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

China's infrastructure sector is changing how it measures the value of a data centre. The number of cards, the scale of compute and the PUE ratio matter less than how many useful tokens a facility can squeeze out of the same electricity. The Chinese outlet 量子位 describes the shift through 中科类脑, which calls its system a "token factory coordinating compute and energy".

The state data administration puts numbers on the scale. Daily token calls in China rose from 100 billion at the start of 2024 to 140 trillion in March 2026, more than a thousandfold over two years. Industry tallies show API prices for some models fell by more than 90 percent in total. Volume is growing fast. The margin on an ordinary token is flattening just as fast.

Three layers instead of one

中科类脑 splits the problem into three levels. The first optimises compute and tokens through heterogeneous scheduling and inference acceleration. The second coordinates energy and capacity. The third handles billing and operations. The result is meant to be a flexible "token production line". A company with a typical demand of 30 cards, and a peak reaching 100 cards, builds only the base part itself and rents the missing 70 cards from the platform at peak. The company's partner, 常峰, explains the value logic directly: one unit of industrial electricity costs a few mao, once converted into compute power the value reaches several dozen yuan, and once turned into tokens some models sell a million tokens for more than 100 yuan.

The platform, which operates under the name BitaHub, now has more than 3 000 compute nodes attached, more than 5 000 P of capacity and more than 80 000 companies and research users. The backing is a "decision brain" developed jointly with a team from the University of Science and Technology of China, which steers the coupling of energy, compute and tokens. In a coordination test stretched between Shanghai, Wuhu and Urumqi, the control response time stayed below 200 seconds, the success rate of migration between regions reached 100 percent, and the accuracy of the energy consumption forecast reached 98 percent.

The market looks at production, not purchases

The same turn is visible at larger players. At an industry summit devoted to heterogeneous inference, a representative of 商汤大装置 described how the average daily number of tokens served rose from 460 billion in February to 4.5 trillion in August 2026, almost tenfold in half a year. The company is shifting the emphasis from the single query to long context, prefix reuse and bursty, long-tail traffic patterns.

The conclusion for the rest of the market is simple. If tokens get more than 90 percent cheaper, the advantage no longer lies in buying cards. It lies in how cheaply and reliably they are delivered to the user.