Huawei has unveiled detailed chipmaking and computing power plans for the first time, setting out a roadmap through 2028 that it says will underpin China’s push for greater technological self-reliance.
The company announced updates to its Ascend chip series and a new generation of supercomputing systems designed to challenge Nvidia’s dominance in artificial intelligence hardware. The disclosure marks a break from years of secrecy around its AI strategy.
Huawei said it will follow its Ascend 910C processor, launched earlier this year, with two versions of the Ascend 950 in 2026. Those will be succeeded by the Ascend 960 in 2027 and the Ascend 970 in 2028. Notably, the company said the 950 will employ its own high-bandwidth memory — a technology China has been working to develop domestically to reduce reliance on overseas suppliers.
On the systems side, Huawei set out plans for two “supernodes” in its Atlas range. The Atlas 950, due in the fourth quarter of 2025, will support up to 8,192 Ascend chips. The Atlas 960, expected in late 2027, will scale that to 15,488 chips. Both will follow the Atlas 900, also known as the CloudMatrix 384, which links 384 of the 910C processors.
The company said the Atlas systems will outperform comparable Nvidia platforms on selected metrics, citing compute capacity, memory size, interconnect bandwidth, and total card count. Independent validation of those claims has not yet been provided.
China has made chip self-sufficiency a policy priority in response to U.S. export controls, which restrict access to leading-edge semiconductors and manufacturing equipment. Huawei has become a focal point of this effort, particularly in artificial intelligence and telecommunications hardware.
Industry reports suggest that Huawei has improved the production yield of its Ascend 910C chips to around 40 percent, up from 20 percent a year earlier. Higher yields lower costs and improve the viability of scaling production, though levels still trail global benchmarks. At the same time, China is reportedly planning new fabrication plants dedicated to AI chips, with one linked to Huawei expected to be operational by the end of this year and two more scheduled for 2026. Ownership structures of those facilities have not been made public.
Huawei’s announcement comes amid competition for market share in AI infrastructure. Nvidia’s GB200 NVL72 system currently sets the industry standard for large-scale model training, supported by its mature software ecosystem. Huawei’s ability to match or surpass that capability will depend not only on hardware performance but also on the maturity of its supporting tools and developer adoption.
The roadmap underscores Huawei’s role in China’s semiconductor drive, but questions remain over the scale at which the company can deliver advanced memory and compute systems under current trade restrictions. Analysts say the next two years will be critical in determining whether its plans can move from prototypes to mass deployment.
