DeepSeek and Huawei release open-source tools for Ascend AI chips

In short: DeepSeek says it worked with Huawei to publish open-source software that helps developers use Huawei’s Ascend AI chips.

What happened

DeepSeek said on September 30, 2026 that it collaborated with Huawei to build and open-source programming tools for Huawei’s Ascend AI accelerators. AI accelerators are special computer chips built to run AI work faster, like a calculator designed just for one kind of math.

The release includes TileLang, which DeepSeek describes as a high-level programming language (a simpler way to write instructions for a computer). DeepSeek said TileLang is meant to be easier to use than Nvidia’s CUDA, which is the popular software system many developers use to run AI on Nvidia chips. DeepSeek also said TileLang is not a chip and it is not, by itself, a full replacement for CUDA.

DeepSeek and Huawei also said they built “compute and communication libraries,” which are bundles of ready-made software parts that handle calculations and help chips share data with each other. The companies said Huawei provided substantial support and that they optimized the software to run well on Ascend hardware.

DeepSeek added that the two companies advanced a “supernode” design using 128 Ascend 950 chips, with work focused on both computation and communication between chips.

Why it matters

AI chips are only useful if developers can easily write software for them. Think of it like having a powerful kitchen but no recipes or tools. By making these tools open source (free to use and modify), DeepSeek and Huawei may make Ascend chips more practical for Chinese companies and researchers who want alternatives to Nvidia.

This announcement does not show that Huawei has replaced Nvidia in real-world AI work. Competing also requires strong performance, lots of compatible software, and a large community of users.

Source: NYTimes