Huawei restricts next-gen AI accelerators to China due to domestic demand
Huawei will not offer its upcoming Ascend 900-series AI accelerators outside China, citing insufficient capacity to meet domestic demand. The company's rotating chairman said there are no plans for a full international expansion, though limited shipments to select countries may occur. The new chips promise significant performance gains over previous generations, with FP4 inference performance reaching up to 28 PFLOPS in future models.
Huawei's rotating chairman Eric Xu stated at the Huawei Connect conference that domestic Chinese demand alone exceeds the company's production capacity for its upcoming Ascend 900-series accelerators. The company will therefore prioritize the Chinese market, with only limited shipments to select countries showing particularly strong demand.
The newly unveiled roadmap shows substantial generational improvements. The Ascend 960DT and 960PR, expected in 2027, will double FP8 training performance to 2 PFLOPS while offering 4 and 8 PFLOPS of FP4 inference respectively. Later models, the Ascend 970 and 980, are projected to reach 14 and 28 PFLOPS of FP4 inference. Despite these gains, the chips remain behind Nvidia's offerings — the Ascend 980's projected 7.2 FP8 PFLOPS trails the R200's 17.5 PFLOPS.
Huawei's decision to keep its accelerators domestic could reshape the global AI hardware market. Companies outside China may face fewer alternatives to Nvidia and AMD, potentially reinforcing their market dominance. Meanwhile, Chinese firms could gain access to competitive hardware despite US export controls, possibly accelerating AI development within China. The performance gap with Nvidia suggests Huawei's chips may serve primarily domestic needs rather than challenging global leadership. Data center operators and AI developers worldwide may need to factor this regional divide into their long-term infrastructure planning.