DeepSeek's Computing Power Gambit: A Shift to Huawei's Ecosystem

In a significant strategic announcement, DeepSeek founder Liang Wenfeng revealed that the company will prioritize Huawei's Ascend series chips as the primary computing platform for training its future AI large models. This decision initiates a substantial technical migration—moving the core training infrastructure from NVIDIA's CUDA-dominated ecosystem to Huawei's self-developed CANN architecture.

The Imperative for Domestic Computing Power

Liang framed this transition as a necessary response to the evolving global technological landscape. With increasing export controls on high-end GPUs, reliance on imported chips poses growing supply chain risks. Building a self-sufficient, domestic computing foundation has become a strategic imperative for China's AI industry.

"To ensure long-term autonomy in computing power, this technological path must succeed," Liang emphasized, acknowledging the challenges ahead while affirming the direction as non-negotiable.

The Migration Challenge: Costly Underlying Reconstruction

Technically, this shift is far more complex than a simple hardware swap. It requires extensive rewrites and adaptations of deep learning frameworks, operator libraries, compilation toolchains, and other low-level code—a massive engineering undertaking with substantial R&D investment.

Industry sources indicate Huawei's next-generation training chips are scheduled for delivery between late 2026 and early 2027. DeepSeek plans to leverage this domestic computing power for training its next-generation models, creating a tight timeline for software ecosystem preparation before hardware arrival.

Opportunities Amid Uncertainty

Analysts note that if DeepSeek successfully migrates and efficiently runs large-scale models on Huawei chips, it could become a landmark case for "software-hardware co-design" in China's AI sector. This would demonstrate domestic computing power's capability to support cutting-edge AI research while providing a viable alternative path for the industry.

Yet significant uncertainties remain:

  • Ecosystem Maturity: Huawei's CANN ecosystem still trails NVIDIA's CUDA in developer tools and community support
  • Performance Optimization: Achieving training efficiency comparable to or better than previous platforms
  • Delivery Timelines: Alignment between chip production schedules and development plans
  • Long-term Evolution: Sustainable iteration of the domestic computing system

Regardless, DeepSeek's bet has accelerated China's pursuit of computing power autonomy. Its outcome will influence not just one company's trajectory but potentially the pace of domestic AI infrastructure development nationwide.