HONG KONG/SHENZHEN, China, June 20 (Reuters) – Psst! Where can a Chinese buyer buy top-end Nvidia ( NVDA.O ) AI chips in the wake of U.S. sanctions?
Visiting the famous Huaqiangbei electronics district in the southern Chinese city of Shenzhen is a good bet – especially the SEG Plaza skyscraper whose first 1[ads1]0 floors are packed with shops selling everything from camera parts to drones. The chips are not announced, but it works discreetly to ask.
They don’t come cheap. Two suppliers there, who spoke to Reuters personally on condition of anonymity, said they could provide a small number of A100 artificial intelligence chips made by the American chip designer, and priced them at $20,000 apiece – double the usual price.
While it is not illegal to buy or sell high-end US chips in China, US export restrictions have created a de facto underground market with vendors keen not to draw scrutiny from either US or Chinese authorities.
President Joe Biden’s administration in September ordered Nvidia to stop exporting its two most advanced chips — the A100 and the newly developed H100 — to mainland China and Hong Kong, part of efforts to thwart Chinese AI and supercomputer development amid growing political and trade tensions. It was then followed up with a number of semiconductor-related export controls.
But as AI grows across the globe following the runaway success of OpenAI’s ChatGPT, demand for high-end chips has skyrocketed, especially for Nvidia’s microprocessors, which are considered the best at handling machine learning tasks.
“We’re talking to two vendors now to get some,” said Ivan Lau, co-founder of Hong Kong’s Pantheon Lab which is trying to buy 2-4 new A100 boards to run the startup’s latest AI models.
Those vendors, who bought the chips outside the US, quoted HK$150,000 ($19,150) per card, he said, adding: “They told us straight up that there will be no warranty or support.”
Reuters spoke to 10 suppliers in Hong Kong and mainland China who described being able to easily source a small number of A100s. Their information highlighted both the intense demand in China for the chips and the relative ease with which Washington’s sanctions can be circumvented for small-batch transactions.
Reuters was unable to estimate total volumes of Nvidia A100 and H100 chips flowing into China or determine the extent to which the transactions taking place are moving towards meeting demand.
Buyers are typically app developers, startups, researchers or gamers, the suppliers said, declining to be identified because the imports run afoul of U.S. trade restrictions. One supplier said buyers also included Chinese local governments.
Nvidia said in a statement to Reuters that it did not allow exports of the A100 or H100 to China, instead providing replacements with reduced capacity that comply with US law.
“If we receive information that a customer is violating the agreement with us and exporting products with restrictions in violation of the law, we will take immediate and appropriate measures,” the statement said.
A US Commerce Department spokesperson said in a statement to Reuters that export control measures have had a “significant impact” on China’s availability of advanced chips.
The statement also said that reports of parties attempting to obtain these chips through illegal means were “not a surprise”, adding that “allegations of violations are being investigated”.
China’s State Council Information Office and China’s Ministry of Industry did not respond to requests for comment.
Nvidia said in September that $400 million in third-quarter sales could be lost if Chinese firms decided not to buy alternative Nvidia products.
The new China-tailored slower variants – the A800 and H800 – developed to mitigate this effect are now being bought by major Chinese technology firms such as Tencent Holdings ( 0700.HK ) and Alibaba ( 9988.HK ), which have the deep pockets to buy huge quantities.
The Chinese suppliers said they primarily acquired the chips in two ways: snapping up excess inventory that finds its way into the market after Nvidia ships large quantities to large U.S. firms, or importing through companies locally incorporated in places like India, Taiwan and Singapore .
This means that the quantities they can secure are small, far from what is needed to build a sophisticated AI large-language model from scratch.
A model similar to OpenAI’s GPT would require more than 30,000 Nvidia A100 cards, according to research firm TrendForce. But a handful can run complex machine learning tasks and improve existing AI models.
According to an electronics sourcing website that listed around 40 sellers of A100s, most were located in the Huaqiangbei electronics area. But listings for the A100s can also be found on Alibaba’s ( 9988.HK ) Taobao e-commerce site, on Instagram-like Xiaohongshu, as well as on Douyin, the Chinese version of TikTok.
Alibaba, Xiaohongshu and Douyin owner ByteDance did not respond to requests for comment.
Some of the suppliers warned that fraud had become common with refurbished chips being labeled as A100s.
Nvidia’s more advanced H100 chips, only on the market since March, seem to be much harder to come by.
Vinci Chow, a lecturer in economics at the Chinese University of Hong Kong whose department has acquired four A100 boards from local suppliers for research purposes, said he had been told that some packs of eight H100 chips were available for purchase. But only one of the 10 suppliers Reuters spoke to said they could procure H100s.
The U.S. is probably not too keen on small transactions of the chips, said Charlie Chai, a Shanghai-based analyst at 86Research.
“Only if/when China poses a greater threat after significant acquisitions will we see stricter enforcement,” he said.
He added that the premiums currently commanded by Chinese vendors for A100 and H100 chips could collapse in the future as many of the Chinese AI startups that led to purchases would eventually withdraw from the market.
($1 = 7.8307 Hong Kong dollars)
Reporting by Josh Ye in Hong Kong, David Kirton in Shenzhen and Chen Lin in Singapore; Additional reporting by Fanny Potkin in Singapore and David Shepardson in Washington; Editing by Brenda Goh and Edwina Gibbs
Our standards: Thomson Reuters Trust Principles.