In the bustling center of Shenzhen’s Huaqiangbei electronics district, a discreet inquiry may unearth a hidden treasure: Nvidia’s high-end A100 AI chips. The AI processors, sold for a whopping $20,000 each, which is double the average price, are not the typical electronic devices on display. This exchange occurs in the clandestine corners of the tech market as a result of U.S. sanctions, which spawned an unadvertised, under-the-table market in China for this coveted chip source.

In September, the U.S. government under President Joe Biden restricted the export of Nvidia’s most advanced chips, the A100 and its more recent sibling, the H100, to mainland China and Hong Kong. Amidst escalating political and trade tensions, this action intended to slow China’s burgeoning AI and supercomputing industries. Despite these limitations, the global AI growth, fueled by achievements such as OpenAI’s ChatGPT, has increased the demand for these high-end chips, especially Nvidia’s microprocessors, renowned for their adeptness at machine learning tasks.

Ivan Lau, the co-founder of Hong Kong’s Pantheon Lab, is one of the individuals navigating this gray market. Lau is endeavoring to acquire 2-4 new A100 cards from non-U.S.-based vendors to run the startup’s most recent AI models. These vendors quote $19,150 per card, making it plain that no warranty or support is included with the sale.

Several vendors in Hong Kong and mainland China have corroborated their ability to acquire small quantities of A100s, demonstrating the soaring demand for these chips in China and the ease with which U.S. sanctions can be circumvented for low-volume transactions.

Who is the market? According to these vendors, app developers, entrepreneurs, researchers, and gamers are among the primary purchasers. Some purchasers are even Chinese government officials. Nvidia maintains, however, that it does not allow exports of the A100 or H100 to China and instead offers alternative products with reduced capabilities that are compliant with U.S. law.

Nvidia asserted that it would take immediate and appropriate action in response to allegations that customers had violated agreements and exported restricted products in violation of the law. The U.S. Department of Commerce also commented on the issue, stating that export control measures have significantly impacted China’s access to high-end processors and that alleged violations are being investigated.

In response to the export restrictions, Nvidia introduced slower variants, the A800 and H800, designed explicitly for the Chinese market. These alternatives mitigate potential revenue losses if Chinese companies choose not to purchase Nvidia’s other products. Due to their financial acumen, tech titans such as Tencent Holdings and Alibaba have purchased these alternatives in large quantities.

The underground vendors acquire the processors by purchasing surplus stock that surfaces after Nvidia’s extensive shipments to major U.S. companies or by importing via local companies in regions such as India, Taiwan, and Singapore. These methods, however, only allow them to obtain small quantities, which are insufficient to build a sophisticated AI large language model from scratch, but sufficient to perform complex machine-learning tasks and improve existing AI models.

Additionally, the shadow market extends to the digital domain. A website for electronic procurement lists numerous A100 vendors, most of whom are located in the Huaqiangbei electronics district. A100 listings are also available on Alibaba’s Taobao e-commerce platform, Xiaohongshu (similar to Instagram), and Douyin, the Chinese equivalent of TikTok.

Despite the flourishing underground market, vendors warn that fraud is prevalent, with refurbished chips frequently misrepresented as A100s. The more difficult-to-obtain, advanced H100 processors from Nvidia are only recently available on the market. Vinci Chow, an economics lecturer at the Chinese University of Hong Kong, mentioned a few packets of eight H100 chips for sale. However, only one of ten vendors validated their ability to acquire H100s.

In the grand scheme of things, the United States may not be unduly concerned about these small-scale chip transactions, according to 86Research analyst Charlie Chai of Shanghai. Nevertheless, he anticipates stricter enforcement if China substantially closes the technology gap. In addition, he predicts that the premium prices charged by Chinese vendors for A100 and H100 processors may fall in the future, as a number of Chinese AI startups that have been driving these purchases may eventually leave the market.

Despite official silence from China’s State Council Information Office and China’s industry ministry and unresponsiveness from companies such as Alibaba, Xiaohongshu, and Douyin’s proprietor ByteDance, the narrative of this covert market continues to develop. It reflects the intricate dance between technology, policy, and market forces in our increasingly interconnected world and is a testament to the insatiable demand for sophisticated AI capabilities.