DeepSeek just released software designed to make it easier for Chinese AI labs to test Huawei hardware without rebuilding every low-level part of their AI stack.
On September 30, DeepSeek open-sourced six software components tailored to Huawei’s Ascend AI chips, extending infrastructure it had previously developed around Nvidia hardware. The release includes tools for programming, computation, and communication—the boring-sounding plumbing that determines whether expensive AI chips actually spend their time doing useful work. Reuters reported that DeepSeek developed the tools with Huawei’s support as part of an effort to build what the company described as a more independent computing ecosystem.
That ambition makes for a neat Nvidia-versus-Huawei headline. The more interesting story starts one level down.
DeepSeek has attacked one of the costs of choosing Huawei: the software work required to get there. Whether that turns Ascend into a dependable alternative to Nvidia depends on a much less tidy set of questions involving chip supply, engineering effort, reliability, and full-scale model performance.
The code shipped. Now someone has to prove the whole system works.
- DeepSeek is attacking Nvidia’s stickiest advantage
- This effort actually started a year ago
- Kernel benchmarks are only part of the test
- There’s a credible case that the software could improve quickly
- Better software cannot solve a chip shortage
- Export controls make the hardware choice even messier
- Here’s what would actually prove the shift is happening
DeepSeek is attacking Nvidia’s stickiest advantage
Nvidia’s grip on AI computing has never been just about making fast chips. A huge part of the advantage is CUDA, the software platform developers have spent roughly two decades using to program Nvidia GPUs.
Once a company has models, libraries, debugging workflows, infrastructure, and engineers built around that ecosystem, switching hardware becomes more complicated than buying a different accelerator. Every incompatible piece creates another engineering bill.
That wider platform strategy is a big reason Nvidia increasingly sells AI as an entire computing stack rather than a box of processors, something we explored in The Neuron’s breakdown of Nvidia’s GTC 2026 strategy.
DeepSeek’s new tools are designed to shrink that software switching bill.
Its new DeepGEMM-Ascend repository ports a library of optimized matrix-multiplication kernels to Huawei’s Ascend platform while keeping the same API used by DeepGEMM elsewhere. Kernels are the small, highly optimized programs that perform much of the repetitive mathematical work inside AI training and inference.
DeepSeek also added Huawei support to TileKernels, which can select either an Nvidia or Ascend backend while exposing the same Python APIs to developers.
That gives developers a common interface across two hardware families. In practice, code written against that interface has a better chance of moving between them without engineers rewriting everything underneath it.
But portability at the API layer is different from parity underneath it.
This effort actually started a year ago
The September 30 announcement can sound as though DeepSeek and Huawei suddenly produced a new alternative to CUDA. The timeline is more gradual.
TileLang-Ascend was already open-sourced in September 2025, giving developers a domain-specific language for writing high-performance workloads on Huawei processors. Today’s release is better understood as an expansion of that effort.
DeepSeek is taking more pieces of the infrastructure used around AI workloads and adapting them to Huawei hardware. An ecosystem becomes more useful as developers get more of the components they already depend on—not simply because someone announces another programming tool.
There is another wrinkle, too. “Open source” does not automatically mean freedom from vendor dependence.
DeepGEMM-Ascend requires Huawei’s Ascend hardware and CANN, or Compute Architecture for Neural Networks, Huawei’s underlying software toolkit. TileKernels similarly lists separate Nvidia CUDA and Huawei CANN requirements depending on which backend is running.
Developers may gain flexibility at one layer of the stack while remaining heavily tied to the platform underneath it.
That makes the practical competitive question narrower: How much engineering work does DeepSeek’s software actually remove from choosing Ascend?
Independent developers have not answered that yet.
Kernel benchmarks are only part of the test
DeepSeek’s published DeepGEMM-Ascend results show selected operations running efficiently on supported Huawei hardware. The repository includes tests, benchmark results, and working code—substantially more evidence than a slide deck promising future support.
Those measurements still come from the project’s developers, and they measure individual kernels rather than the economics of running an entire AI system.
A lab buying compute cares about a much uglier spreadsheet.
Can the same model complete training reliably? How long does the job take? How much power does the cluster consume? How many engineers spend their week fixing hardware-specific problems? What happens when a framework changes? How often does the system fail? And, after all of that, what did the finished workload cost?
A technically successful port can still be an unattractive business decision if everything surrounding it becomes more expensive.
That is especially important because Huawei is competing at the system level, combining processors, networking, memory, and software into large clusters rather than asking customers to compare one Ascend processor against one Nvidia GPU.
