Legacy Nvidia Silicon Prices Triple in China Amid Smuggling Crackdown
A sweeping smuggling crackdown and customs freeze have tripled prices for five-year-old Nvidia A100 servers in China, driving figures up to $82,000.
A severe crackdown on grey-market smuggling routes and a customs freeze have sent shockwaves through China's underground AI hardware market, according to the Financial Times. Prices for legacy enterprise servers built around Nvidia's five-year-old A100 accelerators have roughly tripled since late last year. Chinese companies are paying as much as 600,000 Chinese Yuan, or about $82,000, for systems that previously cost around 200,000 yuan. This dramatic price surge reflects the intense pressure facing local firms and research institutions as alternative supply channels dry up.
The squeeze extends across multiple tiers of restricted architectures within the secondary market. According to the Financial Times, Nvidia's flagship DGX B300 system has doubled in price over the past six months, now commanding more than 8 million yuan, or roughly $1.1 million, compared to its U.S. retail price of nearly $400,000. Restricted Blackwell hardware like the RTX 6000 Pro workstation card has similarly jumped from roughly 50,000 yuan at the start of the year to as much as 130,000 yuan. Even renting hardware offers no relief, with local GPU rental rates now matching or exceeding U.S. pricing.
Regulatory pressures have intensified significantly across international borders. Washington tightened enforcement at the end of last year, and a Supermicro co-founder was charged in March over an alleged $2.5 billion scheme routing Nvidia AI servers to Chinese buyers. Subsequent smuggling investigations launched by authorities in Taiwan and Malaysia effectively choked off the re-export pathways traditionally exploited by traders. Commenting on the situation, Nvidia noted that building data centers from smuggled silicon remains a dead-end since the company provides no support or repairs for restricted products.
Policy actions from Beijing have compounded the hardware shortage from the opposite direction. Following the approval of H200 exports to China, Chinese customs officials were instructed to block the chips at the border, and Commerce Secretary Howard Lutnick later confirmed that Nvidia had not sold a single H200 unit to a Chinese company months afterward. These combined restrictions leave engineering teams with few viable ways to scale high-performance computing clusters without relying on aging infrastructure or modified gaming GPUs pulled into inference workloads.
These bottlenecks drive buyers toward domestic silicon options such as Huawei, which launched its Ascend 950PR inference chip in March. According to the Financial Times, the 950PR is currently undergoing testing at large data center clients across China. However, production output remains limited, and the native CANN software stack substantially trails Nvidia's mature CUDA platform. Furthermore, rising component costs driven by ongoing DRAM and HBM shortages across the hardware stack complicate any transition away from established Nvidia architectures.
Until Huawei successfully scales production of the Ascend 950PR or Beijing alters its stance on H200 imports—an outcome considered highly unlikely—prices for remaining A100 inventory within China are expected to remain elevated. For infrastructure engineers deploying machine learning workloads, the ongoing hardware scarcity highlights the severe switching costs and software ecosystem barriers that continue to protect Nvidia's legacy dominance in the region.