Huawei shelves global AI chip rollout as domestic demand outstrips supply

Huawei this week outlined its AI accelerator roadmap, revealing significant performance gains for the upcoming Ascend 960, 970, and 980 neural processing units (NPUs) over the existing Ascend 910C and 950-series. The Ascend 960DT and 960PR, targeted for 2027, are expected to deliver FP8 training performance of 2 PFLOPS and FP4 inference performance of 4 PFLOPS and 8 PFLOPS, respectively. Successors Ascend 970 and Ascend 980 are projected to reach FP4 performance of 14 PFLOPS and 28 PFLOPS in subsequent years.

Speaking on the sidelines of the Huawei Connect conference, rotating chairman Eric Xu told Reuters: “Since we do not have enough capacity to even satisfy the demand in China, we do not have a plan to expand into the international market in a fully-fledged way.” He added that Huawei supplies limited volumes to “some countries where demand is particularly strong,” without elaborating.

The company’s Atlas clusters, which can scale to 120 EFLOPS, use optical networking to link up to 15,488 chips, positioning Huawei to compete with Nvidia’s GPU-based systems. However, the capacity constraints mean that even as Huawei’s Ascend 960-series may rival offerings from Nvidia and AMD, the vast majority of the world will remain reliant on Western suppliers for the foreseeable future.

Why This Matters

The decision highlights the extreme shortage of AI chip production capacity in China, even as Huawei advances its technology. This reinforces Nvidia’s dominance outside China.
Global customers seeking alternatives to Nvidia’s AI accelerators will have fewer options, potentially slowing AI infrastructure buildout in markets that are wary of supply concentration.
The move signals that US export controls and domestic demand are reshaping the global AI chip landscape, forcing Huawei to focus on the Chinese market first.

Background

Huawei’s chip development has been under pressure since US sanctions cut off access to advanced fabrication and EDA tools. The company has since pivoted to designing its own AI accelerators, the Ascend series, and building a domestic ecosystem to rival Nvidia’s CUDA platform. While Huawei has made progress in performance, manufacturing capacity—likely constrained by reliance on Chinese foundries like SMIC—remains a bottleneck. The company’s decision to prioritize domestic demand reflects both the surging need for AI compute in China and the difficulty of scaling production.

Key Perspectives

Huawei: Seeks to maximize impact by supplying the largest market (China) first, while making limited exceptions for strategically important countries. The company emphasizes that is a capacity issue, not a product quality issue. Nvidia and AMD: Benefit from Huawei’s absence in international markets. However, Huawei’s rising performance in China could eventually challenge their dominance if capacity expands. Global buyers: Companies outside China may find it harder to diversify their AI chip supply, potentially leading to higher prices and longer lead times for Nvidia’s products.

What to Watch

Whether Huawei announces plans to build additional fabs or partners with other foundries to increase capacity.
Independent benchmarks of Ascend 960-series performance against Nvidia’s B300 and AMD’s MI400 series.
US export control updates that could further tighten access to equipment needed for Huawei’s chip production.

ZOTPAPER