DeepSeek released six open-source software components for Huawei Ascend AI chips on Wednesday, addressing a barrier to reducing Chinese AI developers’ reliance on Nvidia: access to tools for programming chips, performing calculations and moving data between them. The release offers developers more software to try on Ascend hardware. It does not establish that moving a model is straightforward or that a Huawei system matches an Nvidia one.
That distinction is the central test of the release. Software can make an alternative chip more useful only if developers can integrate it into complete workloads and run those workloads reliably. DeepSeek has published components for that work, but its reported performance results concern selected operations, not independently verified comparisons of complete systems.
Why an alternative chip needs alternative software
Choosing an AI chip also means choosing the tools used to build and run models on it. Programs need a way to specify calculations for the chip; a deployment using multiple chips also needs software that coordinates their work. Nvidia’s position therefore rests partly on its software ecosystem, not just its processors.
DeepSeek’s Ascend release comprises TileLang, DeepGEMM-Ascend, DeepEP-Ascend, TileKernels, FlashMLA and DeepSelect. It extends tools DeepSeek had previously released for Nvidia hardware. The components address different parts of the job rather than offering a single replacement for everything a developer uses on Nvidia systems.
TileLang is a higher-level way to write programs for chip calculations. DeepGEMM-Ascend addresses matrix multiplication, an operation on arrays of numbers used in AI models. DeepEP-Ascend addresses communication between chips during training and inference, when a model generates outputs. The release’s mix of programming, compute and communication tools reflects how many pieces a developer may need to move together.
A programming tool can help a team express what it wants a chip to do. A compute library can supply implementations of particular calculations. Communication software matters when those calculations must be divided among chips and their results brought back together. Improvements in one part do not, by themselves, resolve problems in the others.
DeepSeek says its programming approach could make development simpler. That is a claim about the experience of using the tools, not a demonstrated reduction in the time or expense of moving a model. Publishing the code gives outside developers something to examine and adapt; it does not show how much adaptation their own applications will require.
The possible benefit is nevertheless concrete. A developer considering Ascend needs more than access to Huawei hardware if its existing work depends on software built for Nvidia systems. Tools tailored to Ascend could provide a starting point for that move. Whether they make Ascend practical for a particular team depends on the team’s complete workload, not on the existence of the tools alone.
What the release does not prove
DeepSeek has reported performance results for particular calculations and data-transfer operations. Those results describe the company’s tests under their stated conditions. They are not independent findings about how quickly a complete model trains or serves users, and they do not establish performance parity with Nvidia systems.
The distinction between a component and a complete workload is consequential. A calculation may run efficiently in isolation while an application still spends time moving data, waiting for other chips or dealing with software integration. Likewise, a promising communication result does not show that every calculation the application requires performs equally well.
Comparing systems would require more than placing DeepSeek’s component figures beside a result from another machine. The chips, software versions, model workloads and conditions would need to be specified. The comparison would also need to account for reliability and the engineering work required to get an application running, not just the speed of one operation.
Nor does an open-source release establish that a tool is ready for every developer’s production use. Outside teams must be able to set it up, fit it into their applications and maintain reliable service with it. Those questions cannot be answered by the release announcement or by DeepSeek’s own selected benchmarks.
This limits a tempting reading of the news. DeepSeek has made software available for Ascend; it has not demonstrated, through independent end-to-end tests, that Chinese AI developers can replace Nvidia systems without a substantial change in performance or engineering effort. The release is a potential means of reducing reliance, not evidence that reliance has already fallen.
Free newsletter
Get the morning briefing
Start each day with the AI stories that matter and why — a short, free email from our newsroom.
What outside evidence can—and cannot—settle
Zheng Yu’s July 9 preprint reports that two large-model inference workloads ran on Huawei Ascend 910 hardware. But Yu also documented the work needed for reliable operation: 12 source-level patches, the disabling of some high-throughput features and additional operational safeguards. The preprint has not been peer-reviewed, and it examines an earlier chip and software setup—not DeepSeek’s six new components.
Separately, a developer reported getting a CANN reference implementation for DeepSeek-V4 inference running on eight Ascend 950PR devices. The account describes configuration and software-stack requirements. It is a report of work with a related implementation, not an independent evaluation of DeepSeek’s newly released libraries or a comparison with Nvidia systems.
Those findings make the practical question sharper without answering it for this release. Yu’s study shows that large-model workloads can run on the Ascend setup it examined, while documenting substantial integration work. The developer account shows that another Ascend configuration could run the reported reference implementation. Neither establishes that the new suite removes those hurdles, and difficulties in earlier setups do not establish that it will fail.
The strongest case for the release is therefore practical but conditional: shared programming and library code could spare developers from building every Ascend-specific component themselves. Its limitation is equally practical. The available evidence does not measure how many developers will use that code, how much work it will save or whether it will change their Nvidia purchases.
As Semafor reported, the DeepSeek-Huawei effort sits within a wider push by Chinese technology companies to reduce reliance on U.S. chip suppliers. That strategic goal explains why the software matters. It is not a measure of what the release has accomplished.
The test for a credible alternative
An independent evaluation would start with the released components, not merely with Ascend hardware. Developers would need to run representative training and inference applications, describe their configuration and report what worked, what required changes and what failed. Comparing those applications with equivalent Nvidia deployments would then put DeepSeek’s component results in a useful context.
For teams weighing a move, the important measures include complete-workload performance, reliability and the effort needed to adapt and operate the software. A narrow benchmark could still identify a valuable improvement, but it cannot substitute for those measures. Nor can one successful deployment establish that the tools cover every workload another team might bring.
DeepSeek has supplied more of the software needed to try Ascend chips. The next consequential evidence will come from outside developers using that software in complete applications. Until those results exist, the clearest conclusion is that Huawei hardware has gained another potential route into AI development—not a proven substitute for Nvidia systems.