--- title: "Is the Core of AI Competition Shifting from Talent Acquisition to Computing Power? Computing Costs Have Become the Largest Expense Item" type: "News" locale: "en" url: "https://longbridge.com/en/news/284159050.md" description: "Latest data from Epoch AI reveals that among three leading AI companies—Anthropic, Minimax, and Z.ai—computing expenditure accounts for 57% to 70% of total costs, completely overshadowing talent compensation. Among them, Anthropic's total projected expenditure for 2025 reaches $9.7 billion, with computing alone consuming $6.8 billion. More alarmingly, the expenditure scale for all three companies ranges from 2 to 3 times their revenue" datetime: "2026-04-27T06:01:01.000Z" locales: - [zh-CN](https://longbridge.com/zh-CN/news/284159050.md) - [en](https://longbridge.com/en/news/284159050.md) - [zh-HK](https://longbridge.com/zh-HK/news/284159050.md) --- # Is the Core of AI Competition Shifting from Talent Acquisition to Computing Power? Computing Costs Have Become the Largest Expense Item Compared to talent, computing power is becoming the heaviest financial burden for AI companies. According to the latest data from Epoch AI, among three leading AI enterprises—Anthropic, Minimax, and Z.ai—computing expenditure occupies an absolutely dominant position in total costs. At Anthropic, the largest of the three, total annual expenditure for 2025 is estimated at $9.7 billion, with computing costs alone reaching $6.8 billion, covering both model training and inference stages. This figure far exceeds the combined expenditure scale of Minimax and Z.ai during the same period. The sharp expansion of computing expenditure reflects the highly capital-intensive nature of cutting-edge AI model development and deployment. Epoch AI estimates that **the current expenditure scale of these three companies is approximately 2 to 3 times their revenue.** ## Computing Dominates Cost Structure; Talent Expenditure Takes a Back Seat Based on Epoch AI data, for Anthropic, Minimax, and Z.ai, **combined R&D computing and inference computing expenditures account for 57% to 70% of their respective total expenses,** exceeding the sum of employee compensation and other operating costs in every case. This proportion is particularly striking for Z.ai—the company's 58% of expenditure is directly linked to the computing power required for model development and training, presenting the most distinct cost structure oriented toward R&D computing. Despite top AI labs paying some of the highest salaries in the technology sector to engineers and researchers, talent costs have failed to breach the half-way threshold of total expenditure across all three companies. This means that in the context of the current AI arms race, the strategic value of chips and computing infrastructure has comprehensively surpassed that of human resources on the financial level. ## Divergent Paths Between Chinese and US AI Companies; Open-Source Strategies Lower Cost Thresholds It is worth noting that both Minimax and Z.ai release a significant number of models in open-source formats, allowing anyone to freely download, modify, and run the model weights. In terms of data caliber, Anthropic's figures are based on reports from The Information and carry a degree of speculation; Minimax and Z.ai's data come from IPO prospectus documents released in January 2026, making them relatively more reliable. The statistical periods for the three companies also differ: Anthropic covers the full year of 2025, Minimax covers the first three quarters of 2025, and Z.ai covers the first half of 2025. Epoch AI states that its total expense figures encompass operating expenditure, cost of goods and services, and non-cash items such as equity incentives. Collectively, the above data paint a clear picture: under the current circumstances of sustained high investment in AI infrastructure and unverified profit models, the ability to acquire and allocate computing resources is emerging as a key variable determining the competitive standing of AI companies. ### Related Stocks - [MSFT.US](https://longbridge.com/en/quote/MSFT.US.md) - [NVDA.US](https://longbridge.com/en/quote/NVDA.US.md) - [00100.HK](https://longbridge.com/en/quote/00100.HK.md) - [XSW.US](https://longbridge.com/en/quote/XSW.US.md) - [NVDL.US](https://longbridge.com/en/quote/NVDL.US.md) - [NVDY.US](https://longbridge.com/en/quote/NVDY.US.md) - [IGV.US](https://longbridge.com/en/quote/IGV.US.md) - [CLOU.US](https://longbridge.com/en/quote/CLOU.US.md) - [XLK.US](https://longbridge.com/en/quote/XLK.US.md) - [NVDX.US](https://longbridge.com/en/quote/NVDX.US.md) - [NVDU.US](https://longbridge.com/en/quote/NVDU.US.md) - [IGPT.US](https://longbridge.com/en/quote/IGPT.US.md) - [02513.HK](https://longbridge.com/en/quote/02513.HK.md) ## Related News & Research - [Anthropic Targets EU Data Center Market With A Six-Figure Recruitment Drive: Report](https://longbridge.com/en/news/283841321.md) - [Axe Compute Secures Record $260 Million AI Infrastructure Deal](https://longbridge.com/en/news/283669492.md) - [SUSE Launches SUSE AI Factory with NVIDIA](https://longbridge.com/en/news/283446429.md) - [08:00 ETNota AI Wins Grand Prize at NVIDIA Nemotron Hackathon, Proving MoE Quantization Prowess with Synthetic Data Technology](https://longbridge.com/en/news/283996925.md) - [LiveRamp integrates Nvidia AI infrastructure for GPU-optimized clean rooms](https://longbridge.com/en/news/284215903.md)