A recent report has revealed that Chinese AI companies, including Huawei, are now testing domestically produced High-Bandwidth Memory (HBM), a critical component for AI chips.
This move is a direct response to U.S. export controls. In December 2024, the U.S. government specifically targeted HBM, effectively turning it into a 'chokepoint' to slow down China's AI progress. By developing its own HBM, China is attempting to bypass this critical bottleneck.
The groundwork for this development was laid over several months. First, back in September 2025, Huawei publicly announced its Ascend AI chip roadmap, which explicitly included plans for “in-house HBM.” Soon after, by late 2025, reports emerged that Chinese memory maker CXMT was already supplying HBM3 samples to partners like Huawei. These early steps showed a clear, strategic intent to achieve self-sufficiency.
Second, momentum accelerated significantly in early 2026. In March, Huawei launched its Atlas 350 AI accelerator, which was presented with up to 112 GB of HBM, providing tangible proof that their integration efforts were working. This wasn't just a lab experiment anymore; it was part of a new product. This progress, combined with financial stability and growing investor confidence in companies like Iluvatar, created the perfect environment for this breakthrough.
Interestingly, the market's reaction has been measured. Following the news, the stock prices of leading HBM manufacturers like SK hynix, Samsung, and Micron actually increased. This suggests that investors believe China's path to mass-producing high-quality HBM will be challenging and gradual. Global demand for AI chips is so high that established players are expected to retain their market dominance for the time being.
In conclusion, this isn't about China instantly solving its semiconductor challenges. Rather, it represents a significant erosion of the U.S. chokepoint strategy. The question is no longer whether China can develop its own HBM, but rather how quickly it can achieve high-volume production with good manufacturing yields.
- HBM (High-Bandwidth Memory): A type of high-performance memory essential for AI processors. It stacks memory chips vertically to achieve much faster data transfer speeds than conventional memory.
- Chokepoint: A strategic point of congestion or blockage. In geopolitics, it refers to controlling access to a critical resource or technology to influence another nation.
- Yield: In manufacturing, this refers to the percentage of non-defective products produced in a batch. A high yield is crucial for cost-effective mass production.