Charted: Compute Costs More Than Talent in AI
See visuals like this from many other data creators on our Voronoi app. Download it for free on iOS or Android and discover incredible data-driven charts from a variety of trusted sources.
Key Takeaways
- Compute is the largest cost for all three AI firms in the dataset, accounting for 57% to 70% of total spending.
- At Anthropic, compute spending reaches $6.8 billion in 2025 across model training and inference.
- In this dataset, compute costs exceed staff and other expenses, highlighting how infrastructure—not talent—drives AI spending.
For leading AI companies, the biggest expense is not talent. It is compute.
This chart from Visual Capitalist’s AI Week, sponsored by Terzo, uses Epoch AI data to compare spending at Anthropic, Minimax, and Z.ai across R&D compute, inference compute, and staff plus other costs.
In every case, compute accounts for the majority of total spending, underscoring how capital-intensive it has become to build and serve frontier AI models.
How AI Company Costs Break Down
Despite differences in scale, all three companies allocate the largest share of their budgets to a single category: compute.
The data below compares spending composition across Anthropic, Minimax, and Z.ai. Anthropic’s figures are for 2025, while Minimax’s are from Q1 to Q3 of 2025 and Z.ai’s are for H1 2025.
| Costs Category | Anthropic | Minimax | Z.ai |
|---|---|---|---|
| R&D Compute (Billions, USD) | 4.10 | 0.14 | 0.18 |
| Inference Compute (Billions, USD) | 2.70 | 0.04 | 0.01 |
| Staff and Other (Billions, USD) | 2.90 | 0.14 | 0.12 |
| Total (Billions, USD) | 9.70 | 0.32 | 0.31 |
| R&D Compute Share | 42% | 44% | 58% |
| Inference Compute Share | 28% | 13% | 3% |
| Staff and Other Share | 30% | 44% | 39% |
Across all three AI companies, compute is the main cost center. Epoch AI estimates that R&D compute and inference compute together account for 57% to 70% of total spending, making infrastructure more expensive than staff and other costs in every case.
Among the three, Z.ai has the most R&D-heavy profile, with 58% of spending tied to compute powering model development and training.
Anthropic stands out for sheer scale. Epoch AI estimates the company spent $9.7 billion in 2025, including $6.8 billion on compute alone across training and inference.
Its costs are significantly higher than Minimax’s and Z.ai’s, even if the two Chinese AI companies’ figures were annualized to match Anthropic’s full-year period.
Both Chinese companies release many of their models as open source, meaning the model weights are freely available for anyone to download, modify, and run. This strategy helps them compete with better-funded U.S. labs by building developer adoption at a fraction of the cost.
AI Talent Costs Less Than Chips and Compute
One of the clearest takeaways is that talent costs less than compute in this comparison. Even though top AI labs pay some of the highest salaries in tech, staff and other costs still account for less than half of total spending at each of the three firms.
While the chart focuses on costs, Epoch AI estimates these labs are currently spending around 2–3x more than they generate in revenue, even as some expect economics to improve over time.
How These Estimates Were Built
This dataset comes with a few important caveats. Anthropic’s figures are based on reporting from The Information and are more speculative, while Minimax and Z.ai figures come from IPO filings released in January 2026.
The time periods also differ: Anthropic data is for the full year of 2025, Minimax covers 2025 Q1–Q3, and Z.ai covers 2025 H1. Epoch AI says its expense totals include operating expenses, cost of goods and services, and non-cash items such as stock-based compensation.
Learn More on the Voronoi App 
If you enjoyed today’s post, check out The Soaring Revenues of AI Companies on Voronoi.
