Anthropic's S-1 shows compute per revenue dollar falling fivefold, which makes the $800B commitment the open question
The first public frontier-lab accounts. Hyperscalers, chip vendors and lenders hold the other side of the compute bill.
Competitive · S-1 filed, compute cost per dollar under 50 cents
Written for: Investor or analyst
Next step
Thesis: lab inference is a high-margin business whose risk now sits in the commitment schedule, not in unit cost. Confirming signal: the first quarterly report after listing shows compute per revenue dollar still falling while API list prices keep dropping. Already building: Broadcom's financing SPV and its bank syndicate, plus FT, Stratechery and Latent Space doing filing-level analysis.
Why now
Our recent card asked whether the price war would break lab margins under a fixed compute bill. The S-1 answers the operating half: inference margins are wide and still widening. The open question is now the commitment schedule. Stratechery argues competition may eat the profit and that investors face endless dilution. GPT-6.1 Terra halving API prices the same week tests that argument at once. Broadcom had to accept higher rates and give loan guarantees to get banks into the $63B chip SPV, so lenders are pricing that risk too.
What became possible
Audited lab unit economics are now public. Per the FT's read of the S-1, Anthropic spent about $2.50 on compute and infrastructure per $1 of revenue in Q1 2025 and under 50 cents in Q2 2026. Quarterly revenue is $11.6B, the first operating profit has arrived (Latent Space puts it at $1B in Q2), and existing customers spend nine times what they did a year ago. Against that sit roughly $800B in long-term compute commitments, and CNBC reports $80B in first-half losses.
Two ways in
- Commitment map
A data product that ties each lab's disclosed take-or-pay compute commitments to the vendors, SPVs and lenders on the other side, updated with every filing and financing.
- Customer
- Credit desks and public-equity analysts covering AVGO, cloud providers and AI-exposed banks
- Model
- Annual data subscription per seat
- Lab cost-curve tracker
Normalises compute cost per revenue dollar across Anthropic's S-1 and future lab filings, alongside list-price cuts like Terra's, to show whether cost falls faster than price.
- Customer
- Crossover funds preparing for the next lab IPO
- Model
- Research subscription plus custom models
Competition
The risks are that SemiAnalysis, Bloomberg Intelligence or a credit-data incumbent adds this as a feature, and that the market for it is small. A narrow, fast-moving data feed sold to credit and equity desks looks like the realistic opening.
Full competition notes (9 competitors)
The cost-curve tracker has less room. SemiAnalysis already sells lab unit economics and token-price modelling to institutions, Epoch gives away compute-spend data, and once the S-1 is public anyone can compute ratios like Anthropic's revenue per compute dollar (about 0.4x to 2.1x). Kepler and several Substacks did exactly that within days. A newcomer would need a methodology that is more rigorous and auditable than those. One example would be a consistent split of training from inference across filings; analysts already note the S-1 does not give one. Another would be lining up those costs against list-price cuts. With only one public lab filing, plus xAI inside the SpaceX S-1, there is not yet much of a time series to sell. The commitment map is the more open idea. What exists is fragmented. Bloomberg's circular-deals graphic is a one-off, the Wikipedia table is crowd-edited, the neocloud funding trackers cover only one side, and the general deal databases do not link a lab's take-or-pay line items to the specific vendor, SPV and lender on the other side. Even the headline number is unsettled. Secondary sources put Anthropic's commitments anywhere from $417B to $750B, the SEC copy leaves the aggregate blank, and the split between cancellable and non-cancellable amounts differs by counterparty: about 80% non-cancellable overall, but xAI capacity can be cancelled on 90 days' notice. A reconciled, sourced map filed on every 10-Q, 8-K, bond deal and SPV financing would be new. Examples to cover include the Broadcom equipment leases, Blackstone's loan package tied to Google chips, and Nvidia's guarantee for the Ohio campus. Credit desks covering AVGO, ORCL, CRWV and the lending banks would find it useful. The risks are that SemiAnalysis, Bloomberg Intelligence or a credit-data incumbent adds this as a feature, and that the market for it is small. A narrow, fast-moving data feed sold to credit and equity desks looks like the realistic opening.
