An Open Financial Model for AI Labs
For this exercise, I used a mix of public reporting, estimates, and audited filings to assemble a reconstruction of potential frontier-lab financials. The full Excel model is also embedded in this page, which allows you to toggle around the inputs and recompute the math depending on your views.
This analysis is obviously highly illustrative (not investment advice), and is a living / breathing model that's subject to change. Feedback and corrections are welcome! Email me here if you have any thoughts.
Key findings from the model
As opposed to the traditional SaaS business model, frontier labs comprise a hybrid of SaaS margins, cloud-infrastructure capex and project finance. My colleague Braeden showed that serving is ultimately gross-margin positive across the board, and through this project I hope to elucidate the question of whether run-rate revenue can turn into durable cash flow after capacity commitments and price compression.
Change the assumptions
Sliders override the workbook's input cells and re-run its formulas in the browser. Edited cells turn amber. The tiles, charts, and workbook grid recompute together.
Full workbook
How this was built
From Excel to this page
The workbook's cells, formulas, formats, and comments were extracted verbatim. A spreadsheet engine re-evaluates all ~1,400 formulas in the browser; it reproduces Excel's own computed values exactly. Edits stay in the browser. Refresh to reset.
Reading the model
Standard banker convention: blue cells are hardcoded inputs with the source in a cell note; black cells are formulas. Pasted-over values and conservative plugs are listed in the Audit Log tab.
Sources
Reported and leaked figures: The Information, WSJ (chart estimates flagged as such), CNBC, FT-verified Ed Zitron financials, J.P. Morgan EOTM, Coatue, company announcements. Audited figures: Z.ai and MiniMax HKEX listing documents. Serving economics: SGLang GLM-5.2 B200 benchmarks + SemiAnalysis TCO constants; the GLM-5.2 serving-cost reconstruction behind the Margin Model draws on my colleague Braeden's token-economics work, with the frontier comparisons and price-cut scenarios built on top of it. Each figure's link lives in its cell note.
Known limitations
Anthropic Q3/Q4 2026 revenue is plugged at the minimum required to reach the company's $18B target, to show how little H2 revenue the target assumes. The monthly path is a recognized-revenue run-rate proxy, not true ARR. OpenAI's 2026–30 gross margin is not forecast and shows 0%; see the Audit Log.
Caveats
This is a public reconstruction of private companies from reporting of varying quality; treat outputs as directional. Estimates read from article chart images are flagged chart_pixel_estimate in the WSJ Pull tab. Nothing here is investment advice.
Download
Download the .xlsx. It is the same file this page runs.