The core of AI competition is shifting from vying for talent to competing in computing power? Computing power has become the largest cost item.

Compared to talent, computing power is becoming the heaviest financial burden for AI companies. The latest data from Epoch AI shows that, among the three leading AI companies—Anthropic, Minimax, and Z.ai Zhiyu—spending on computing power dominates the total costs. At the largest of them, Anthropic, total spending for 2025 is estimated at $9.7 billion, with computing power alone accounting for $6.8 billion, covering both model training and inference. This figure far exceeds the overall expenditure of Minimax and Zhiyu during the same period. The rapid expansion in computing power expenditure reflects the highly capital-intensive nature of frontier AI model development and deployment. Epoch AI estimates that the current expenditure for these three companies is about 2 to 3 times their revenue, and the industry as a whole is still in the stage of large-scale burning of money. Computing power dominates cost structure; talent expenditure takes a back seat According to Epoch AI data, for Anthropic, Minimax, and Zhiyu, the combined cost of R&D computing power and inference computing power accounts for 57% to 70% of their total spending, in every case exceeding the combined costs of employee salaries and other operational expenses. This ratio is especially prominent at Zhiyu—58% of its expenses are directly tied to computing power for model development and training, showing the strongest R&D computing-oriented cost structure. Although top AI labs pay engineers and researchers among the highest salaries in the tech industry, talent costs at the three companies fail to surpass half of their total spending. This means that, in the context of the current AI arms race, the strategic value of chips and computing infrastructure has fully outweighed talent resources on the financial level. Diverging paths between Chinese and American AI firms; open-source strategies lower the cost threshold Notably, both Minimax and Zhiyu have released many of their models as open source, with model weights available for anyone to download, modify, and run. Regarding the data sources: Anthropic’s figures are based on reporting from The Information and are somewhat speculative; Minimax and Zhiyu's numbers come from IPO prospectuses released in January 2026 and are relatively more reliable. The reporting periods also differ: Anthropic is for the full year 2025, Minimax covers the first to third quarters of 2025, and Zhiyu is for the first half of 2025. Epoch AI states its totals include operational costs, the cost of goods and services, and non-cash items such as equity incentives. Together, these data paint a clear picture: In a time when investment in AI infrastructure remains high and profitability has yet to be proven, the ability to acquire and manage computing resources is becoming the key variable determining AI companies' competitive positions. Risk Warning and Disclaimer The market involves risks, and investments should be made cautiously. This article does not constitute personal investment advice, nor does it consider the specific investment goals, financial situation, or needs of individual users. Users should consider whether any opinion, view, or conclusion in this article suits their particular circumstances. Investment based on this article is the user's responsibility.