SAN FRANCISCO — Chinese artificial intelligence developers have largely maintained their reliance on Nvidia chips for training large language models, despite a strong government push to adopt domestic hardware — held in place by the engineering costs of switching software platforms.
Nvidia's Compute Unified Device Architecture, or CUDA, is the industry-standard software ecosystem for AI development, offering a comprehensive suite of tools and libraries optimized for Nvidia's graphics processing units.
Huawei Technologies' alternative, Compute Architecture for Neural Networks, or CANN, requires developers to rewrite and optimize large sections of existing code. James Wang, an AI model developer at a research institute affiliated with a Shanghai-based university, said his team's training pipelines depend on CUDA. He estimated that migrating current workflows to Huawei's Ascend chips could increase both time and costs by at least 50 percent.
By 2025, Beijing had signaled a clear intent to reduce reliance on U.S. chip suppliers. State media began labeling Nvidia's China-compliant H20 chips as potentially unsafe or compromised, escalating pressure on domestic tech firms. Major Chinese tech companies including Alibaba and ByteDance were reportedly instructed to cancel new orders for Nvidia GPUs. The AI startup DeepSeek announced in August that its next model would be designed to run on China's domestic AI chips.
Huawei's Ascend line has emerged as the primary domestic contender. The Ascend 910B matured under U.S. sanctions and became a default option for companies unable to secure Nvidia GPUs. A Huawei official said the 910B outperformed Nvidia's A100 — the company's top chip from 2020 — by roughly 20 percent in some 2024 training tasks.
The 910B still uses HBM2E memory, an older high-speed memory standard. It holds about one-third less data in memory and transfers data between chips approximately 40 percent more slowly than Nvidia's H20.
Nvidia's competitive advantage stems from its integrated hardware and software stack. Its software ecosystem, combined with memory capacity and interconnect bandwidth, creates switching costs that domestic rivals have not yet overcome.
China is pursuing a two-track strategy: engaging Washington for access to advanced U.S. chips while building out domestic alternatives. Huawei is investing in rack-scale supercomputing clusters that pool thousands of chips to expand aggregate computing power. Domestic players including Alibaba, Baidu and Cambricon are also investing in an independent AI stack, though matching Nvidia's integrated hardware-software performance at scale remains the central challenge.
