Huawei used its annual Huawei Connect conference in Shanghai on Thursday to lay out an accelerated AI hardware roadmap aimed squarely at Nvidia. The company unveiled the Ascend 960 SuperPoD cluster and said the Ascend 960DT processor has been pulled forward to the first quarter of 2027, three quarters earlier than originally planned. The Ascend 960PR is expected to follow in the third quarter of 2027.
The new Atlas 960 SuperPoD builds on the Atlas 950 SuperPod introduced only months ago. Huawei says the system improves both AI training and inference. It integrates near-packaged optics, UnifiedBus interconnect technology, and Hi-ONE optical systems to connect up to 4,096 processors in a single supernode. Huawei also introduced a broader architecture called Peerium, which is designed to eventually link as many as one million processors into one computing system.
Rotating chairman Eric Xu said demand for Huawei's AI chips is outpacing supply inside China, so the company has no major plans to expand international sales. He framed the push as a response to US export controls. "We cannot accept a destiny where we cannot control our fate being determined by others in terms of willingness to sell chips to China or not," Xu said.
Huawei says its Ascend chips now hold a larger market share in China than Nvidia, and more than 40 AI models have been trained on its platform. The company also highlighted a commitment from DeepSeek, which is building a data center in Inner Mongolia that will run on about 160,000 Ascend 950DT processors. Still, analysts note the performance gap remains significant: Bernstein has estimated the Ascend 950 delivers only about 6% of the performance of Nvidia's upcoming VR200 superchip.
