Despite all the efforts the US government has made to restrict Chinese entities’ access to high-end AI hardware, there are enough AI processors in China to build the most powerful training cluster in the world, SemiAnalysis reports. Huawei, probably the most sanctioned company in the IT sector, has not only circumvented US sanctions by using proxies but also produces many of its own Ascend 910 processors domestically, the report said.
It’s no secret that even Nvidia’s top AI and HPC GPUs, like the H100 and H200, are 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 higher price. If these entities don’t want to illegally smuggle GPUs, they can also acquire Nvidia’s China-targeted DGX H20 processors, which are shipped to the country without restrictions.
SemiAnalysis estimates that Nvidia will officially sell 900,000 HGX H20 processors to China this year alone, and then ship well over a million B20 processors to the country by 2025. It is unknown how many ‘re-exported’ H100/H200 processors will end up in China. , but analysts believe that ‘many’ will reach the country. Chinese companies would obviously like to have more of Nvidia’s high-end AI processors, but they can also access them in the cloud.
Huawei may not be particularly interested in 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 The company’s Ascend 910 processor was originally made by TSMC (and it appears that Huawei ultimately managed to continue making this design’s computer chiplet through a proxy at TSMC), and there was also an Ascend 910B made by SMIC using its N+2 manufacturing technology (7nm class), together with the Ascend 910C made based on SMIC’s N+3 manufacturing technology (6nm class).
How can SMIC continue to produce chips based on advanced process technologies, two years after the US imposed strict restrictions on the sale of advanced wafer fabrication tools to China’s most advanced factories? Apparently, these restrictions can easily be circumvented by formally supplying advanced wafer fabrication equipment to unrestricted factories. For example, SMIC has an unlimited legacy factory and a limited advanced logic factory connected by a wafer bridge that allows wafers to travel back and forth, effectively turning the two factories into a single facility. This setup allows one factory to continue producing advanced chips for AI applications, such as the Kirin 9000S and Ascend 910B, while the other factory, classified as an older factory, can legally import advanced tools.
Huawei has joined forces with SMIC and CXMT in a comprehensive, state-backed effort to build a domestic manufacturing network that could produce everything needed for an AI hardware supply chain, ranging from advanced logic and HBM memory to advanced packaging and silicon photonics interconnects.
Huawei plans to spend a significant $7.3 billion on foreign wafer production equipment in 2024 alone, including tools from the US, Japan and Europe. If purchases of SMIC and CXMT are included, China would rank second globally in equipment spending, behind only Taiwan’s TSMC.
This size of investment points to weaknesses in current US export controls, which rely on entity lists that SMIC and Huawei can circumvent through technical differences. Analysts propose revising the Foreign Direct Product Rule to cover equipment with U.S. content instead of just equipment that meets the 25% threshold, tightening control over technology exports.
The setback from U.S. fab tool suppliers complicates this strategy. Many suppliers claim that stricter controls are harmful to their businesses and are lobbying for relaxed regulations. However, analysts warn that easing restrictions could undermine U.S. technological leadership and national security in the long term, as these sanctions give China an accelerated path to self-sufficiency in AI and semiconductors.
