Huawei limits new Ascend AI chips to China due to capacity constraints
Huawei announced that its upcoming Ascend 960‑series neural processing units will be sold only within China because the company cannot meet domestic demand with its current production capacity. Rotating chairman Eric Xu said at the Huawei Connect conference that the firm may supply limited volumes to a few countries with strong demand, but there is no plan for a full international rollout.
Key points
- Huawei will sell its next‑gen Ascend 960‑series NPUs only in China, citing insufficient capacity for domestic demand.
- The Ascend 960DT aims for 2 PFLOPS FP8 training and 4 PFLOPS FP4 inference; 960PR targets 8 PFLOPS FP4, still below Nvidia’s comparable GPUs.
- Atlas SuperPoDs can pack up to 15,488 NPUs delivering up to 120 EFLOPS FP4, but power efficiency lags Nvidia, limiting export appeal.
The Ascend 960DT and 960PR are projected to reach 2 PFLOPS FP8 training and 4‑8 PFLOPS FP4 inference performance by 2027, with later Ascend 970 and 980 models targeting up to 28 PFLOPS FP4. Even so, benchmark figures place these chips well below Nvidia’s H200, B300 and R200 GPUs. Huawei’s Atlas SuperPoDs, which can house up to 15,488 NPUs and deliver up to 120 EFLOPS FP4, rely on massive system‑level scaling that consumes far more power than competing Nvidia solutions, limiting their attractiveness outside China.
Huawei shelves global AI chip rollout as China's own demand outstrips supply — 15,488-chip Atlas clusters leverage optical networking to counter Nvidia, scales to 120 EFLOPS
Tom's Hardware · 21 September 2026
Loading the full article…
This text was published by Tom's Hardware and written by Anton Shilov. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Hardware & Compute
All →- Applied Materials and Intel expand AI chipmaking collaboration at epic Center · 1 src
- Xeal to deploy 100,000 Nvidia GPUs in EV charging pods across US · 1 src
- Meta’s Muse boosts AMD stock as AI agents shift workload to CPUs · 1 src
- Nokia CEO says data center construction could double without supply constraints · 1 src
- Ghost launches $3,499 AI agent computer with $11M seed · 1 src
Comments
via GitHub Discussions