- SemiAnalysis Tokenomics Model and institutional research
A paid model for institutional clients that links lab revenue and token throughput for OpenAI, Anthropic and others to hardware demand. It has published lab gross-margin work, such as an estimate of Anthropic's API margins. It is the closest overlap with the cost-curve tracker, and it already sells to the crossover and hedge fund buyers.
- Epoch AI – AI Companies Data Hub
Free and CC-licensed. Tracks revenue, funding and compute spend for frontier labs from disclosures and press reports. The page I found was last updated in September 2025. Any paid normalised cost-curve product will be measured against it.
- Bloomberg – AI Circular Deals graphic
A one-off 2026 graphic of the deal loops between Microsoft, OpenAI, Nvidia and others. It covers counterparties but is not a feed, and it does not break out take-or-pay terms or lenders. Terminal users may see other Bloomberg Intelligence coverage.
- Wikipedia – AI build-out financing
A free crowd-edited table of major disclosed stakes and the commercial commitment attached to each. It shows how much of the basic counterparty map is already free.
- Newmarketpitch Neocloud Funding Tracker
Tracks 83 neocloud financings across 50 companies. It covers the vendor and financing side but does not tie deals to lab commitments.
- Augment.market
Publishes analysis that splits OpenAI's commitments from its debt, a light version of commitment mapping.
- Kepler.ai S-1 breakdown
An AI-driven filing analysis tool. It has already pulled compute-to-revenue numbers out of Anthropic's S-1 (compute was 55% of revenue in H1 2026), so it could extend that to future lab filings.
- Is AI Profitable Yet?
A non-institutional site that compares cumulative spend with revenue across major AI companies and says it updates monthly. The numbers are estimates.
- Dealogic / PitchBook (generic deal databases)
No product specific to AI compute showed up, but credit desks already use these to track SPV and private-credit deals such as Meta's Beignet/Hyperion. The linked piece is press coverage of that SPV debt.
Checked 2026-10-27
Risks
- The headline numbers conflict between sources: an operating profit versus $80B in first-half losses. Read the S-1 line items before building on any single figure
- A thin float and the founder-control structure may decouple the stock from fundamentals
- Price cuts like Terra's could compress gross margin faster than compute cost falls
Strong market pull and team fit; open space is the weak point (4).
tech 8 · pull 9 · space 4 · fit 9 · build 6
Exposed
Companies that could benefit or be threatened. News and context, not investment advice.
Momentum of its main topic
Sources
- S-1: Anthropic spent ~$2.50 on compute and infrastructure for every $1 of revenue in Q1 2025, falling to under 50 cents in Q2 2026, as it seeks profitability (Financial Times)Techmeme
- Anthropic opens its books for Wall Street, showing record revenue and mounting obligationsThe Decoder
- [AINews] Anthropic files for $2T IPO in 2 weeks: 1400% growth rate, 48% of revenue through cloud partners, 85% margin, $1B op profit in Q2, $134B cash in bank and $10B lifetime burnLatent Space
- Anthropic S-1, The Cost of Compute, Who Is Paying?Stratechery
- Anthropic's IPO plans: 900% revenue retention, but $80 billion in first-half lossesCNBC Technology
- Broadcom arranges a $63B deal with BofA, Blackstone, Citi, and Morgan Stanley to finance the purchase of chips for Anthropic via a new special purpose vehicle (Lynn Doan/Bloomberg)Techmeme
- Sources: Broadcom CEO Hock Tan held 11th-hour talks and agreed to higher rates and loan guarantees to secure bank participation in the $63B Anthropic debt deal (Wall Street Journal)Techmeme
- GPT-6.1 Terra halves OpenAI's API prices and still gains ground on its larger modelsThe Decoder
- Anthropic officially files for its public debutTechCrunch AI
Status history
- 2026-10-27 Flagged as buildable