According to a review of tender documents by Reuters, Chinese military bodies, state-run artificial intelligence research institutes, and universities have procured small quantities of Nvidia semiconductors that were banned by the U.S. from being exported to China.
This highlights the challenges faced by Washington in completely restricting China’s access to advanced U.S. chips, which could potentially contribute to advancements in AI and sophisticated military computers.
It is important to note that the purchase or sale of high-end U.S. chips is not illegal in China. The publicly available tender documents reveal that numerous Chinese entities have acquired and received Nvidia semiconductors since the implementation of restrictions.
These purchases include Nvidia’s A100 and the more powerful H100 chips, which were banned from export to China and Hong Kong in September 2022.
Additionally, the slower A800 and H800 chips, which Nvidia specifically developed for the Chinese market, were also banned in October of the same year.
Nvidia’s graphic processing units (GPUs), which are a type of chip, are widely recognized as being significantly better than competing products for AI tasks due to their ability to efficiently process large amounts of data required for machine learning.
This is evident in the continued demand for and availability of banned Nvidia chips, highlighting the lack of viable alternatives for Chinese companies despite the emergence of rival products from Huawei and others.
Nvidia Faces Challenges in China as 90% AI Chip Market Share
Prior to the bans, Nvidia held a dominant 90% share of China’s AI chip market.
Among the purchasers of these banned Nvidia chips were prestigious universities and entities subject to U.S. export restrictions, such as the Harbin Institute of Technology and the University of Electronic Science and Technology of China.
These entities have faced accusations of involvement in military affairs or affiliation with military bodies contrary to U.S. national interests.
The former bought six Nvidia A100 chips in May for the purpose of training a deep-learning model, while the latter purchased one A100 chip in December 2022, with no specified purpose.
Despite attempts to seek comments from the purchasers mentioned in the article, none of them responded.
Additionally, a Reuters review revealed that neither Nvidia nor approved retailers were identified as suppliers, raising questions about how these suppliers obtained their Nvidia chips.
Following the imposition of U.S. restrictions, an underground market for these chips has emerged in China. Chinese vendors have previously claimed to acquire excess stock that becomes available after Nvidia supplies large quantities to major U.S. companies.
They also import chips through locally incorporated companies in countries like India, Taiwan, and Singapore.
Reuters attempted to obtain comments from 10 suppliers listed in tender documents, including those mentioned in the article, but received no responses.
Nvidia has stated that it adheres to all relevant export control laws and expects its customers to do the same. The company spokesperson emphasized that if they discover any customer engaging in illegal resale to third parties, they will promptly take appropriate action.
The U.S. Department of Commerce has chosen not to provide any comment on the matter.
In an effort to tighten export restrictions, U.S. authorities have taken steps to restrict access to chips by Chinese companies located outside of China, aiming to impede China’s AI development.
However, Chris Miller, a professor at Tufts University and author of “Chip War: The Fight for the World’s Most Critical Technology,” believes that achieving completely foolproof export restrictions is impractical due to the small size of chips, which makes them easily smuggled.
Ongoing Chip Acquisitions Despite Quantity Limitations
The review encompasses over 100 tenders in which state entities have acquired A100 chips, as well as numerous tenders since the ban in October that reveal purchases of the A800.
Tenders released last month also indicate that Tsinghua University procured two H100 chips, while a laboratory operated by the Ministry of Industry and Information Technology procured one.
According to military tenders from a database, one unnamed entity of the People’s Liberation Army based in Wuxi, Jiangsu province sought three A100 chips in October and one H100 chip this month.
However, due to heavy redaction in Chinese military tenders, it remains unknown who won the bids or the purpose behind these purchases.
Although most tenders specify that the chips are being utilized for AI, the quantities procured are relatively small and insufficient for constructing a sophisticated AI large language model from scratch.
Research firm TrendForce estimates that a model similar to OpenAI’s GPT would necessitate over 30,000 Nvidia A100 cards. Nonetheless, a few chips can still execute complex machine-learning tasks and enhance existing AI models.
For instance, the Shandong Artificial Intelligence Institute awarded a contract worth 290,000 yuan ($40,500) to Shandong Chengxiang Electronic Technology for five A100 chips last month.
Many tenders require suppliers to deliver and install the products before receiving payment, and most universities have published notices confirming the completion of these transactions.
Tsinghua University, often referred to as China’s Massachusetts Institute of Technology, has been an active participant in tenders, having procured approximately 80 A100 chips since the ban in 2022.
In December, Chongqing University issued a tender for one A100 chip, explicitly stating that it must be brand new and not second-hand or disassembled. A notice confirmed that the delivery was finalized this month.