At its October 2026 release, DeepSeek's DeepInfer model-serving software ran on Nvidia SM100 data-center GPUs and Huawei Ascend 950 chips