China's compute-leasing market has quietly crossed a line this year. It is no longer a speculative theme; it is producing delivered earnings, and the people paying attention include some unlikely names.
At the Hong Kong Global Financial Leaders' Investment Summit, Blackstone President and COO Jon Gray made a comparison that stuck with me. AI compute demand, he said in a November 8 report by Hong Kong News, now runs well ahead of supply — the mirror image of the late-1990s telecom boom, when networks were built first and demand followed. Dong Guanzhi, a professor at Jinan University and dean of the China New Industrial Economy Research Institute, put it more bluntly in the same report. Leasing has restructured how compute is supplied, demand is penetrating every industry, and yet competition will become a contest of end-to-end delivery capability, concentrating gains among a few leaders. Read those two together and the picture is odd, demand everywhere, winners scarce.
The order book backs that up. Beijing's No.1 token factory has passed a daily output of 1.4 trillion tokens in its first phase, heading toward a long-term target of 10 trillion per day. Then in October, Runjian announced that its Wuxiang Yungu token factory signed a one-year, US$200 million token-service order with an overseas client. The compute sits entirely in Chinese data centers; billing happens by tokens across borders. It is the first major overseas token deal won by a Chinese firm. A China IRP report sketches the size of the pond, projecting China's compute-leasing market at roughly 260 billion yuan of potential revenue in 2026, growing more than 20% a year.
Compute-leasing market at a glance
260B yuan — projected 2026 China compute-leasing revenue potential (China IRP), +20% a year
1.4T tokens/day — phase-one output of Beijing's No.1 token factory
US$200M — overseas token-service order signed by Wuxiang Yungu (announced October)
Supply, though, is splitting into a K shape. Dong noted that low-end general compute is abundant, crowded and price-competitive, with margins shrinking by the quarter. High-end AI compute is a different animal, constrained by hardware, compliant facilities and professional operations, which is why quality capacity sits with a few early movers. Founder Securities frames the same split through three drivers — 密集 large orders, cloud price hikes and domestic hardware substitution. Scarcity at the top, involution at the bottom.
Local policy tools keep widening the competitive radius. Fuzhou's Mindou Ruci token factory went live on November 3, aggregating 14 mainstream models and linking compute clusters in Pingtan, Suzhou (Anhui), Zhongwei and Xining for cross-region scheduling. Compute vouchers and token credit keep lowering the entry barrier for smaller firms. For lessors the contest has shifted. It is no longer about scale — it is scheduling efficiency, and delivery into real scenarios.