How DeepSeek’s Huawei Toolkit Could Challenge Nvidia’s CUDA

DeepSeek’s software release targets a practical barrier to using Huawei’s Ascend AI chips: developers need more than hardware. They also need tools to program the chips and coordinate work across them. DeepSeek and Huawei have open-sourced Ascend programming infrastructure, including TileLang and compute and communication libraries, and report a jointly developed 128-chip Ascend 950 system. 1813

The aim is to make Huawei’s platform a more workable alternative to Nvidia’s CUDA ecosystem. But an open-source toolkit and a large-chip configuration are starting points, not evidence that developers have switched or that the system matches CUDA-based platforms in performance.

What the Ascend toolkit includes

TileLang is a programming language for developing AI software. Its role is to give developers a way to write programs for Ascend hardware without relying on Nvidia’s CUDA platform. DeepSeek’s Ascend version of TileLang is among the tools made available for free. 68

The accompanying libraries address two other parts of the job:

• Compute libraries provide software for carrying out AI operations on the chips.
• Communication libraries help chips exchange data and coordinate work when a task is distributed across multiple accelerators. 11325

Together, the programming layer and libraries are intended to support both work on individual chips and coordination across a larger system. Their availability gives developers components to build with, but it does not establish that every model or workflow will run smoothly on Ascend.

What the 128-chip system shows—and what it doesn’t

DeepSeek and Huawei have jointly advanced a “supernode” based on 128 Ascend 950 chips, with computing and communication optimized together, according to reports on the partnership. 1324

That number signals the scale of the configuration they say they have worked on. It should not be read as proof that the system can replace any Nvidia cluster, or that it delivers comparable training speed, reliability, or cost: the available reports do not provide those comparisons.

Why Huawei’s support matters

DeepSeek said Huawei provided full support in developing the programming infrastructure. 1 That backing matters because software for an AI accelerator has to work with the hardware it targets; the partnership brings chip and software development together rather than treating them as separate pieces.

The effort also fits a broader push among Chinese technology companies to build alternatives to Nvidia’s ecosystem. U.S. export controls are part of the context for Chinese developers seeking domestic options, although the toolkit’s release does not by itself remove hardware or software constraints. 114

The real test is whether developers adopt it

Making the tools open source and free to download lowers one barrier to trying them. But switching an established workflow also depends on whether the tools are dependable, useful for developers’ workloads, and supported well enough to keep improving. The reports establish the partnership, release, and 128-chip configuration; they do not establish broad developer adoption or parity with CUDA. 1813

That is the distinction to keep in mind: DeepSeek and Huawei have laid out a credible attempt to make Ascend more programmable and usable at scale. Whether it becomes a serious CUDA alternative will depend on how well the toolchain works in practice—and whether developers choose to build with it.