DeepSeek Teams Up with Huawei to Open-Source Ascend Programming Toolchain; Domestic AI Compute Replaces Nvidia CUDA

On September 30, 2026, DeepSeek announced on WeChat that it is cooperating with Huawei to open-source the complete programming toolchain for Ascend 950 chips, supporting 128-card “supernode” clusters.

Factual Recap

DeepSeek and Huawei jointly developed and open-sourced the programming infrastructure for Ascend series chips, with Huawei providing full support during development. The two sides also advanced a “supernode” project based on 128 Ascend 950 chips. Bloomberg, Yahoo Finance, and other media outlets reported the news the same day. DeepSeek previously mainly used Nvidia H800 chips and gradually shifted to domestic hardware in 2026; its V4-Flash model has been optimized for Huawei chips.

Mechanism Breakdown

The open-sourced toolchain includes the TileLang high-level language and core libraries such as DeepGEMM and FlashMLA, aiming to enable smooth migration for developers familiar with CUDA. The toolchain directly targets cluster deployment for Ascend 950 chips, addressing compatibility issues from hardware to software stack. DeepSeek’s shift from reliance on H800 to Huawei chips reflects that under export controls, companies are advancing domestic substitution through simultaneous hardware switching and toolchain adaptation.

Industry Impact

This collaboration gives Chinese AI labs, for the first time, a developer toolchain that can fully replace Nvidia CUDA. Huawei’s Ascend chips had previously launched a new product line and are planned for widespread use in model training by 2027. After the toolchain is open-sourced, other Chinese tech companies can reduce dependence on a single supplier and accelerate the building of local AI training clusters. Market observations show that this move, against the backdrop of Beijing introducing multiple policies to support domestic chips, will further promote coordination across the upstream and downstream of the industry chain.

Strategic Assessment

(Analysis, not fact) From compute “bottlenecks” to toolchain independence, the DeepSeek-Huawei collaboration may shorten the deployment cycle of domestic AI infrastructure, but the actual effect still depends on developers’ willingness to migrate and the maturity of the subsequent ecosystem. Historical experience shows that only simultaneous breakthroughs in hardware and software can create sustainable substitution; the current open-source move provides a verifiable starting point for this path.