Despite all the efforts that the U.S. government has put into restricting the access of Chinese entities to high-performance AI hardware, there are enough AI processors in China to build the world's most powerful training cluster, reports SemiAnalysis. Huawei, which is probably the most sanctioned company in the IT sector, has not only circumvented U.S. sanctions by using proxies, but it produces plenty of its own Ascend 910-series processors domestically, the report says.
It's no secret that even Nvidia's highest-end AI and HPC GPUs like the H100 and H200 are being smuggled into China from other countries, including India, Malaysia, and Singapore. As a result, entities that need to build a server or even a rack powered by Nvidia's Hopper GPUs can do so, albeit at a premium price. If these entities don't want to smuggle GPUs illegally, they can also acquire Nvidia's cut-down DGX H20 processors aimed at China and shipped to the country without restrictions.
SemiAnalysis estimates that Nvidia will officially sell 900,000 HGX H20 processors to China this year alone, and will then supply over a million B20 processors to the country in 2025. How many 're-exported' H100/H200 processors will end up in China is unknown, but the analysts believe that 'many' will make it into the country. Chinese companies would of course like to have more of Nvidia's high-end AI processors, but they can also access them in the cloud.
Huawei may not be exactly interested in getting Nvidia's hardware, as it has its own Ascend 900-series hardware platform for AI. The company's Ascend 910 processor was originally made by TSMC (and it looks like eventually Huawei managed to continue making the compute chiplet of this design at TSMC via a proxy), and it also has the Ascend 910B made by SMIC using its N+2 manufacturing technology (7nm-class), along with the Ascend 910C made on SMIC's N+3 fabrication technology (6nm-class).
How can SMIC keep producing chips on advanced process technologies two years after the U.S. imposed tight restrictions on sales of advanced wafer fab tools to China's most advanced fabs? Apparently, these curbs are easily circumvented by formally supplying sophisticated wafer fab equipment to fabs that are not restricted. SMIC, for example, has an unrestricted legacy fab and a restricted advanced logic fab interconnected by a wafer bridge that enables wafers to travel back and forth, essentially making the two fabs into a single facility. This setup enables one fab to continue producing advanced chips for AI applications, like the Kirin 9000S and Ascend 910B, while the other fab, classified as a legacy fab, can legally import advanced tools.
Huawei has joined forces with SMIC and CXMT in a comprehensive, state-backed effort to build a domestic fabrication network that could produce everything needed for an AI hardware supply chain, starting from advanced logic and HBM memory, to advanced packaging and silicon photonics interconnects.
Huawei plans to spend a substantial $7.3 billion on foreign wafer fabrication equipment in 2024 alone, including tools from the U.S., Japan, and Europe. If SMIC and CXMT's purchases are included, China would rank second globally in equipment spending, behind only Taiwan's TSMC.
This scale of investment highlights weaknesses in current U.S. export controls, which rely on entity lists that SMIC and Huawei can evade through technical distinctions. Analysts suggest revising the Foreign Direct Product Rule to cover equipment with any American content rather than only those meeting the 25% threshold, thus tightening control over tech exports.
The pushback from U.S. fab tools suppliers complicates this strategy. Many suppliers argue that tighter controls harm their businesses, lobbying for relaxed regulations. However, analysts warn that easing restrictions could undermine U.S. technology leadership and national security in the long term, as these sanctions workarounds are giving China an accelerated path toward AI and semiconductor self-sufficiency.
