Do compute costs really dominate AI company budgets?

Epoch AI's Luke Emberson and Yafah Edelman report that at the three AI companies where they can make estimates — Anthropic, Minimax and Z.ai — compute is the dominant expense: "R&D and inference compute together make up 54% to 62% of costs." Staffing, by contrast, is a comparatively small share: "Despite AI labs offering some of the highest salaries in tech, spending on staff accounts for less than 25% of total spending."

The authors categorize each company's spending into R&D compute, non-R&D (inference) compute, staffing, and other, expressed as a share of total operating expenses, including non-cash items like stock-based compensation — except for Anthropic, where staffing could not be separated out from the other categories. Their data sources differ by company: Anthropic's figures come from The Information's reporting of the company's "optimistic" projections as of December 2025 ($5.2B in EBITDA losses on $4.5B of revenue, implying $9.7B in total expenses; $4.1B of that spent on training, taken as R&D compute; a 40% gross margin implying $2.7B of inference compute against the $4.5B revenue). Minimax's and Z.ai's figures come from the detailed financial documents each company released for its January 2026 IPO.

This sits alongside Are AI agents now doing more research work than humans?, where a median daily agent inference cost above $600 is one visible symptom of the same compute-heavy cost structure Epoch AI finds at the balance-sheet level. It also contrasts with the buyer side captured in Are enterprises actually spending more on AI that works?: enterprises are moving AI spending into recurring IT budgets, while the labs selling them that compute carry the R&D and inference cost on their own books. And it sharpens the ambiguity in Why do individual AI gains not translate to business results? — if compute is this large a share of a lab's own costs, whether AI pays for itself depends as much on compute pricing and lab margins as on enterprise productivity gains. It also pairs with Do AI labs spend far more than they earn?, the companion finding from the same breakdown.

The estimate covers only three companies with unusual data availability — Anthropic via secondhand press reporting, Minimax and Z.ai via disclosures required for their IPOs — and Epoch AI flags its own Anthropic figures as "somewhat speculative" given that secondhand accounting. The breakdown says nothing about cost structure at labs that neither IPO nor have their financials leaked to the press, so this is a claim about three disclosed cases rather than a general law of AI-company economics, though the consistency of the pattern across three differently situated companies suggests compute dominance may generalize further than the sample alone can show.