A Chinese AI developer is preparing to migrate its massive computing requirements to local hardware to end its dependence on US-made accelerators by 2027. DeepSeek has ordered 160,000 next-generation Ascend 950DT and 960 chips from Huawei to move its heavy pre-training workloads to domestic silicon. The company currently uses a hybrid strategy, employing Nvidia GPUs for its 33T token pre-training phase while utilizing Huawei Ascend chips for inference and post-training, such as RLHF and SFT. The full transition depends on the next generation of hardware supporting FP4 low-precision training formats.

To facilitate the migration, DeepSeek open-sourced a software stack including TileLang and DeepGEMM that mirrors its Nvidia-based tools one-to-one. The DeepGEMM library reached 99.8% of the theoretical hardware limit for Dense GEMM in official tests. This effort, conducted alongside Huawei, includes the development of a 128 chip supernode based on the Ascend 950 to build an independent GPU software ecosystem. While the tools simplify writing compute kernels to break Nvidia's CUDA moat, supply chain shortages for top-end memory may cap Huawei's 950DT output this year.