DeepSeek plans to use at least 160,000 Huawei Ascend 950DT chips to train its V5 model in a new data center in Ulanqab, Inner Mongolia. Pre-training work could begin in late 2026, close to the chip’s planned commercial launch in the fourth quarter of that year. Huawei reportedly embedded engineers with DeepSeek to adapt the hardware and software to the laboratory’s requirements after stability and performance problems with earlier Ascend generations. Each Ascend 950DT has 144 GB of high-bandwidth memory and offers around 4 TB/s of bandwidth, but Liang Wenfeng says four Ascend 950 chips are needed to match the functionality of a single Nvidia GB300 GPU. The shift toward domestic hardware aims to reduce DeepSeek’s reliance on Nvidia amid US export restrictions on advanced AI chips bound for China.
Source: Read the original article
