DeepSeek completes full migration to Huawei Ascend, eliminates NVIDIA dependency
Liang Wenfeng spent months migrating DeepSeek's codebase to Huawei's domestic silicon, proving Chinese AI infrastructure can operate independently of U.S. hardware despite engineering costs and release delays.
DeepSeek founder Liang Wenfeng completed a months-long migration of the company's entire codebase from NVIDIA GPUs to Huawei's Ascend chips. The move was strategic, not urgent—DeepSeek had already demonstrated strong performance on NVIDIA hardware—but Wenfeng chose to prove Chinese AI infrastructure could operate without American silicon as U.S. export controls tightened.
The project consumed significant engineering resources and delayed the next model release. Wenfeng's team worked under pressure with no guarantee the port would succeed at scale, yet delivered a fully functional system on domestic chips with no reported quality loss. DeepSeek now runs inference and training on Ascend without architectural compromises, establishing Huawei's accelerators as viable for production AI workloads. Other Chinese AI labs now have a reference implementation for moving off NVIDIA hardware, a shift that could accelerate as export controls expand. The migration marks the first major Chinese foundation-model company to publicly operate entirely on domestic accelerators, reshaping procurement options across the Chinese AI sector.
