DeepSeek

AI Chips · Semiconductors · Open Source Software

Launch

DeepSeek open-sources TileLang toolkit to rival Nvidia's CUDA on Huawei chips

A six-module release targets the four million CUDA developers that keep Nvidia locked into China's AI chip market.

• DeepSeek open-sourced six software modules for Huawei's Ascend AI chips, led by TileLang, a programming language pitched as a simpler alternative to Nvidia's CUDA.
• Hangzhou-based DeepSeek built the toolkit with Huawei's full support, and the pair jointly optimized a 'supernode' cluster linking 128 Ascend 950 chips.
• The six released projects — TileLang, DeepGEMM-Ascend, DeepEP-Ascend, TileKernels, FlashMLA and DeepSelect — mirror DeepSeek's existing Nvidia-focused toolkit component for component.
• DeepSeek's benchmarks show DeepGEMM-Ascend hitting up to 99.8% compute utilization on dense matrix operations running on Ascend 950DT chips.
• Nvidia's CUDA lock-in stems from an estimated four million developers worldwide who build on it, a moat rivals like AMD haven't managed to cross even with competitive hardware.
• The release comes two weeks after Huawei unveiled its next-generation Ascend processors and supernode systems, which it expects to see wide use in model training in 2027.
• By making the toolkit free and API-compatible with its Nvidia-based tools, DeepSeek is trying to erase the switching cost that has kept Chinese AI developers tied to Nvidia hardware.