The fastest-growing segment of demand for Nvidia’s high-performance AI accelerators is no longer the official data center market—it’s the underground supply channel into China. According to the Financial Times, prices for servers equipped with the latest GPUs, which are banned for export, have doubled compared to U.S. price levels.
Specifically, servers based on the B300 architecture are being sold on illegal markets for roughly 7 million yuan (approximately 1 million U.S. dollars)—about twice the price available through legal supply chains in the United States. This reflects an acute shortage of top-tier computing solutions used for training AI models.
Price increases accelerated after new regulations from the U.S. Department of Commerce took effect in late May. These restrictions closed previously existing loopholes that allowed high-performance chips, including the Blackwell and Rubin series, to reach Chinese companies through overseas intermediaries.
At the same time, illegal supply channels existed well before these measures: estimates suggest that in 2025 alone, Nvidia AI chips worth around 1 billion U.S. dollars may have been smuggled into China. The new restrictions did not fully halt the flow, but they significantly increased the complexity of deliveries and drove chip prices higher.
Chinese companies continue to purchase such solutions due to intense competition in AI model development, where Nvidia accelerators remain the industry standard.
In parallel, the country is accelerating its own semiconductor production efforts in an attempt to reduce reliance on American technology and import restrictions.
Earlier, Nio Vice President Ma Lin publicly disclosed important information. Nio has been using Nvidia chips for several years, with peak annual purchases of Nvidia chips reaching 300 million U.S. dollars, prompting the company to begin developing its own chips.
As previously noted, Chinese companies’ dependence on Nvidia continues to decline, and the nation is strengthening its local supply chain for AI accelerators.
About the Author
