you the lab is exposed if demand does not.

A sensible habit is to keep your prompts and tool definitions portable across at least two model vendors, and to test the fallback path before you need it. That costs a few hours now and removes most of the supply-side risk the filing numbers hint at.

Why this matters even if you only build with Claude

You do not need to care about bond covenants to care about this story. Three practical links:

1. Capacity is the product. Rate limits, model availability and latency all trace back to how much compute a lab can bring online. The scale of these deals, which stretch to 2029, 2033 and in one report 2045, is the supply side of every capacity decision; see also our look at Anthropic weighing a new model to counter GPT-6 Astra.
2. Fixed costs make labs want volume. A lab with large take-or-pay commitments and consumption-based revenue has every reason to keep GPUs and TPUs busy. That favors aggressive pricing, batch discounts and generous free tiers, not price hikes, as long as demand holds. The filing coverage does not announce any price change, and we are not predicting one.
3. Counterparty risk is shared. If a neocloud supplier cannot borrow, capacity arrives late. Our neocloud shakeout and AI bubble fact-check posts cover the lender side, and the SpaceX GPU debt post covers one supplier.

Politics and systemic risk

The size of the numbers has reached Washington. A House Progressive Caucus resolution against AI bailouts cited the Anthropic prospectus, as we covered in our resolution explainer. The IMF's stability report flagged circular financing among AI suppliers and customers, covered in our IMF post. Neither claims Anthropic is in trouble; both show why a loose number like "$800B" gets repeated.

How to check the filing yourself

1. Open the S-1 on SEC EDGAR and search the document for "purchase commitments", "commitments and contingencies" and "leases".
2. Compare the table of future minimum payments by year with any outlet's total. Interest and executory costs are often excluded or included differently.
3. Separate signed from "under negotiation" language; only the first belongs in a commitment total.
4. Check the filing index for an S-1/A, which fills blanks and can change totals.
5. When an article gives a total but not its parts, add the parts it does name. If they fall short, as with Finimize's roughly $431.5 billion, ask where the rest came from.

What to watch next

• An S-1/A with the price range and finalized commitment tables.
• Whether SpaceX's $84.5 billion and $120.5 billion figures are reconciled.
• Any lab pricing or capacity announcement that responds to the filing.
• Reaction from rating agencies and lenders to the neocloud suppliers named in the filing.

Related reading

• Anthropic S-1 explained: revenue, losses, compute and the ANT IPO
• Anthropic's reported $517B compute commitments
• Anthropic Q3 revenue: claimed vs verified
• Has the AI bubble burst? Three rough weeks, fact-checked
• The neocloud shakeout
• IMF on AI financing and circularity

Figures and reports are as of October 31, 2026. Filing numbers can change in amendments; check EDGAR before relying on any total.