Huawei’s answer to Nvidia’s data center silicon will reach customers sooner than advertised. The Ascend 960DT, a training part, is now set for the first quarter of 2027, pulled forward from the fourth quarter of next year.
An inference-tuned sibling, the 960PR, follows in the third quarter of 2027, with the 970 slated for 2028 and the 980 for 2029. Wang Tao, Huawei’s rotating chairman, said the line came together early and roughly doubles the compute of parts shipping today.
Huawei argues the contest is one of systems, not single dies. Wang pointed to UnifiedBus, its chip-to-chip protocol, and said deployments of Atlas SuperPoD systems have passed 1,000 units among roughly 370 customers. Clusters are designed to hold as many as 100,000 accelerators, with networking asked to make up for gaps against Nvidia on a per-chip basis.
Optics enters next. Wang said modules built on near-packaged optics technology ship within the coming quarters, moving data between accelerators and servers faster than pluggable parts allow.
Interest is not the problem. Huawei says orders for its AI chips exceed what it can supply, and DeepSeek alone is fitting out an Inner Mongolia data center with at least 160,000 of the earlier 950DT parts. Prices for Chinese accelerators have climbed too as HBM gets scarce.
Beyond China, Huawei is courting Malaysia and Egypt and selling its stack as an alternative to US silicon. That pitch sharpens as Nvidia’s Jensen Huang predicts Chinese firms could field their own advanced lithography tools by 2030 and Washington weighs fresh controls.
