Chinese AI developers that rely on scarce top-end servers now face hardware costs set by layered brokers, raising training budgets and pushing some toward rentals or consumer graphics cards.
Quotes for an eight-GPU server that 36Kr calls the X-type rose from about RMB4 million late last year to about RMB15 million in September and above RMB20 million in the past month. 36Kr published the figures in a long report on China’s market for top-end AI servers, drawn from buyers, brokers and people who build computing clusters. The outlet does not name the chip behind its X-type and Y-type labels, and every person it quotes speaks under a pseudonym.
Large Chinese AI companies will pay about RMB15 million for such a machine, according to an intermediary 36Kr calls Lin Yue. Most of them have stopped buying above RMB20 million, the outlet reported, although requests from one leading model developer still circulate. Rental has climbed as well: monthly rent on a one-year lease had reached as much as RMB190,000, which 36Kr attributed to public reports, or more than RMB2 million a year.
The prices are hard to verify, and 36Kr says so. Some come from brokers and some from public reports, without a source for each, and Lin Yue told the outlet that many rumored deals never close. What the report adds is a description of how the price is formed.
Other reporting points to the same steep curve
The 36Kr figures line up with earlier reporting on Nvidia’s B300, the company’s most advanced AI accelerator, which US rules bar from sale to China without a license. In April, Reuters reported that a B300 server with eight GPUs cost about RMB7 million in China, nearly double the roughly RMB4 million of late last year, against about US$550,000 in the United States. That story, carried by The Standard, cited four industry sources and put one-year rentals at RMB190,000 a month, the same rent 36Kr gives. Nvidia told Reuters that "unlawful diversion is a recipe for failure" and that it gives no service or support for such systems.
On November 10, Bloomberg reported that financing records valued 32 Super Micro servers with B300 chips, linked to a service contract with Moonshot AI, at RMB432 million. That implies about RMB13.5 million per server, roughly 350% above the US rate, according to the copy published by Yahoo Finance. Moonshot disputed the documents, and Nvidia said the entities involved were not its customers. EastFrontier has followed how B300 financing in China is structured through leasing.
Each layer of brokers adds its own markup
36Kr describes a chain with equipment holders at the top and model developers, internet groups and computing service providers at the bottom. Between them sit brokers who find buyers for holders, brokers who find machines for buyers, and agents who buy on commission. Each handoff adds a margin to the previous quote. Buyers often give a broker their ceiling price and ask for any machine in stock to be locked immediately, which leaves little room for those who negotiate.
Some sellers now attach conditions. A buyer told 36Kr that purchasing a high-end server was "like buying Hermès, you have to take add-ons," with X-type machines sold only alongside Y-type ones. The Y-type eight-GPU server trades at about RMB6.5 million on the spot market, about RMB2 million more than the price through a manufacturer’s allocation.
Zhao Cheng, as 36Kr calls a contractor who deploys and maintains servers for buyers, said "most model companies have given up buying." He told the outlet that sellers may break contracts to resell at higher prices, so even a locked order is no guarantee. Some large companies have begun asking about consumer graphics cards such as the RTX 5080.
Demand data from MIIT explains the pressure
36Kr ties the squeeze to rapid growth in China’s computing build-out. China’s intelligent computing capacity reached 2,185 exaflops at the end of June, up 177% from a year earlier, according to the Ministry of Industry and Information Technology (MIIT) figures reported by Xinhua. 36Kr, citing MIIT, said the country had built 52 computing facilities with more than 10,000 accelerator cards each. By the end of September capacity had grown to 2,592 exaflops, up 146%, MIIT spokesperson Xie Cun said on October 20, according to 21st Century Business Herald.
Supply is not keeping pace, 36Kr argued, citing SK Hynix and TSMC on tight memory and equipment supply. Zhao Cheng estimated that, in an optimistic case, ready-made modules for assembling Y-type servers could last about half a year, after which prices would rise further. The next figure to watch is whether quoted prices hold above RMB20 million into December, or whether large buyers’ RMB15 million ceiling pulls them back.