Huawei plans to launch its Ascend 960DT AI chip in the first quarter of 2027, about six months earlier than previously planned; the earlier target was the third quarter of 2027. The company aims to double performance with the Ascend 960 family compared with the previous generation. Huawei is developing its Peerium Computing Architecture, which combines numerous accelerators in a single system, in the fields of AI accelerators and large-scale computing, where Nvidia is strong. UnifiedBus connects processors to memory, storage and networking hardware. The first products will be Atlas 950 SuperPoD and Atlas 950 SuperCluster; up to 256 thousand accelerator cards can be connected in the SuperCluster. According to Eric Xu, the architecture will enable the construction of larger systems for training and inference of AI models.
Why it matters
Bringing the timeline forward signals Huawei’s effort to strengthen its position in AI accelerators and large-scale computing sooner. The architecture under development brings into focus not only the capacity of a single chip but also the scale of systems in which numerous accelerators will work together with memory, storage and networking components. The development therefore concerns organizations that assess hardware selection and system design together for infrastructure used to train and run large AI models. However, the extent to which the planned performance increase and accelerated launch schedule will be achieved in practice will become clear when the products are launched.
Background
Nvidia is not a new name in the FikirPilot archive: over the past 90 days, we have published 27 articles mentioning the name, the latest dated September 18, 2026.