Across three AI companies where we can make estimates, compute is the dominant expense: R&D and inference compute together make up 54% to 62% of costs. Despite AI labs offering some of the highest salaries in tech, spending on staff accounts for less than 25% of total spending.
| Company | Year | Metric | Value mid (billions) |
|---|---|---|---|
| Anthropic | 2025 | Revenue | 9.70 |
| Anthropic | 2025 | Compute (R&D) | -4.10 |
| Anthropic | 2025 | Compute (Non-R&D) | -2.70 |
| Anthropic | 2025 | Other | -2.90 |
| Z.ai | 2025-H1 | Revenue | 0.03 |
| Z.ai | 2025-H1 | Compute (R&D) | -0.18 |
| Z.ai | 2025-H1 | Compute (Non-R&D) | -0.01 |
| Z.ai | 2025-H1 | Staff | -0.07 |
| Z.ai | 2025-H1 | Other | -0.05 |
| Minimax | 2025 Q1 to Q3 | Revenue | 0.05 |
| Minimax | 2025 Q1 to Q3 | Compute (R&D) | -0.14 |
| Minimax | 2025 Q1 to Q3 | Compute (Non-R&D) | -0.04 |
| Minimax | 2025 Q1 to Q3 | Staff | -0.05 |
| Minimax | 2025 Q1 to Q3 | Other | -0.09 |
Currently, spending at each of these labs is 2-3x higher than revenue. While Minimax and Z.ai both indicate that they will remain unprofitable for the foreseeable future, Anthropic expects to become profitable by 2028.
Epoch's work is free to use, distribute, and reproduce provided the source and authors are credited under the Creative Commons BY license.
Learn more about this graph
We break down spending at 3 notable AI companies where data availability permits: Anthropic, Minimax, and Z.ai. Spending is categorized into R&D compute, non-R&D compute, staffing, and other, although for Anthropic we aren’t able to separate staffing expenses.
We report spending categories by the share of total operating expenses, and include non-cash expenses like stock-based compensation.
Data
Our data sources differ by company. For Anthropic, data comes from reporting at The Information:
- As of December 2025, Anthropic’s “optimistic” projections were for $5.2 billion in EBITDA losses on $4.5B of revenue for 2025, suggesting total expenses of $9.7B. Since these figures were near year-end, we expect them to be close to final reports.
- In the same article, they report that Anthropic spent $4.1B training its models in 2025, which we take to be R&D compute costs. It also reports a gross margin of 40%, suggesting that the $4.5B of revenue involved $2.7B in inference compute expense.
Minimax and Z.ai each made IPOs in January 2026, which required the public release of detailed financial documents. These documents provided detailed information on revenues and spending across categories.
| Anthropic (2025) | Minimax (2025 Q1-Q3) | Z.ai (2025 H1) | |
|---|---|---|---|
| Revenues | 4500 | 53 (p. I-5) | 27 (p. I-4) |
| Expenses | 9700 | 314 (p I-5) | 312 (p. I-4) |
| R&D compute | 4100 | 142 (p. 423) | 183 (p. 258) |
| Inference compute | 2700 | 38 (p. 418) | 9 (p. 256 and 259) |
| Staff | 2900 | 48 (p. I-34) | 72 (p. I-37) |
| Other | 86 (residual) | 48 (residual) |
All figures in millions of USD
Note that when tabulating total expenses, we include operating expenses and the cost of goods and services, but omit other line items like the cost of financing and on-paper changes in the fair value of financial liabilities.
Assumptions
Second-hand reports of financial statements carry a degree of uncertainty over the exact financial accounting being used. For that reason, our estimates for Anthropic are somewhat speculative.
Download this data
Explore this data
AI company data, with data on revenue, funding, staff, and compute for many of the key players in frontier AI.
Related insights
