China’s AI Accelerator Industry Is ‘Nothing’ Without Access to Foreign HBM, as Domestic Firms Heavily Rely on Pre-Export Control Stockpiles

Muhammad Zuhair
Huawei Preps Ascend 910C To Tackle NVIDIA's H100 In China's Domestic AI Market 1
Image Credits: Huawei

China's main hurdle towards scaling up AI chip production isn't just capacity constraints from the likes of SMIC, but a lack of domestic HBM technologies has a much more significant role to play.

China Is Far From Achieving Sufficient HBM Production, But This Could Change If US Export Controls Aren't Ramped Up

Beijing is rapidly enabling domestic firms to switch towards in-house tech stacks such as those from Huawei and Cambricon, but the real question is whether the nation has enough resources to cater to the enormous demand. Usually, when we talk about the constraints China faces in scaling up compute capabilities, the lack of semiconductor production certainly comes to mind, but according to a report by SemiAnalysis, reliance on foreign HBM has played a massive role over the span of the AI hype.

Related Story NVIDIA RTX 6000D “Blackwell Pro For China” Tested: Features 17% Fewer Cores, 14% Less VRAM, & Lower Clocks

The report claims that China is facing an 'HBM bottleneck', and the firm is heavily dependent on the stockpile of HBM, which it managed to accumulate before the US export controls. One of the significant beneficiaries of China's HBM demands was Samsung, as the Korean giant reportedly shipped out 11.4 million stacks of HBM to China, which is a massive chunk of the nation's inventory right now. While the demand for foreign HBM might be catered to through 'grey channels', it is evident that the flow of HBM from other nations to China has reduced dramatically.

Huawei can easily produce 805,000 units of the Ascend 910C chips based on capacity from TSMC and SMIC, but the nation doesn't have enough HBM to cater to the production request. This shows that the memory type has restricted Beijing's efforts to expand its influence in the AI compute market. Without sufficient HBM production, China's domestic AI chip efforts are of no use, and Western manufacturers like NVIDIA/AMD will have a huge advantage over Huawei.

In terms of China's domestic efforts to scale up HBM production, firms like CXMT face a huge constraint regarding the equipment required to convert standard DRAM into HBM technologies, which is why Beijing is pushing for relaxation in this segment. However, considering the investments being poured into the domestic HBM sector, coupled with the rather 'unspecific' sanctions related to this segment, China could scale up to HBM3E by 2026, as predicted by SemiAnalysis, which means that the HBM bottleneck will be short-lived if action isn't taken.

It would be interesting to see how China's AI chip industry shapes up under US export controls, considering that domestic firms are ramping up efforts to reduce their reliance on Western alternatives. However, they have yet to create a suitable supply chain to cater to the demand of the Chinese AI industry.

Follow Wccftech on Google to get more of our news coverage in your feeds.

Button