Huawei unveiled a new generation of Ascend chips and Atlas SuperPoD systems during the Huawei Connect conference in Shanghai, in a move aimed at providing hardware infrastructure for training and running artificial intelligence models. The announcements come as Chinese companies seek to develop domestic alternatives to advanced hardware whose export to China is restricted by the United States.
Launch roadmap extending through 2029
Wang Tao, Huawei’s vice president, said the Ascend 960DT chip, designed for training artificial intelligence models, will launch in the first quarter of 2027, while the Ascend 960PR, intended for inference operations—that is, processing data after the model has been trained—will arrive in the third quarter of the same year.
The company also announced that the next two generations, Ascend 970 and Ascend 980, will launch in 2028 and 2029, respectively. Huawei says the Ascend 950, currently in use, helps close part of the gap created by U.S. export restrictions on Nvidia’s advanced chips, alongside the computing and networking equipment provided by the company.
New architecture said by Huawei to overcome scaling limitations
Huawei based its presentation on a principle it calls “Tao’s Scaling Law,” which it introduced in May as an approach to overcoming some limitations associated with Moore’s law. According to the company, the approach relies on a folded architecture that shortens signal latency rather than relying exclusively on reducing component dimensions.
Wang explained that Huawei aims by 2031 to develop semiconductor circuits with performance equivalent to 1.4-nanometer chips, without requiring advanced lithography machines subject to similar export restrictions on China. He also claimed that the architecture could significantly increase processor performance, bandwidth, memory capacity, and communication speeds without comparable process-node shrinkage.
Huawei had previously indicated that it began using this architecture in Mate 2 phones, but its application to artificial intelligence chips remained uncertain. According to the available details, the presentation did not provide independent test results proving these claims or fully explain the manufacturing method for the chips in question.
Optical interconnects inside Atlas SuperPoD
The company also unveiled the Atlas 960E SuperPoD system, which will use NPO technology, or near-package optics, to connect circuits and modules through optical links instead of electrical connections. This architecture is intended to improve data transfer among a large number of graphics processing units within a single cluster.
Huawei says the system was designed to accelerate the training and running of artificial intelligence models when they reach 10 trillion parameters. However, the source does not mention the system’s commercial availability date, detailed specifications, or performance results that can be compared with competing platforms.
Why does this announcement matter?
The actual change is not limited to adding a new chip; rather, it lies in presenting an integrated roadmap that combines processors, connections between computing units, and networking equipment. This could give Chinese companies a more integrated option for developing artificial intelligence systems amid restrictions on access to some foreign chips and equipment.
Nevertheless, most of the long-term figures, such as the 2031 target, remain statements from Huawei rather than independently verified results. The practical value of the announcement will therefore depend on when the Ascend 960 reaches the market, its actual ability to compete with alternatives, and the extent to which NPO technology succeeds in large-scale production systems.