Huawei is set to expand Ascend AI chip output, seizing on Nvidia's shrinking position in China as US export curbs and Beijing's security mandates squeeze the American company's market access. According to Bloomberg, Huawei plans to produce about 600,000 Ascend 910C chips in 2026 — more than twice its 2025 volume — with total Ascend output climbing to 1.6 million dies, a sign that Huawei and partner SMIC are alleviating capacity constraints.
AI players, including DeepSeek and Alibaba, are driving demand, yet Nvidia's shipments have slumped. The company sold more than one million H20 chips in China in 2024 but logged no sales in the second quarter of 2025, hit by US export controls and Beijing's push to sideline its products.
Huawei has laid out a three-year roadmap through 2028, with phased launches of the Ascend 950, 960, and 970. The company is also expected to debut the 950DT by late 2026, a four-die package with a planned run of 100,000 units.
Bernstein research underscored the performance gap: the Ascend 950 is expected to reach just 6% of Nvidia's upcoming VR200, while Huawei still depends on SMIC's improved 7nm node. Nvidia, by comparison, manufactures on TSMC's 4nm process, leaving Huawei trailing by at least two generations.
Huawei is still investing heavily in AI chip R&D, including upgraded 7nm designs for model training. Commercial rollout may not come until after 2027, but the company is banking on scale and architecture to counter weaker process technology.
Huawei rotating chairman Eric Xu unveiled the UnifiedBus interconnect, enabling up to 15,488 Ascend chips to operate as one cluster, an approach designed to offset limited single-chip performance with massive parallelism.
Analysts say Huawei's aggressive chip R&D and capacity plans sharpen its rivalry with Nvidia while reinforcing China's broader drive for semiconductor self-sufficiency amid the intensifying US-China tech standoff.
Article translated by Levi Li and edited by Jack Wu
