DeepSeek will use Huawei Ascend chips to train smaller versions of its R2 family. The company is testing accelerators from multiple Chinese vendors as it tries to cut dependence on Nvidia. The choice reflects wider moves in China to build a domestic AI hardware base.
DeepSeek will still rely on Nvidia for its largest R2 model. The company sees Nvidia GPUs as the most reliable option for very large models today. DeepSeek faces engineering and timing challenges with R2 and has delayed its wide release while work continues.
Why the shift
DeepSeek must balance speed cost and supply. Nvidia remains dominant for top tier models. Nvidia GPUs offer proven software tools and broad industry support. DeepSeek will therefore keep Nvidia for R2 sized training jobs for now. At the same time the company will train and tune smaller model variants on Ascend and other domestic chips. That will reduce the number of Nvidia hours DeepSeek needs and lower exposure to export limits.
Testing several chips also helps DeepSeek learn how models run on different hardware. Reports say the firm is evaluating accelerators from Huawei Baidu and Cambricon. Each chip family has different performance trade offs and software stacks. The learning will guide which models are moved off Nvidia and which must stay.
The move is not without risk. Earlier attempts to train large models on domestic chips showed limits. DeepSeek paused or delayed parts of R2 work after tests with Ascend did not meet internal targets. That pushed the company to rely on Nvidia for higher scale work while it keeps tuning Ascend for smaller tasks.
What to watch
Nvidia has noted the new competition and said the market will choose the best stack for each use case. That view underscores how customers must weigh software support ecosystem maturity and real performance data from independent tests. DeepSeek and its peers will likely publish more details as they qualify chips for production workloads.
DeepSeek is moving carefully. The change reduces single vendor risk and aligns with local supply goals. It also reflects the engineering reality that large and small models have different hardware sweet spots. Expect gradual shifts with clear testing milestones and with more public signals about which chips power which models in the months ahead.