Thought Leadership
Huawei is bringing forward the market launch of the Ascend 960 AI chip, aiming to move closer to Nvidia.
Chinese technology group Huawei unveiled an updated, accelerated roadmap for its Ascend series of AI accelerators at its Huawei Connect conference in Shanghai. The Ascend 960DT variant, originally scheduled for late 2027, was reportedly moved forward by three quarters to the first quarter of 2027. According to its own statements, Huawei aims to further close the gap in the global race for AI chips, which is currently led by US vendor Nvidia.
Roadmap with Annual Cadence
Following the commercially available Ascend 950 (in the 950PR variant for inference tasks and 950DT for training and decoding), the Ascend 960 is set to follow in 2027, the Ascend 970 in 2028, and the Ascend 980 in 2029. According to the company, each new generation is expected to achieve roughly double the computing power of its predecessor. According to several trade media reports, the Ascend 960 is expected to roughly match the performance class of Nvidia’s current Blackwell architecture—putting Huawei mathematically about two years behind Nvidia, compared to a gap of five or more years around 2020.
Optical Networking Instead of Pure Chip Performance
At the center of the presentation was the so-called Ascend 960 Supernode, which Huawei described as the world’s first supernode featuring near-packaged optics (NPO) technology. The system is designed to support models with up to ten trillion parameters and, in one configuration, bundle 4,096 computing units with a latency of around two microseconds. In addition, Huawei introduced the Ascend 960 SuperPoD, which can interconnect up to 4,000 processors in a cluster. The UnifiedBus interconnect technology used for this relies on optical connections to achieve microsecond-range latencies; the near-packaged optics technology is intended to reduce signal loss and heat generation. Analysts view this strategy as an attempt to offset disadvantages in pure single-chip performance through advantages in networking entire computing clusters.
High Demand Despite Export Restrictions
Meanwhile, demand for domestic computing power in China remains high: Chinese AI provider DeepSeek is reportedly planning to deploy 160,000 Ascend 950DT chips in a new data center in Inner Mongolia. US export restrictions directly impact Huawei’s chip access: After memory manufacturers SK Hynix, Samsung, and Micron ceased shipments of their latest AI memory chips to Huawei, prices for the Ascend 950DT have reportedly surged by around 60 percent due to component shortages. US investment bank Morgan Stanley estimates that China’s spending on computing infrastructure could rise to as much as 646 billion yuan (around 96 billion US dollars) by 2030. Beyond China, Reuters reports that Huawei is also exploring business opportunities in countries such as Malaysia and Egypt.
(Editorial Team)
