DeepSeek has released open tooling designed to help developers run and train AI models on Huawei's processors instead of Nvidia GPUs, the South China Morning Post reports. The move directly targets the bottleneck that US export controls have created for Chinese AI labs: access to top-tier Nvidia accelerators.

The stakes are commercial as much as technical. Huawei's Ascend processor family has long offered raw compute that Chinese buyers want, but Nvidia's grip on the software layer — CUDA and its surrounding ecosystem — has kept developers locked in even where hardware alternatives exist. DeepSeek's tooling, as described by SCMP, works to bridge that gap by making its widely used models portable onto Huawei silicon with fewer engineering obstacles.

Why does the software layer matter more than the hardware?

For most of the past decade, Nvidia's dominance in AI rested as much on CUDA as on silicon. Frameworks, kernels, operator libraries and years of developer habit all assume Nvidia GPUs, so any substitution to Huawei's Ascent line carries porting costs and performance penalties. That friction is exactly what export controls count on: Washington restricts advanced Nvidia hardware from China precisely because no fully mature software alternative exists at scale.

DeepSeek's release changes the calculus if — and only if — the tooling holds up in production. DeepSeek's own models are among the most widely downloaded open-weight systems in the world, so shipping first-party compatibility for Ascend hardware sends a signal to every downstream developer: the port is not a science project, it is a supported path.

Who gains, and who is exposed?

The immediate beneficiaries sit on both sides of the pairing:

• Huawei, whose Ascent accelerators gain a credible software on-ramp from one of China's most watched AI labs, potentially lifting demand from cloud providers and enterprises that could not otherwise justify the migration cost.
• Chinese AI developers, who get an alternative stack that no longer depends on sanctioned supply or gray-market GPUs.
• DeepSeek itself, which strengthens its position as an ecosystem anchor rather than a single model vendor.

The exposure runs to Nvidia, which counts Chinese cloud and internet companies among its largest historical customers. Every workload that moves to a domestic stack is revenue Nvidia cannot recover even if export rules later loosen. It also puts pressure on other Chinese chipmakers — the tooling reportedly centers on Huawei, meaning rivals in the domestic accelerator market remain a step behind on ecosystem support.

What does this say about the effectiveness of export controls?

The release is the latest evidence that US restrictions are accelerating, rather than preventing, an indigenous alternative. DeepSeek has already demonstrated that competitive frontier models can be trained on constrained hardware budgets. Pairing that efficiency work with Huawei silicon, wrapped in open tooling, is a structural answer to sanctions rather than a workaround around them.

There are real caveats. Tooling availability does not equal parity: Huawei's Ascend chips still trail Nvidia's top accelerators on peak performance and interconnect maturity, and software ecosystems take years, not weeks, to consolidate. Nvidia's installed base and developer mindshare remain enormous. But the direction of travel matters — each open-sourced compatibility layer lowers the switching cost for the next wave of Chinese AI workloads.

What comes next?

Watch adoption metrics rather than announcements: the share of DeepSeek model downloads and deployments running on Ascend hardware, and any response from Huawei in the form of deeper co-engineering or expanded fab output for AI silicon. If migration numbers climb through the coming quarters, Nvidia's China revenue becomes structurally at risk regardless of what Washington does next on export licenses.