Huawei is preparing to launch its new generation Ascend 960DT artificial intelligence chip about six months earlier than the previously announced schedule. According to the information shared by the company at the Huawei Connect event, Ascend 960DT is planned to be ready in the first quarter of 2027. Huawei previously targeted the third quarter of 2027 for this chip. The company spokesperson stated that the Ascend 960 family will arrive ahead of schedule and that the performance is aimed to be doubled compared to the previous generation. This change in calendar shows that Huawei is accelerating its work in the field of artificial intelligence accelerators and large-scale computing systems, where Nvidia is strong.
Rather than focusing solely on the performance of individual processors, Huawei aims to combine multiple AI accelerators into a single large computing system. The company describes this approach as Peerium Computing Architecture. The architecture is based on UnifiedBus technology, which connects processors with memory, storage and networking hardware. According to Eric Xu, Huawei’s rotating chairman of the board, the new architecture will allow the establishment of larger systems for both training artificial intelligence models and inference processes. This approach also reveals Huawei’s desire to expand its own hardware ecosystem, as rivals such as Nvidia offer not only chips but integrated solutions through high-speed connections and large data center clusters.
Huawei Atlas 950 systems will be able to combine hundreds of thousands of accelerators
The first products developed based on Peerium Computing Architecture will be the Atlas 950 SuperPoD and Atlas 950 SuperCluster systems. According to the information provided by Huawei, up to 256 thousand accelerator cards can be connected to each other in an Atlas 950 SuperCluster. A system of this scale emphasizes the speed of data transfer between systems, latencies, and the efficiency of the software infrastructure rather than the raw performance of individual processors. Since thousands or even tens of thousands of processors need to work simultaneously in training large language models and similar artificial intelligence systems, the connection architecture is becoming more and more decisive. For this reason, Huawei positions UnifiedBus technology as one of the key components of its new product family.
However, there are some question marks about the scale targets announced by the company. Analyst Rui Ma, who follows the Chinese technology sector, stated that Huawei had previously said that the Atlas 960 SuperPoD system could be scaled up to 15 thousand 488 Ascend 960 chips, and in the last statement, a structure with 4 thousand 96 chips was mentioned. According to Ma, although it is a positive development that Ascend 960 chips have been brought forward significantly in the calendar, the announced SuperPoD system falls behind the previously stated scale. Therefore, it is not yet clear whether the acceleration on the chip side means that equally large systems will be ready in a short time. If Huawei provides more detailed information about actual system configurations, energy consumption, connection performance and software compatibility in the coming period, it may become more visible how to explain this difference.
Huawei’s semiconductor work continues under US sanctions that limit its access to China’s advanced chip and manufacturing technologies. Despite this, the company continues to develop its own artificial intelligence processors, connection technologies and data center systems. Rui Ma also argues that China’s goal of self-sufficiency in semiconductors has become a stronger priority due to current conditions, and that it is difficult for external restrictions to completely stop investments in this direction. In addition, the real competitiveness of the Ascend family will be determined not only by technical features, but also by production capacity, software support, developer tools and the extent to which large customers will adopt the systems. Considering the software ecosystem that Nvidia has built around CUDA over the years, Huawei will need to support progress on the hardware side with a similarly powerful software layer.
The timing of the statement coincides with a period when the competition between the USA and China around artificial intelligence and semiconductor technologies continues. According to the information provided by the source, US President Donald Trump and Chinese President Xi Jinping are expected to meet in Washington on September 24. While Trump opposes calls to reduce the pace of artificial intelligence development for security reasons, he argues that the United States must maintain its technological superiority over China. Eric Xu states that Chinese companies need to accelerate their work in the field of artificial intelligence in order to close the gap with the USA. The Ascend 960DT’s calendar, which is set for the first quarter of 2027, shows that Huawei is trying not only to prepare a more powerful processor, but also to create an infrastructure that can build very large artificial intelligence clusters with its own hardware and connection technologies.