Breaking

DeepSeek Releases Huawei Ascend Programming Tools to Reduce Reliance on Nvidia CUDA

September 30, 2026 at 03:41 EDT

Infra & Chips
Open Source
Software Dev & Coding

DeepSeek has publicly released Ascend-oriented AI chip programming infrastructure developed with Huawei, including the high-level language TileLang and related compute and communication libraries. The move gives developers a more open route to program Huawei's Ascend accelerators and marks another step in China's effort to reduce dependence on Nvidia's CUDA software ecosystem.

Sep. 30, 2026 · DeepSeek and Huawei

A CUDA Alternative Takes Shape

DeepSeek has opened Ascend programming tools built with Huawei, giving developers a higher-level route to target domestic AI accelerators.

The Planned Hardware Scale

10,000
chips

160,000
planned Ascend accelerators

The software release arrives alongside a much larger shift toward Ascend-based inference capacity.

160,000

Ascend chips planned for deployment

source-level patches in one field study

384

Ascend 910C NPUs in CloudMatrix384

What the Stack Tries to Replace

Model kernels

Runtime compile

Compute libraries

Communication

TileLang sits at the programming layer, while supporting libraries and runtimes aim to make real LLM workloads portable to Ascend.

Developer Upside

A higher-level language can lower the effort needed to write model-specific kernels for non-Nvidia accelerators.

Open Question

Independent deployments still report reliability work, feature gaps and ecosystem fragmentation on Ascend systems.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.