Business news |
By Brian Tristam Williams
Chinese buyers are paying sharply higher prices for restricted AI hardware, with Nvidia A100 servers reportedly tripling in price as US enforcement and Chinese import barriers squeeze both illicit and authorised supply routes.
Servers built around Nvidia’s five-year-old A100 accelerator have risen from about 200,000 yuan to as much as 600,000 yuan, or roughly $82,000, since late 2025, the report says. Demand has also pushed buyers towards repurposed gaming GPUs and older Ampere-generation systems that can still be used for inference workloads.
The pricing reflects a constrained market rather than a sudden improvement in the A100 itself. The accelerator was launched in 2020 and has been overtaken by Hopper and Blackwell-class hardware, but it remains useful for AI training, fine-tuning and inference where software compatibility, memory capacity and CUDA support are more important than owning the newest chip.
The squeeze has moved up the product stack. Reuters previously reported that Nvidia B300 servers were trading in China at about 7 million yuan, or roughly $1 million, in April as black-market supply tightened. The Financial Times now reports that DGX B300 systems have moved above 8 million yuan, while RTX 6000 Pro workstation cards have also climbed sharply.
The shortage has two sides. US authorities have stepped up enforcement against alleged diversion of controlled AI systems, including a March indictment in which three people were charged with conspiring to unlawfully route advanced AI servers to Chinese customers. The US Department of Justice said advanced AI accelerator chips and servers containing them require export licences for China and Hong Kong.
At the same time, the legal route for newer H200 GPUs has not become a working supply channel. As previously reported by eeNews Europe when Trump said H200 exports could be permitted, the proposal still needed clear policy mechanics. Reuters later reported that US Commerce Secretary Howard Lutnick said Nvidia had not yet sold H200 chips to Chinese companies, citing Beijing’s focus on domestic alternatives.
That leaves Chinese AI firms facing higher costs for any Nvidia-compatible capacity they can obtain, while domestic alternatives from Huawei and others continue to scale. As eeNews Europe has also tracked in relation to cloud GPU access and export controls, Washington is increasingly treating AI compute as a controllable resource, whether it moves as a physical server or is rented remotely.
The result is a distorted compute market in which restricted hardware gets older but not necessarily cheaper. Nvidia A100 servers are now a stopgap for buyers locked out of newer supply, and that makes even 2020-era accelerators valuable in a market still hungry for AI capacity.
If you enjoyed this article, you will like the following ones: don't miss them by subscribing to :
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。