DeepSeek, Huawei Open-Source Ascend Chip Toolkit

DeepSeek publicly released a software toolkit co-developed with Huawei, including TileLang, an open-source alternative to Nvidia's CUDA, to program Huawei's Ascend AI accelerators. The move deepens the two Chinese firms' partnership as they work to build a domestic AI hardware-software stack that can substitute for Nvidia's ecosystem.

2 effects · 1 potentially affected company · The full cascade analysis is part of the Pro plan. Create free account

This release targets the software layer that has historically locked developers into Nvidia's CUDA, a key pillar of Nvidia's pricing power and moat in China. By open-sourcing TileLang for Huawei's Ascend 950 chips, DeepSeek and Huawei aim to lower switching costs for Chinese AI developers, accelerating domestic adoption of non-Nvidia accelerators. Combined with DeepSeek's reported plan to deploy 160,000 Ascend chips in Inner Mongolia, this signals a credible, well-funded effort to build an end-to-end Chinese AI compute stack ahead of DeepSeek's expected public listing near a $74 billion valuation. For Nvidia, this incrementally weakens long-term China exposure and could set a template other markets watch as export-control pressure persists. For Huawei, broader software tooling support could materially expand Ascend's addressable developer base domestically and potentially in other CUDA-restricted markets.