– Click the card above to follow me-
Get free access to a collection of large model learning materials compiled over the past year
A new report from SemiAnalysis indicates that while Huawei is producing hundreds of thousands of AI chips, the real bottleneck is not chip manufacturing, but the shortage of High Bandwidth Memory (HBM).
As the United States attempts to maintain its advantage in artificial intelligence by restricting China’s access to computing power, China is pushing for complete technological independence, with Huawei as the national champion. However, SemiAnalysis believes that Huawei’s surge in AI chips is built on a shaky foundation. The real issue is not the manufacturing of logic chips, but obtaining sufficient HBM.
Huawei’s Production Relies on Stockpiled Chips
SemiAnalysis estimates that Huawei will ship approximately 507,000 Ascend AI accelerators in 2024, with an expected shipment of 805,000 units in 2025. However, the report states that these figures are misleading because Huawei has stockpiled over 2.9 million Ascend chips manufactured by TSMC to circumvent export controls. This stockpile is driving production in 2024 and 2025, but analysts expect it to be depleted within nine months.
Meanwhile, Huawei is heavily investing in vertical integration to control its supply chain. The company has launched a chip manufacturing tool company called SiCarrier and is expanding its own factories. By next year, these factories may rival or even exceed the capacity of China’s largest contract chip manufacturer, SMIC.
SMIC is No Longer the Major Bottleneck
Huawei currently outsources chip production to SMIC, which has been a limiting factor. However, SemiAnalysis expects this situation to change by the end of the year. SMIC is expanding its advanced process capacity (7nm and below), aiming to produce 45,000 wafers per month by the end of 2025 and 80,000 wafers per month by 2027. Even a portion of this capacity would be sufficient to utilize millions of Ascend chips annually.
SMIC’s focus on AI chips has become possible partly because other Chinese companies are still using TSMC to produce mobile chips, alleviating pressure. Analysis also indicates that Huawei continues to obtain advanced data center CPUs and network components manufactured by TSMC through shell companies. Another factor is the slow adoption of U.S. regulatory measures by Japan and the Netherlands, allowing Chinese companies to stockpile equipment.
HBM is the Real Fatal Weakness
SemiAnalysis points out that HBM supply is the true weakness of China’s AI efforts. Notably, Beijing is pushing for the relaxation of HBM restrictions in trade negotiations, rather than those on lithography tools or access to TSMC.
China is attempting to gain an edge by stockpiling 13 million HBM stacks, most of which (11.4 million) come from Samsung. Reports indicate that Samsung alone shipped 7 million stacks within a month of announcing this news, just before the U.S. began implementing restrictions on HBM2E in December 2024. This is enough to produce approximately 1.6 million Ascend 910C chips, but SemiAnalysis expects the stock to be depleted by the end of the year. After that, without foreign HBM, Huawei may not be able to produce even 1 million Ascend chips in 2025, as domestic supply cannot fill this gap.
China’s largest DRAM manufacturer, CXMT, is rapidly catching up but cannot meet demand. SemiAnalysis estimates that CXMT will produce only 2 million HBM stacks next year, enough to produce 250,000 to 300,000 Ascend 910C chips. This highlights the effectiveness of HBM sanctions, and analysts suggest adding CXMT to the sanctions list as well.
NVIDIA’s Role and Computational Diplomacy
The U.S. government’s decision to allow NVIDIA to export its H20 chips to China is part of a complex strategy. SemiAnalysis suspects that the Chinese public’s lack of interest in the H20 is a tactic to gain approval for more powerful Blackwell series chips (B30A). Reports of models like Deepseek performing poorly on Huawei hardware highlight that Chinese companies still rely on Western chips.
The U.S. faces a strategic dilemma: selling chips ties China closely to U.S. technology, slowing the development of domestic alternatives, but also accelerates China’s AI advancements. The report suggests that stronger chips should only be approved when China can manufacture its own comparable chips at scale. Until then, Chinese companies like ByteDance must rely on cloud providers in places like Malaysia to access top GPUs—access that the U.S. can cut off at any time. Analysts estimate that for NVIDIA, selling improved chips to China could generate up to $30 billion in revenue annually.