Huawei announces new AI chips to compete with Nvidia

Chinese Huawei announced two new processors for artificial intelligence, which it plans to launch in 2027, and introduced a new generation of systems for combining a large number of chips into a single computing platform. The company continues to build its own AI infrastructure amid China's limited access to Nvidia's most powerful products.

Huawei presented its new developments on September 17 during the Huawei Connect 2026 conference in Shanghai. According to the company's rotating chairman David Wang, the 960DT processor is scheduled for release in the first quarter of 2027, and the Ascend 960PR in the third quarter.

At the same time, Huawei continues to develop UnifiedBus technology, which allows a large number of processors to be combined and made to work as one large computing system. For modern AI models, this approach is becoming especially important, since the computing power required for their training far exceeds the capabilities of an individual chip.

According to the company, 11 semiconductor solutions have already been developed for large systems based on UnifiedBus. Huawei also stated that it has shipped more than 1000 so-called supernode systems, which combine tens and hundreds of AI processors. They are used by more than 370 customers.

At Huawei Connect, the company also introduced the new Atlas 960 SuperPoD computing cluster and an updated version of UnifiedBus. In the new architecture, Huawei for the first time applied near-packaged optics technology, which is intended to increase data transfer speed between processors while reducing power consumption.

The combination of UnifiedBus with the new Hi-ONE optical system, according to Huawei's estimates, will allow one SuperPoD to combine up to 4000 processors. The company claims that the Atlas 960 SuperPoD with Ascend 960 chips should increase training performance by about 2.3 times and inference performance for models with parameters up to 10 trillion by about 2.5 times compared to the previous generation Atlas 950.

Huawei's strategy is not only to try to close the performance gap of individual chips, but also to scale the number of processors that can work as one system. The company previously reported plans to create a SuperCluster with more than one million AI processors.

The development of its own accelerators has become especially important for China after the tightening of US export restrictions on advanced AI chips and equipment for their production. This has created an opportunity for Huawei to capture a larger share of the domestic market, where Nvidia systems remain the main technological benchmark.

Based on materials from: Nikkei Asia, South China Morning Post, Huawei