Huawei plans to release its next-generation Ascend 960DT AI accelerator in the first quarter of 2027, six months earlier than its previous third-quarter target. The revised schedule intensifies its attempt to supply China with alternatives to Nvidia hardware under US semiconductor restrictions.
The company says the new chip will double performance, though the announcement does not provide independent benchmark results. Huawei is pairing the accelerator with its Peerium Computing Architecture, which links processors to memory, storage and networking through a technology called UnifiedBus. The design is intended for large training and inference clusters.
Its first systems under the architecture are the Atlas 950 SuperPoD and SuperCluster. Huawei says a SuperCluster can connect as many as 256,000 accelerator cards. However, analyst Rui Ma noted that the newly described SuperPoD is smaller than an earlier plan: Huawei had previously discussed 15,488 Ascend 960 chips in one system, while the current announcement refers to 4,096.
The accelerated chip launch does not remove manufacturing and scaling constraints created by export controls. It does show that Huawei is pursuing performance through both faster individual accelerators and very large interconnected systems, a strategy aimed at narrowing China’s gap in available AI computing capacity.