Rebuilding the AI labs’ financials

An interactive financial model of Anthropic, OpenAI, Z.ai and MiniMax, rebuilt from filings, press reporting and estimates, running live in this page. With the labs preparing 2027 IPOs, the question is whether run-rate revenue becomes durable cash flow once compute commitments and price compression bite. Only Z.ai and MiniMax report audited financials; the frontier figures are reconstructions, and the limitations log lists what is estimated. isaiprofitable.com tracks headline profitability; Braeden’s token-economics work underpins the Margin Model tab. Change any input and the model recomputes.

Highly illustrative and not investment advice. Feedback is wanted and welcome! This is a living, breathing analysis.

Key findings from the model

Bullish and bearish read on the frontier labs' profitability.

Benchmarking the open Chinese labs against the frontier

Z.ai and MiniMax report audited financials in their HKEX listing documents; the frontier columns are the model's reported and leaked estimates. Chinese-lab figures are annualized from the latest filed period.

R&D weighs far more heavily for Z.ai and MiniMax because they own their compute (Z.ai's audited capex ran ~40% of FY24 revenue) and expense training as incurred. In contrast, the frontier labs rent capacity through partner commitments that sit outside the P&L until used.

Change the assumptions

The sliders map to inputs in the workbook, and as you drag, the full model will recompute.

Full workbook

Method & sources

Sources

• Z.ai (Zhipu) HKEX listing application — audited financials
• MiniMax HKEX listing application — audited financials
• Ed Zitron — FT-verified OpenAI financials
• J.P. Morgan, Eye on the Market (Jun 2026) — contracted compute rates
• CNBC — Anthropic Q1/Q2 2026 results
• Anthropic announcements — raises, ARR, compute partnerships
• The Information — lab revenue and margin reporting
• WSJ — OpenAI and Anthropic IPO finances — chart estimates, flagged chart_pixel_estimate in the Press Pull tab
• SGLang GLM-5.2 B200 benchmarks
• SemiAnalysis InferenceX — GPU TCO constants
• Braeden Norris — GLM-5.2 token economics — the serving-cost work behind the Margin Model

Each cell in the workbook carries its specific source in the cell note.

Limitations log