Chinese AI leader DeepSeek has made a significant strategic move, signaling its next frontier model will launch by prioritizing native Chinese hardware like Huawei's Ascend chips.
This decision to defer support for Nvidia's CUDA and AMD's ROCm platforms is not a technical limitation but a calculated response to the intense U.S.-China tech rivalry. The timing is critical, coming just days after U.S. officials alleged that DeepSeek trained this very model on banned high-end Nvidia Blackwell chips. Launching with CUDA support would have implicitly validated this claim, escalating legal and political risks. By sidestepping U.S. technology at launch, DeepSeek effectively de-risks its debut.
Furthermore, this move aligns perfectly with Beijing's long-term goal of technological self-sufficiency. DeepSeek has already proven its capability to operate on domestic hardware. In 2025, it successfully shipped model variants with day-one support for Huawei's CANN software stack. This established a credible alternative to the CUDA-dominated ecosystem, making a 'CANN-first' launch not just a political statement, but an operationally viable strategy. It’s a classic case of turning a constraint into an opportunity.
Adding to the complexity is the uncertain status of Nvidia's export-compliant H200 chips in China. While Beijing granted 'conditional approvals' for their import, no major purchase orders have materialized. This ambiguity reduces the immediate commercial incentive for DeepSeek to invest resources in optimizing for CUDA at launch. Why prioritize a platform whose hardware availability is uncertain? The decision reflects a pragmatic hedging against supply chain and policy risks.
Ultimately, DeepSeek's announcement is a watershed moment. It highlights how U.S. sanctions, intended to slow China's AI progress, are instead accelerating the development of a parallel, non-U.S. technology stack. While the immediate financial impact on Nvidia and AMD might be limited, the long-term implication is the fragmentation of the global AI ecosystem, with China determinedly carving out its own path.
- Glossary
- CUDA: A parallel computing platform and programming model created by Nvidia. It allows software developers to use a CUDA-enabled graphics processing unit (GPU) for general purpose processing.
- CANN (Compute Architecture for Neural Networks): Huawei's software stack for its Ascend AI processors, designed as an alternative to Nvidia's CUDA for developing and running AI applications.
- Nvidia Blackwell chips: Nvidia's next-generation, highly advanced GPUs designed for AI and high-performance computing, which are subject to strict U.S. export controls to China.