High Bandwidth Memory (HBM) supply constraints are driving up the cost of Chinese-made AI accelerators, as domestic vendors like ChangXin Memory (CXMT) face component scarcity.
The reported cost pressure on Chinese AI accelerators is coming not from the headline processor itself, but from the component that sits beside it: high-bandwidth memory, or HBM. As AI data centers demand more advanced memory and access to certain HBM products faces restrictions, the economics of China's alternatives to Nvidia are becoming harder to manage.
The pressure has surfaced with Huawei's Ascend 950DT and the forthcoming Cambricon 690. While product-level price figures remain reported rather than official public price lists, the broader conclusion is clear: designing an accelerator is only part of the challenge. The memory that feeds it can become the expensive, difficult-to-source half of the system.
HBM is a form of DRAM built from vertically stacked memory dies and placed close to an accelerator. That short physical path helps move large volumes of data quickly during AI model training and inference—exactly the kind of workload where data movement becomes as important as raw compute. An accelerator without enough fast memory is like a powerful engine waiting for fuel. The processor may be designed domestically, but the surrounding memory technology still determines how efficiently the complete system can be supplied and operated.
The HBM market is unusually concentrated. In Q2 2026, SK hynix accounted for 50% of global HBM revenue, Samsung for 33%, and Micron for 18%. The distinction from broader DRAM markets matters: CXMT holds 10% of global DRAM revenue, but this does not establish a comparable position in HBM, where the market is dominated by the three foreign suppliers.
On December 2, 2024, the U.S. Bureau of Industry and Security announced controls covering certain HBM products, including U.S.-origin HBM and certain foreign-produced HBM subject to U.S. export rules, as part of a wider semiconductor-control package aimed at limiting China's ability to produce advanced technologies for military applications. The scope depends on the products and rules involved, but restrictions matter in a market where a small group of suppliers accounts for most measured HBM revenue. At the same time, AI data-center demand is pulling memory toward large-scale training and inference systems. Production decisions, conventional-memory demand, and export controls add pressure from different directions. The situation is more precisely described as a supply-and-demand squeeze centered on specialized memory rather than a simple "AI chip shortage."
CXMT has reportedly begun producing HBM3E in small quantities, and Alibaba's T-Head and Cambricon have reportedly tested it with their processors. That is an important development for China's semiconductor ambitions, but it is not the same as demonstrating large-scale, commercially competitive HBM production. The available public picture does not establish CXMT's comparable production scale, yield, customer volume, or performance. Nor does the reported activity establish that the memory technologies associated with Huawei's Ascend 950PR and Ascend 950DT—HiBL 1.0 and HiZQ 2.0, respectively—are manufactured domestically. A first domestic supply channel can reduce dependence at the margin, but it does not instantly replace the capacity, process maturity, and customer validation of the dominant suppliers.
China can develop its own accelerator designs and still face a bottleneck in the memory subsystem. HBM must sit close to the processor, move data at high speed, and arrive in sufficient volume for the complete accelerator platform to make economic sense. A constraint in any one of those links affects the finished system. That is why the reported pressure around Huawei Ascend 950DT and Cambricon 690 matters beyond two product names—it exposes a dependency that cannot be solved merely by producing another processor. Until domestic HBM reaches a scale and performance that can support those systems consistently, China's AI-chip strategy remains tied to a concentrated global memory supply chain. That dependency is now showing up in cost.