Huawei quickens AI chip pace, promises next entrant 3 quarters early

Huawei Technologies announced on Thursday that it will launch its next-generation artificial intelligence chip in Q1 2027, nine months earlier than initially planned, as the company accelerates its ecosystem growth in China's self-sufficiency drive. The Ascend 960DT chip, specifically designed for model training, promises to "double performance" and launch a quarter ahead of schedule.

Meanwhile, the Ascend 960PR, aimed at AI model inference, is set to debut in Q3 2027, a month earlier than expected. At the Huawei Connect 2026 conference in Shanghai, David Wang Tao, Huawei's rotating and acting chairman, disclosed that the Ascend series will adhere to an annual upgrade cycle, with the Ascend 970 and 980 expected in 2028 and 2029, respectively.

These developments come as a backdrop to anticipated high-level trade talks between President Xi Jinping and US President Donald Trump, with AI and semiconductor exports expected to dominate the technology agenda. The Ascend 950, Huawei's most potent current processor, has been serving as an alternative to Nvidia's top-performing graphics processing units (GPUs), which remain inaccessible in China owing to US export controls.

In May, following Trump's visit to China, Huawei unveiled its Tau Scaling Law and LogicFolding architecture, semiconductor advancements aiming to match chips built on a 1.4-nanometre process by 2031 without reliance on advanced lithography tools unavailable in China. At the Thursday event, Huawei also unveiled the Atlas 960 SuperPoD computing cluster, slated for late 2027.

Powered by Ascend 960 chips, this system is projected to enhance performance for training and inference of 10-trillion-parameter models by 2.3 and 2.5 times, respectively, compared to the Atlas 950 SuperPoD, which is scheduled for deployment by the end of this year. AI training, a computationally intensive process, involves models ingesting vast data sets to discern fundamental patterns.

Inference, on the other hand, pertains to the application of a completed model to address user queries and produce output.