DeepSeek Opens Huawei Ascend Tools to Challenge CUDA

DeepSeek has open-sourced six infrastructure components for Huawei Ascend AI accelerators, strengthening China’s domestic AI-computing ecosystem and offering an alternative to Nvidia’s CUDA platform. The release includes TileLang Ascend, a CUDA-like programming layer, alongside FlashMLA, DeepEP, DeepGEMM, TileKernels and DeepSelect for attention, distributed communication, matrix operations, vector computing and TopK selection. DeepSeek said the Ascend toolchain supports much of the operator stack used to train its models. Developed with Huawei, it targets a 128-chip Ascend 950 supernode. DeepGEMM supports BF16, FP8 and FP4 operations, while some APIs remain compatible with Nvidia versions. DeepSeek had previously adapted its V4 model to Huawei hardware and is reportedly planning a data centre in Inner Mongolia using at least 160,000 Ascend accelerators. Performance claims remain subject to limitations. Tests used unreleased proof-of-concept hardware and manual configurations. Full bandwidth is expected with Huawei’s commercial Atlas 850E development kit, planned for October 2026. DeepEP reached about 90%–95% of physical bandwidth with expert parallelism capped at 32, but larger-scale communication remains under optimisation. Some Ascend features are experimental, FlashMLA’s fused operator remains Nvidia-only, and Ascend DeepSelect currently supports only BF16. For crypto traders, the DeepSeek release has no direct impact on cryptocurrency fundamentals or prices. It may support sentiment around Chinese semiconductor and AI infrastructure companies, but software compatibility, execution risks and delayed hardware availability limit the near-term market effect. The longer-term significance is a possible shift in AI hardware demand away from CUDA, although this is unlikely to create a direct crypto catalyst.