The New York Times reports that DeepSeek and Huawei are working to erode one of the central pillars of Nvidia's dominance in artificial intelligence: the software ecosystem that keeps developers tied to its hardware.

Nvidia's market position rests not only on the raw performance of its GPUs but on CUDA, the proprietary software platform that has become the default environment for training and running AI models. Developers who build on CUDA effectively commit to Nvidia silicon, and that lock-in has allowed the company to command premium pricing across its data-center product families. Any credible challenge to that position would have to come through software as much as through chips.

That is precisely where the two Chinese companies are directing their effort, according to the Times. DeepSeek, the startup whose low-cost model training runs drew global attention this year, has an engineering incentive to reduce its dependence on Nvidia's stack, particularly as US export controls narrow its access to top-tier GPUs. Huawei, meanwhile, fields its own AI accelerator line in the Ascend family and has been building out the supporting software layer, CANN, as a domestic alternative to CUDA.

The geopolitical context matters here because it directly shapes the commercial one. Washington's successive rounds of export restrictions have cut Chinese buyers off from Nvidia's most advanced accelerators, pushing cloud providers and AI labs in China toward domestic hardware. A software ecosystem that lets those customers move workloads off CUDA would lower the switching costs that currently protect Nvidia even in markets where its chips cannot legally be sold at the high end.

The Times frames the effort as targeting a "key source" of Nvidia's dominance rather than its silicon itself — a distinction worth noting. Huawei's Ascend accelerators remain behind Nvidia's flagship parts on raw performance, and DeepSeek's influence comes from model engineering rather than chip design. But if the two can make non-CUDA software a practical path for large-scale AI workloads, they would attack the part of Nvidia's franchise that export controls cannot reach: developer habit.

The outcome is far from settled. CUDA has well over a decade of accumulated tools, libraries and optimized code behind it, and ecosystems built on that kind of incumbency tend to shift slowly. Still, the combination of a model developer with proven cost discipline and a hardware vendor with state-backed capacity gives China its most credible attempt yet at breaking the software lock-in — and the coming year will show whether developer migration follows.