The reported cost pressure on Chinese AI accelerators is coming from a component that sits beside the processor, not inside the headline chip itself: high-bandwidth memory, or HBM. As AI data centers absorb more advanced memory and access to some HBM products is restricted, the economics of China’s alternatives to Nvidia are becoming harder to manage.
The pressure has been associated with Huawei’s Ascend 950DT and the forthcoming Cambricon 690. Product-level price figures remain reported rather than official public price lists, so the useful conclusion is broader: 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 the memory beside the AI chip
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 that makes data movement 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.
Why the memory market matters
The measured HBM market is unusually concentrated. In Q2 2026, SK hynix accounted for 50% of global HBM revenue, Samsung for 33%, and Micron for 18%. Those rounded figures add up to 101%, so they should be read as market estimates rather than an exact allocation.
| Company | Global HBM revenue share, Q2 2026 | Global DRAM revenue share, Q2 2026 | What the figure measures |
| SK hynix | 50% | 25% | HBM revenue and total DRAM revenue, respectively |
| Samsung | 33% | 38% | HBM revenue and total DRAM revenue, respectively |
| Micron | 18% | 24% | HBM revenue and total DRAM revenue, respectively |
| CXMT | Not reported | 10% | Total DRAM revenue only; this is not an HBM share |
The distinction matters. CXMT’s 10% figure describes global DRAM revenue, not its position in HBM. Treating the two numbers as interchangeable would make China’s memory position look more established in HBM than the data supports.
Export controls meet AI-driven demand
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. The action was part of a wider semiconductor-control package aimed at limiting China’s ability to produce advanced technologies for military applications.
That is not a blanket prohibition on every memory product or every HBM transaction. The scope depends on the products and rules involved. Still, 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. Calling the situation simply an “AI chip shortage” misses the more precise problem: a supply-and-demand squeeze centered on a specialized memory component.
China’s domestic alternative is still at an early stage
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.
The practical takeaway is straightforward: 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.
The strategic consequence for Chinese AI accelerators
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—and that dependency is now showing up in cost.
