DeepSeek: "...open-sourcing the infrastructure components for Huawei's Ascend computing platform, including the TileLang high-level language compiler, computing libraries, and distributed communication libraries. All components correspond one-to-one with the previously open-source components for the NVIDIA platform. TileLang has already been validated on NVIDIA platforms and currently powers the majority of the operators used in training DeepSeek’s state-of-the-art V4 series models." Qwen says: Currently, DeepSeek uses a hybrid training strategy: NVIDIA for the massive 33T token pre-training phase, and Huawei Ascend for post-training (RLHF/SFT) and inference. However, with DeepSeek recently ordering 160,000 next-generation Ascend 950DT/960 chips, the company is building the infrastructure to move pre-training to Ascend by 2027, once the new chips support the necessary low-precision (FP4) training formats.