DeepSeek and Huawei released open-source programming tools for Huawei’s Ascend chips on Wednesday, aiming to reduce reliance on Nvidia’s software ecosystem. We think that aim matters. But publishing code is only the first step: the tools will change developers’ choices only if teams beyond the two partners can use them to run AI workloads reliably, without prohibitive engineering effort.
The release is a meaningful attempt to address the software side of chip competition. It is not evidence that developers have adopted the tools widely, that workloads have moved off Nvidia, or that Huawei’s chips match Nvidia systems in practical use. Those are separate claims, and the available evidence does not establish them.
Code is available; an ecosystem must be used
The South China Morning Post reports that DeepSeek released six software modules and described its goal as an independent software ecosystem. Among the tools are TileLang, a language for writing the small programs that perform calculations on chips, and libraries for computation and communication between chips. That addresses a real obstacle: developers need software that lets them make effective use of hardware, not simply access to the hardware itself.
The scope of the release is specific. DeepSeek’s DeepGEMM-Ascend documentation says its initial version supports Ascend 950 devices and lists requirements for running it. TileLang’s release notes also describe native Ascend 950 support. Those documents establish what the projects say they support; they do not show how readily an outside team can install the tools, adapt an application and keep it running.
Open source makes code available for others to inspect, modify and try. That can help developers build an alternative to software tied to a particular chip supplier. It cannot, by itself, remove the cost of debugging an application, filling gaps in software support or maintaining a working deployment. The South China Morning Post reports that Nvidia remains TileLang’s primary backend, underscoring how different adding Ascend support is from displacing an established option.
Promising figures are not a developer’s result
DeepSeek’s own FlashMLA documentation reports 410 TFlops during prefill, when a model processes its initial input, and 360 TFlops during decoding, when it generates output. Those figures describe trillions of floating-point calculations per second for a particular sparse-attention program under workloads DeepSeek specifies. The project says they represent 95% and 83% of theoretical peak performance, respectively. They are DeepSeek’s measurements, not independently verified comparisons with Nvidia systems on matched tasks.
Free newsletter
Get the morning briefing
Start each day with the AI stories that matter and why — a short, free email from our newsroom.
Tom’s Hardware reports that the companies worked on optimization for a supernode built around 128 Ascend 950 chips. That describes the scale of the work the partners have discussed, not independently established production performance. A fast calculation or an ambitious configuration would still leave developers asking whether their complete applications run correctly and consistently, and what it takes to get them there.
There is reason not to dismiss the effort. In a July 9 arXiv preprint, Zheng Yu reports that two large-model inference workloads—applications serving model outputs—ran on a 16-device Ascend 910 system. Successful service on earlier Ascend hardware is evidence that deployment is possible. It also came with costs: the paper reports twelve source-level patches, disabled performance features and added operational safeguards.
That study is a preprint, not peer-reviewed research, and it examined earlier chips and software. It cannot tell us whether the new Ascend 950 tools solve those problems. Nor should its engineering difficulties be used to declare the new release a failure. Its value is more precise: it identifies the kinds of work that a performance figure alone can conceal.
Show the work required to switch
The strongest case for DeepSeek and Huawei is that an openly available software stack gives developers another route to useful AI computing. If outside teams can reproduce results and deploy workloads without extensive fixes, Nvidia’s software ecosystem could become less of a constraint on their choices. The evidence so far supports the possibility, not the conclusion. None of the available accounts establishes how many outside developers use the new tools or how many have shifted workloads from Nvidia.
Our position is that the release deserves a practical test rather than either a victory lap or a dismissal. DeepSeek and Huawei should publish reproducible comparisons of representative workloads on Ascend and comparable Nvidia systems, including the software versions and conditions used. Independent teams should be able to check correctness, performance and reliability—and report the time and changes required to make each workload run. Evidence of use by developers outside the partner companies would answer a different, equally necessary question: whether the tools are becoming an ecosystem.
For developers weighing chip options, usable software is the stake. A release can open that choice. Only demonstrated use can show whether it has.