DeepSeek Reportedly Training Next-Gen AI Models on Huawei Ascend 950 Chips
Reports indicate that Chinese AI lab DeepSeek is utilizing domestic Huawei Ascend 950 chips to train its frontier large language models. This shift highlights a transition toward locally produced AI hardware amid ongoing trade restrictions on US chip exports. Successfully training state-of-the-art models on Huawei hardware validates China's push for AI hardware independence from Nvidia. It demonstrates that domestic AI chip architectures are becoming viable options for frontier LLM development. Huawei's Ascend 950 series offers high compute density and high-bandwidth memory (HBM) designed to challenge Nvidia's AI accelerators. However, software stack optimization and cluster-scale interconnect efficiency remain major engineering challenges compared to Nvidia's CUDA ecosystem.
## BACKGROUND
US sanctions have restricted leading Chinese AI companies from importing cutting-edge Nvidia GPUs, driving demand for domestic hardware alternatives. In response, Huawei developed the Ascend family of AI processors along with its proprietary CANN compute architecture. DeepSeek, founded by Liang Wenfeng, is a prominent Chinese AI research organization recognized for releasing open-weights frontier LLMs.