The benchmark that ultimately matters is the workload.
There’s a credible case that the software could improve quickly
There is a stronger argument for DeepSeek’s release than today’s benchmark numbers alone.
Putting more of the stack in public repositories lets outside developers test it, find bugs, add hardware support, improve libraries, and expose failures that would otherwise stay inside one company. If enough developers adopt the tools, that feedback loop could mature Huawei’s software ecosystem much faster than its current limitations suggest.
DeepSeek itself has benefited from an open infrastructure strategy before, and Huawei has been investing in broader developer access around Ascend. More users could create more fixes, which attract more users, which create more fixes.
That is the optimistic case—and it is plausible.
What we do not know yet is whether that loop will become large enough to overcome the harder problems underneath it. Shared APIs do not guarantee equal performance. More contributors do not create additional chips. A growing developer community does not automatically make large training runs reliable.
The release creates conditions that could accelerate the ecosystem. The next test is whether developers actually show up and make it better.
Better software cannot solve a chip shortage
This is where DeepSeek’s software push runs into a very physical problem.
Less than two weeks before the release, Huawei said demand for its AI computing equipment exceeded what it could currently supply inside China. That makes hardware availability a constraint even before developers start debating benchmark results.
GitHub can make code abundant. It cannot make chips abundant, and better kernels do not manufacture more high-bandwidth memory.
That distinction matters because easier software and scarce hardware can exist at the same time. DeepSeek’s tools could make Ascend more practical for developers who already have access to the chips while smaller labs remain constrained by what hardware they can actually obtain.
Research group Epoch AI has also identified manufacturing and high-bandwidth-memory capacity as important constraints on Huawei’s ability to scale AI compute. Its projections depend on assumptions about future production and should be treated as estimates, but the underlying constraint is straightforward: software efficiency can stretch available compute; it cannot replace manufacturing capacity.
We’ve seen the same principle across the broader AI boom. As The Neuron’s AI Economics 101 explainer examined, some of the most important limits on AI progress increasingly sit in physical infrastructure rather than model architecture alone.
DeepSeek is trying to push one bottleneck further down the stack. The next one is harder to upload to GitHub.
Export controls make the hardware choice even messier
Chinese demand for alternatives to Nvidia also exists inside a policy environment that keeps changing.
U.S. export controls restrict Chinese access to many advanced American chips, giving companies a strong incentive to develop domestic hardware and software. But describing the market as completely sealed off from advanced Nvidia products would also be too simple.
In January, the U.S. Bureau of Industry and Security shifted Nvidia H200, AMD MI325X, and similar processors to case-by-case license review for exports to China when specified conditions are met.
For an AI lab making a multi-year infrastructure decision, that uncertainty becomes part of the calculation.
Performance matters. So do supply guarantees, compliance requirements, software maturity, engineering availability, and confidence that the hardware a company builds around today will remain accessible tomorrow.
Huawei therefore does not need to erase every Nvidia advantage before its ecosystem becomes strategically important. Reducing the cost and difficulty of maintaining another workable option could itself change how companies plan infrastructure.
How much it changes those decisions is the part that still needs evidence.
Here’s what would actually prove the shift is happening
Another repository release will tell us less about Huawei’s competitive position than what outside developers do with the repositories already available.
The strongest evidence would look more mundane:
- Independent end-to-end tests: The same model workload running on comparable Nvidia and Ascend systems, with training or inference completed successfully rather than isolated operations benchmarked.
- Real customer adoption: AI labs outside DeepSeek and Huawei choosing the stack for meaningful production workloads.
- Measured migration costs: Engineering hours, compatibility problems, debugging effort, and maintenance requirements before and after moving workloads.
- Delivered hardware: Actual chip availability and customer deployments, not planned capacity or demand estimates.
- Total operating cost: Hardware, power, networking, reliability, staffing, and software maintenance measured together.
Those results could still favor different platforms for different jobs. Training a frontier model, serving millions of inference requests, and running a smaller private model are different workloads with different bottlenecks.
And that may ultimately be the most consequential part of DeepSeek’s release.
For many AI developers, Nvidia’s advantage has been reinforced by the fact that choosing its hardware also meant staying inside the software stack their teams already knew. If DeepSeek and Huawei can make that software layer more portable, developers get to push more of the competition down toward hardware availability, reliability, and economics.
That would be a genuine shift even if Nvidia remained stronger across major workloads.
The first convincing sign will not be another declaration of technological independence. It will be an outside AI lab choosing Ascend for an important workload, finishing the job reliably, and deciding the numbers were good enough to do it again.
