What Huawei announced
Huawei rotating chairman Eric Xu says the company cannot currently make enough AI computing equipment to satisfy demand in China and is therefore limiting broader overseas expansion. He said testing of the Ascend 950DT had produced good results and that Huawei was in extensive discussions with Chinese model developers, with wider use of 950DT systems for training expected next year.
Huawei also moved forward its next chip schedule. Reuters reports that rotating chairman David Wang said the Ascend 960DT is now planned for the first quarter of 2027, three quarters earlier than previously planned, while the 960PR is targeted for the third quarter, one quarter earlier. Huawei says it intends to release a new Ascend generation every year, with 970 and 980 families following in 2028 and 2029.
The systems strategy is as important as the chip
Huawei is trying to offset constraints on individual accelerator performance by joining very large numbers of processors into one logical computing system. Its new Peerium architecture is intended eventually to connect as many as one million processors, while an Ascend 960 supernode is planned to contain up to 4,096 accelerators before multiple supernodes are linked into much larger clusters.
The company says communication can consume more than 40% of training time in conventional very large clusters, making interconnect design a central part of its strategy. Huawei's UnifiedBus technology links processors, memory, storage and networking inside these systems. More than 1,000 systems based on the earlier Ascend 910C have already been deployed, according to company material cited by Reuters.
Why it matters for China's AI stack
US export controls restrict Chinese access to Nvidia's most advanced accelerators and semiconductor manufacturing equipment, which makes domestic compute capacity a strategic constraint for Chinese AI developers. Huawei is increasingly positioned as the main local supplier trying to provide both the accelerator and the system architecture required for frontier-scale training.
Xu said he believes Ascend's share of China's AI-chip market now exceeds Nvidia's, but Huawei provided no market-share data to support that estimate. Nvidia also retains a major software advantage through CUDA. Huawei says more than 5,200 developers are active monthly on Ascend software and that more than 40 AI models have been trained directly on its platform, figures that likewise come from the company.
What remains unverified
The roadmap and demand statements were made by Huawei executives at HUAWEI CONNECT 2026 and reported by Reuters, but the company has not yet published detailed independent performance data for the 960DT, 960PR or a million-processor Peerium deployment. The largest system is therefore a planned architecture rather than an operating cluster whose scaling efficiency has been demonstrated publicly.
The material confirmed by the conference reporting is the accelerated roadmap, current supply constraint and systems direction. Claims about relative market share, future cluster performance and how effectively the architecture competes with Nvidia at frontier-model training should remain provisional until customers, benchmarks or deployed systems provide independent evidence.