Epoch AI's latest data reveals that among the leading AI companies Anthropic, Minimax, and Z.ai, the expenditure on computing power accounts for a staggering 57% to 70% of total costs, overwhelmingly surpassing talent compensation. Among these, Anthropic is projected to have a total expenditure of $9.7 billion in 2025, with computing power alone consuming $6.8 billion. More alarmingly, the expenditure scale of the three companies is 2 to 3 times their revenue.

Written by: Zhao Ying

Source: Wall Street Insights

Compared to talent, computing power is becoming the heaviest financial burden for AI companies.

According to the latest data from Epoch AI, in the top three AI companies—Anthropic, Minimax, and Z.ai—expenditure on computing power occupies an absolute dominant position in total costs.

In the largest company, Anthropic, the total expenditure for the entire year of 2025 is estimated to reach $9.7 billion, with computing power alone amounting to $6.8 billion, covering both model training and inference stages. This figure far exceeds the overall expenditure scale of Minimax and Z.ai during the same period.

The rapid expansion of computing power expenditure reflects the highly capital-intensive nature of developing and deploying cutting-edge AI models. Epoch AI estimates that the current expenditure level of the three companies is approximately 2 to 3 times their revenue, and the industry as a whole is still in a stage of large-scale cash burning.

Computing Power Dominates Cost Structure, Talent Expenditure Takes a Backseat

According to Epoch AI data, in Anthropic, Minimax, and Z.ai, the combined expenditure on research and inference computing power accounts for 57% to 70% of their total expenditures, exceeding the sum of employee compensation and other operational costs in every case.

This ratio is particularly pronounced in Z.ai, where 58% of expenditures are directly related to the computing power required for model development and training, showcasing a distinctly research-driven cost structure.

Although top AI laboratories pay some of the highest salaries in the tech industry to engineers and researchers, talent costs have not exceeded half of total expenditure in any of the three companies. This indicates that, in the current context of the AI arms race, the strategic value of chips and computing infrastructure has surpassed that of human resources at the financial level.

Divergence in Paths of US and Chinese AI Companies, Open Source Strategy Lowers Cost Threshold

It is noteworthy that both Minimax and Z.ai have released a large number of models in open-source form, allowing anyone to download, modify, and run the model weights for free.

In terms of data sources, Anthropic's figures are based on reports from The Information and have a speculative nature; while the data from Minimax and Z.ai comes from IPO prospectuses released in January 2026, making them relatively more reliable. The statistical periods for the three companies also differ: Anthropic covers the entire year of 2025, Minimax includes the first to third quarters of 2025, and Z.ai covers the first half of 2025. Epoch AI states that its total expenses include operating expenses, the cost of goods and services, as well as non-cash items such as equity incentives.

Together, these data paint a clear picture: In the current environment where investment in AI infrastructure remains high and profit models have yet to be validated, the ability to acquire and allocate computing resources is becoming the key variable determining the competitive position of AI companies.