Token factories: China shifts data centers from power to token output

The Chinese infrastructure sector is changing how it values data centers. It no longer counts cards, compute scale, or PUE ratios. The metric is now how many useful tokens a facility produces from the same electricity. The Chinese outlet 量子位 illustrates this with 中科类脑, which calls its system a "token factory coordinating compute and energy".

Official data confirms the scale. The state data administration notes daily token calls in China grew from 100 billion in early 2024 to 140 trillion in March 2026. That is a thousandfold jump in two years. Industry data shows API prices for some models fell by more than 90 percent. Volume is surging, but the margin on a standard token is shrinking rapidly.

Three layers instead of one

中科类脑 breaks the problem into three levels. First, it optimizes compute and tokens via heterogeneous scheduling and inference acceleration. Second, it coordinates energy and capacity. Third, it handles billing and operations. The goal is a flexible "token production line". A company needing 30 cards on average, with peaks at 100, builds only the base capacity. It rents the remaining 70 cards from the platform during peak times. Partner 常峰 explains the economics: one unit of industrial electricity costs a few mao. Converted to compute, the value reaches several dozen yuan. As tokens, some models sell a million for over 100 yuan.

The platform, operating as BitaHub, connects over 3,000 compute nodes, 5,000 P of capacity, and 80,000 companies and research users. A "decision brain" developed with the University of Science and Technology of China steers the coupling of energy, compute, and tokens. In a test spanning Shanghai, Wuhu, and Urumqi, control response time stayed under 200 seconds. Migration success between regions hit 100 percent. Energy consumption forecast accuracy reached 98 percent.

The market looks at production, not purchases

Larger players show the same shift. At a summit on heterogeneous inference, a representative of 商汤大装置 noted average daily tokens served rose from 460 billion in February to 4.5 trillion in August 2026. That is nearly tenfold in six months. The company is moving focus from single queries to long context, prefix reuse, and bursty, long-tail traffic patterns.

The takeaway for the market is clear. If token prices drop by more than 90 percent, the advantage is no longer in buying cards. It lies in delivering them cheaply and reliably to the user.