Huawei says China’s AI computing demand exceeds supply as it accelerates two 2027 Ascend chips and plans systems linking 1 million processors in its Nvidia challenge.
Key Points:
- Huawei cannot make enough AI computing equipment to meet demand in China and is limiting most overseas sales.
- The Ascend 960DT is now due in the first quarter of 2027, three quarters earlier than previously planned.
- Huawei’s Peerium architecture is designed to link as many as 1 million processors into a single computing system.
Huawei Chip Roadmap
Reuters reported Sept. 17 that Huawei cannot produce enough AI computing equipment for Chinese customers, forcing the company to limit overseas sales while expanding its Ascend processor roadmap. Domestic demand comes first.
Rotating Chairman Eric Xu said Ascend 950DT testing produced strong results and that Huawei is discussing the chip with Chinese AI model developers that could begin training on related systems next year. Huawei gave no data to support Xu’s separate estimate that Ascend has a larger China market share than Nvidia.
Rotating Chairman David Wang said Huawei plans to have the Ascend 960DT ready in the first quarter of 2027, three quarters earlier than previously planned. The 960PR is due in the third quarter, one quarter ahead of schedule.
Also Read: iPhone Duo Vs Galaxy Z Fold 8: Five Reasons To Pick Apple's Foldable
Nvidia Competition
Huawei is also trying to offset limits on individual Chinese chips by linking large numbers of processors into systems that can train advanced AI models more efficiently. Its new Peerium architecture is designed to connect up to 1 million processors, while an Ascend 960 supernode could combine 4,096 chips before multiple units form much larger clusters.
The strategy matters because China still faces restrictions on access to Nvidia’s most advanced processors and semiconductor manufacturing equipment, while Huawei has faced separate U.S. trade restrictions since 2019. Xu called “full self-sufficiency for chips” the way forward.
Huawei says clusters of about 100,000 chips are becoming common for training some of the largest AI models, where machine communication can consume more than 40% of training time. Faster links could reduce that bottleneck, although Nvidia retains a major software advantage through CUDA, its widely used platform for building and running AI applications.
Huawei’s AI infrastructure push is already well underway. The company says it has deployed more than 1,000 Ascend 910C systems, moved Ascend 950 systems into commercial use and supported more than 40 AI models trained directly on its platform.
Read Next: Tesla Stock Gets Ron Baron Buy-Now Call With 55% FSD Adoption
