When you visit the renowned Huaqiangbei electronics area in Shenzhen, China, you’re likely to find a variety of shops in the SEG Plaza skyscraper offering a wide range of electronic components, including camera parts and drones.
While high-end U.S. chips are not openly advertised, discreet inquiries can lead you to vendors who specialize in them.
However, these chips come at a steep price.
Anonymous vendors in the area revealed that they could provide limited quantities of A100 artificial intelligence chips, designed by a prominent U.S. chip manufacturer, at a staggering cost of $20,000 per chip – twice the usual price.
While buying or selling these high-end U.S. chips is not technically illegal in China, the export restrictions imposed by the U.S. government have created an underground market.
Vendors are cautious not to attract attention from either U.S. or Chinese authorities.
In September, President Joe Biden’s administration ordered Nvidia, the chip designer, to halt the export of its most advanced chips, including the A100 and the recently developed H100, to mainland China and Hong Kong.
These restrictions were part of broader efforts to hinder the progress of Chinese AI and supercomputing development, given the escalating political and trade tensions. Subsequently, various export controls related to semiconductors were implemented.
However, with the global boom in AI, fueled by the success of OpenAI’s ChatGPT, there has been a surge in demand for high-end chips, particularly Nvidia’s microprocessors, which are widely regarded as the best for handling machine learning tasks.
Ivan Lau, the co-founder of Pantheon Lab in Hong Kong, confirmed that he is currently in talks with two vendors to procure 2-4 new A100 cards for the startup’s latest AI models.
These vendors, who acquired the chips outside the U.S., quoted a price of HK$150,000 ($19,150) per card, emphasizing that there would be no warranty or support.
Multiple vendors in Hong Kong and mainland China have reported the availability of small quantities of A100 chips, highlighting both the strong demand for these chips in China and the relative ease with which small-scale transactions can circumvent U.S. sanctions.
It is challenging to estimate the overall volume of Nvidia A100 and H100 chips flowing into China or determine the extent to which these transactions meet the demand.
The buyers classically include app researchers, developers, gamers, startups, and perhaps even Chinese local authorities. However, due to the violation of U.S. trade restrictions, the vendors chose to remain anonymous.
Nvidia clarified in a statement that it does not allow exports of A100 or H100 chips to China. Instead, it provides alternative products with reduced capabilities that comply with U.S. law.
The company also stated that it would take immediate action if it receives information about customers breaching their agreement and exporting restricted products unlawfully.
The U.S. Department of Commerce spokesperson acknowledged the substantial impact of export control measures on China’s access to high-end chips.
They emphasized that allegations of violations are investigated and not surprising given the reports of parties attempting to acquire these chips through illicit means.
Despite reaching out for comments, China’s State Council Information Office and the country’s industry ministry did not respond.
Nvidia had expressed concerns in September that if Chinese companies refrained from purchasing alternative Nvidia products, it could result in a loss of $400 million in sales during the third quarter.
Consequently, Nvidia developed China-specific slower variants, namely the A800 and H800, to mitigate this impact.
Large Chinese tech companies such as Tencent Holdings and Alibaba, with substantial financial resources, are now purchasing these variants in significant quantities.
The Chinese vendors revealed that they primarily acquire these chips in two ways: by purchasing excess stock that reaches the market after Nvidia supplies large quantities to major U.S. economies who have high demand for it.