By Eduardo Baptista
BEIJING, Dec 10 (Reuters) – U.S. President Donald Trump’s move to allow exports to China of Nvidia’s second-best artificial intelligence chip, the H200, will spur strong demand from the country’s tech giants, research institutes, and its defence-industrial complex.
Beijing has not yet confirmed whether it will allow the chip to be sold in China, but a Reuters review of more than 100 tenders and academic papers shows it is already being supplied to domestic buyers via the grey market.
The analysis shows the nature of customers in China who will jump at the chance for legal bulk buys of the H200 chip, which is many times more powerful than any chip Nvidia is allowed to sell there.
ELITE UNIVERSITIES AND RESEARCH INSTITUTES
China’s leading universities all have research teams focused on AI development, and the number of high-end chips they have at their disposal directly affects talent recruitment and research.
One professor at Beijing Jiaotong University advertises that his laboratory owns eight H200 chips, allowing for AI model research.
Researchers at the state-backed Shanghai Artificial Intelligence Laboratory and the Sun Yat-Sen, Tsinghua, and Shanghai Jiao Tong universities used four Nvidia H200 processors to train an AI model designed to detect whether an image is AI-generated, they showed in a paper published last month.
In June, a state-run AI institute in the eastern city of Hefei issued a a tender for a server equipped with eight Nvidia H200 chips to power a “quantum AI model” project.
Dozens of universities and research institutes nationwide have bought or sought to acquire H200 chips, the Reuters review showed.
CHINESE MILITARY AND AFFILIATED ENTITIES
China hawks in Washington have balked at Trump’s moves to reverse the previous administration’s export controls, saying the Chinese military would use Nvidia’s most advanced AI chips to turbocharge its capabilities.
The Reuters review suggests H200 chips are already making their way into the hands of the People’s Liberation Army and closely-linked universities.
In August, the PLA Air Force Medical University in Xian issued a tender for eight Nvidia H200 chips to train a large-language model training platform to support medical AI and biosurveillance research.
On Monday, the School of Cyberspace Security at Beihang University, one of China’s “Seven Sons”, or universities under U.S. sanctions with a history of defence-related research, issued a tender seeking a supplier that could rent out H200-level computing power.
Chinese entities increasingly rent usage time on a server fitted with banned Nvidia chips as a way to access restricted hardware without importing it.
AI DATA CENTER OPERATORS AND BUILDERS
Even before Trump’s announcement, some Chinese AI cluster and data center projects had set their eyes on large quantities of the H200 chip.
In the eastern province of Jiangsu, a firm owned by the Binhai county government issued a tender in July for 48 servers equipped with 384 H200 chips, with deliveries due by year-end.
Larger ambitions are visible in the far western region of Xinjiang, which has seen a massive AI infrastructure buildout in recent years as Chinese companies and authorities seek the cheapest land and electricity for huge data centre projects.
A June 6 tender by Urumqi Jiangsuan outlines a plan for a 20,000-petaflop hub combining more than 8,000 H200 GPUs, 12,000 RTX 4090 GPUs and 4,500 servers fitted with Huawei Ascend 910C processors, the most powerful domestic AI chip now available.
A separate 1.86-billion-yuan project in Burqin county in northern Xinjiang, unveiled in October 2024, details a green-energy compute centre dominated by 1,000 domestic chip servers but supplemented by a smaller H100 or H200 chip cluster of 100 servers.
The H200’s less powerful predecessor, the H100 has been banned from export to China since late 2022.
In the central province of Hubei, Xiaogan Yunqi Data Technology submitted a regulatory filing in October for a computing power project worth 307 million yuan, to deploy 128 H200 servers for telecoms giant China Unicom by next March.
(Reporting by Eduardo Baptista; Editing by Clarence Fernandez)
Comments