Sources: Huawei has told some clients that a large share of its next flagship AI chip, the Ascend 960DT, will run mainly on HBM made by Samsung and SK Hynix
Context & Ripple Effects
Huawei’s Ascend program has moved from 910C testing by Chinese internet companies to client discussions around the 910D and a planned increase in 910C output. A 2025 teardown also found Samsung and SK Hynix components in some 910C chips, following earlier HBM stockpiling by Huawei, Baidu, and other Chinese technology companies.
The reported 960DT memory configuration makes HBM supply a central constraint on Huawei’s next flagship design. That matters because Samsung says its HBM capacity is sold out through 2027, while SK Hynix has committed more than half of its 2027 and 2028 output to long-term buyers.
First-order effects
- Huawei’s clients gain a clearer indication that the 960DT’s performance and availability will depend substantially on Samsung and SK Hynix HBM rather than on chip design alone.
- Samsung and SK Hynix face incremental demand tied to Huawei’s Ascend roadmap while much of their future output is already committed under long-term agreements.
Second-order effects
- Huawei’s planned expansion beyond the 910C, including its targeted 2026 Ascend production increase, makes memory allocation a procurement issue alongside fabrication capacity.
- Other HBM buyers face a tighter allocation market if Huawei converts its client guidance into volume orders, reinforcing suppliers’ leverage in contract negotiations.
Third-order effects
- The Ascend line illustrates how a second source of AI compute can still rely on the concentrated HBM supply chain, limiting how independently chip vendors can scale.
- If flagship AI accelerators increasingly compete for the same constrained memory, HBM suppliers—not only accelerator designers—will shape the pace and distribution of AI infrastructure buildouts.
The trend: AI compute is becoming a memory-constrained market, with HBM allocation determining which accelerator platforms can scale.