Nvidia's advanced AI chips (H100, RTX Blackwell, DGX), banned from export to China by US regulations, are selling on black market at approximately double their official prices, revealing enforcement gaps and sustained demand for premium compute capacity.
Nvidia's China problem persists despite US export controls. Black-market prices reveal that demand for banned chips remains robust—buyers now pay more than double the official retail price, while Beijing simultaneously directs procurement toward Huawei as a domestic alternative.
Jensen Huang, Nvidia's chief executive, warned Washington months ago about the policy's ineffectiveness. At Citadel Securities' Future of Global Markets event, he stated that Nvidia had fallen from roughly 95% of China's advanced AI accelerator market to zero after the export controls tightened, calling the policy "already largely backfired." Recent market data supports this concern.
According to reporting by the Financial Times, Nvidia's banned AI hardware is now selling on China's black market at prices that undermine the restrictions. The DGX B300 system, which retails in the US for approximately $400,000, now fetches more than Rmb8 million—roughly $1.1 million—in China, according to traders interviewed by the FT. Six months prior, the same system sold for around Rmb4 million. Nvidia's RTX 6000 Pro workstation chip has jumped from approximately Rmb50,000 at the start of the year to as much as Rmb130,000.
These markups indicate that a parallel market remains functional despite tightened enforcement around re-export routes by Washington, Taiwan and Malaysia. The FT's reporting also noted that older A100 chips are being repurposed and that GPU rental prices in China now match or exceed US levels—a reversal from the earlier period when smuggled supply made Chinese compute cheaper. This price signal carries more significance than policy language alone.
The Super Micro Computer case demonstrates the scale of circumvention channels. In March, Reuters reported that US prosecutors charged three people tied to Super Micro, including co-founder Yih-Shyan "Wally" Liaw, with diverting approximately $2.5 billion in Nvidia-powered servers to China. The company itself was not charged; Liaw subsequently resigned from Super Micro's board following the indictment. Prosecutors alleged the operation used a Southeast Asian front company and falsified shipping records to move restricted hardware to Chinese buyers.
This case illuminates a fundamental challenge: chip controls are only as effective as the enforcement infrastructure supporting them. AI servers are valuable, portable and difficult to monitor once they pass through brokers, cloud resellers and third countries. A policy premised on Chinese buyers simply ceasing to purchase Nvidia chips lacks realism. Determined buyers will pay more, wait longer, rent capacity, strip older systems for parts or use intermediaries—the predictable behavior of any constrained market.
Huawei presents the complementary pressure. According to Tom's Hardware reporting from late May, Huawei rotating chairman Xu Zhijun publicly stated that "if the United States hadn't forced China and its companies, they wouldn't have built this way," thanking the US for the pressure. This comment followed questions about Huawei's LogicFolding chip architecture, presented as a method to reorganize chip layouts and improve data movement. While Huawei remains unequal to Nvidia, the ban has effectively created a captive domestic customer base and provided patriotic justification for sustained investment.
Huawei's Ascend line has become the primary alternative for Chinese firms, accelerated by both US restrictions and Beijing's own directives to local companies to favor Chinese hardware. The Guardian, citing the Financial Times, reported in January that Chinese customs officials blocked Nvidia H200 shipments despite US approval under a licensing regime tied to a 25% tariff. Chinese authorities simultaneously discouraged domestic technology companies from purchasing the chips absent necessity. Washington cracked the door; Beijing directed its companies elsewhere.
Nvidia's position remains complicated by software's portability. CUDA remains the standard language for many AI engineers, and Chinese companies operating smuggled, rented or legacy Nvidia equipment continue operating within that ecosystem. However, software loyalty erodes as hardware supply becomes expensive and politically unstable. Huang has characterized China as only "nanoseconds behind" in chipmaking—rhetoric that nonetheless captures a real risk: each month of restricted access strengthens the case for domestic suppliers like Huawei and Cambricon to close the gap.
For investors and operators, the black-market premium offers the clearest market signal. A DGX B300 selling for more than twice its US retail price confirms that Chinese AI demand has not been crushed—it has been repriced. Some buyers will continue paying the premium for immediate Nvidia performance. Others will conclude that funding domestic alternatives, accepting lower efficiency and avoiding future customs seizures or indictments represents the better hedge.
The export controls have produced genuine effects: top-end compute is harder to obtain, costs have risen and companies face harder procurement decisions. Yet the evidence suggests the goal of preventing Nvidia-class capabilities from reaching China has not been achieved. Nvidia lost legitimate revenue, brokers captured the spread and Huawei gained the kind of urgency that conventional industrial policy rarely generates. The uncomfortable reality: the chips continue arriving, simply at black-market prices, while companies unable to obtain them build around the shortage.