Huawei Speeds AI Chip Roadmap as It Pushes to Challenge NVIDIA

Huawei Technologies is accelerating its AI chip development, moving the launch of its Ascend 960DT processor to the first quarter of 2027 as it works to reduce China’s dependence on NVIDIA hardware.

The Ascend 960DT had been scheduled for later in 2027. Huawei also plans to release the Ascend 960PR in the third quarter of 2027, followed by the Ascend 970 in 2028 and Ascend 980 in 2029. The company expects each generation to provide roughly twice the computing power of the one before it.

The faster roadmap is part of a larger Huawei strategy that extends well beyond individual processors. Because its chips still lag NVIDIA’s most advanced GPUs in single-chip performance, Huawei is placing a major emphasis on connecting large numbers of accelerators into unified computing systems.

The commercial opportunity is substantial. Morgan Stanley estimates China’s AI computing market could reach 646 billion yuan, or about $96 billion, by 2030. Huawei’s accelerated roadmap shows how it intends to pursue that market: improve its processors each year while using networking and massive clusters to close the performance gap at the system level.

Focus on Scale

The company’s growth efforts include Peerium Computing Architecture, which Huawei is developing to combine processors, memory, storage and networking resources into large AI systems.

The architecture uses Huawei’s UnifiedBus technology to move data among components and across server racks. Huawei says communication between machines can consume more than 40% of model training time in conventional server environments, making faster interconnects a major source of performance gains.

Huawei’s Atlas 950 SuperCluster can connect as many as 256,000 accelerator cards. Huawei has also disclosed an Ascend 960 SuperPoD configuration supporting 4,096 processors, although earlier plans cited a far larger chip system. The company’s longer-term Peerium strategy is designed to support systems containing up to one million processors.

This focus on scale could help Huawei compete even if its individual accelerators remain less powerful than NVIDIA’s leading products. It also highlights the enormous computing requirements of AI training and inference, where system architecture, memory and networking can have a major impact on overall performance.

Huawei already faces more demand than it can supply. DeepSeek plans to use at least 160,000 Ascend 950DT processors at a data center under development in Inner Mongolia, but Huawei currently lacks the production capacity to fill an order of that size.

Huawei executives say the company’s AI chips now hold a larger share of the Chinese market than Nvidia, though Huawei has not released figures supporting that claim.

The shortage is also limiting Huawei’s international ambitions. The company is testing its AI hardware with selected overseas customers but is prioritizing China because domestic demand exceeds production. Huawei has explored opportunities in markets including Malaysia and Egypt, but a broad global expansion is not currently planned.

Supply constraints go beyond processor manufacturing. U.S. export controls restrict Chinese access to advanced semiconductor equipment and high-end AI memory from suppliers including Micron Technology, Samsung Electronics and SK Hynix. Nvidia’s latest Blackwell processors also are not directly available to Chinese customers.

Huawei is responding by developing more of its technology internally. Its Ascend 950 processors are the company’s first AI accelerators to use Huawei-designed memory. The company has also introduced LogicFolding, an architecture that stacks circuitry vertically to increase transistor density without needing the most advanced manufacturing equipment. Huawei expects to use the technology in Ascend processors by 2030.

The limits on supply are creating costs for customers. Huawei recently raised the price of the Ascend 950DT by 60%, citing shortages of components. Eric Xu, Huawei rotating chairman, said China will not catch up with overall demand for AI hardware, including memory and optical components, until 2030.