Why China's Top AI Models Still Train on Nvidia Chips

China's leading artificial intelligence developers — including the teams behind DeepSeek, Kimi K3, and Alibaba's Qwen models — remain heavily reliant on Nvidia chips for training their most advanced large language models, despite intense government pressure and US export restrictions . The core barrier is not hardware performance but software ecosystem lock-in: Nvidia's CUDA platform has become the de facto standard for AI development, and migrating to Huawei's Ascend chips with the CANN toolkit imposes prohibitively high engineering costs .

The CUDA Moat: 15 Years of Developer Dependency

Nvidia's CUDA (Compute Unified Device Architecture) enjoys a 15+ year head start with mature libraries (cuDNN, TensorRT), debugging tools, and a global developer community . Almost all major AI frameworks — PyTorch, TensorFlow, JAX — are optimized for CUDA out of the box .

By contrast, Huawei's CANN (Compute Architecture for Neural Networks) toolkit, even after being open-sourced in August 2025, remains far less mature in terms of operator coverage, optimization, and community tooling . Developers report that many CUDA-optimized kernels have no direct CANN equivalent, requiring manual rewriting .

This self-reinforcing cycle is difficult to break: the larger Nvidia's developer community and the richer its library of pre-optimized code, the harder it becomes for alternatives to gain traction .

The Cost of Switching: At Least 50% More Time and Money

An AI researcher at a Shanghai-affiliated institute estimates that switching training from Nvidia CUDA to Huawei Ascend would add at least 50% in both time and costs . Industry sources confirm this figure as a realistic floor .

This includes:

• Initial porting effort: Training code written for CUDA uses architecture-specific kernel calls, memory management patterns, and parallelization strategies that do not directly translate to Ascend's different NPU architecture .
• Ongoing debugging and tuning: Once ported, models must be re-optimized for performance parity on CANN's smaller operator library .
• Lost productivity: Developers must learn a new toolchain, and the smaller CANN community means fewer shared solutions to common problems.

DeepSeek's Abandoned R2 Plan: A Concrete Case Study

Perhaps the most telling example of the software gap comes from DeepSeek, one of China's most prominent AI labs. DeepSeek reportedly dropped plans to train its R2 model on Huawei's Ascend NPUs, citing unresolved performance and toolchain issues . This is a concrete case of a top-tier lab concluding that the domestic alternative was not ready for their most demanding workloads.

Performance Gaps Remain

While Huawei's Ascend 910C NPU has improved significantly, developers report it still underperforms Nvidia's Hopper-generation GPUs on large-scale distributed training tasks . The Ascend 910C, built on SMIC's enhanced 7nm process, delivers about one-third of the BF16 throughput of Nvidia's B200 . Huawei's upcoming Ascend 950 is expected to narrow this gap further, with Morgan Stanley data showing the performance gap between Chinese chips and Nvidia's export-compliant hardware has narrowed significantly .

The Supply Paradox: Stockpiled Nvidia Chips Create Inertia

Paradoxically, US export restrictions have not fully cut off Nvidia supply to China. The US approved the H200 for sale to China, and Beijing subsequently blocked its acceptance — but Chinese labs already hold massive stockpiles of Nvidia chips . ByteDance, Alibaba, and Tencent collectively spent an estimated $16 billion to stockpile roughly 1.3 to 1.6 million H20 units in anticipation of a ban .

This installed base creates powerful inertia: as long as existing Nvidia hardware continues to work, the high switching cost delays the move to domestic alternatives .

The $295 Billion Response: China's Domestic Chip Push

China is not standing still. In response to these barriers:

• Chinese chipmakers aim to triple domestic AI chip output in 2026
• Beijing is reportedly drafting a ~$295 billion (2 trillion yuan) plan to link every major Chinese data center into a single AI computing grid by 2028, with at least 80% of the underlying chips required to be domestic
• Huawei open-sourced CANN in August 2025, releasing core components including operators, domain libraries, and graph-engine technology
• ByteDance signed a $5.6 billion contract with Huawei for Ascend AI chips in December 2025, with deliveries starting in Q1 2026

Progress is real: domestic chip training share reportedly grew from 8% in early 2025 to 42% by April 2026 . But much of this adoption appears concentrated in less critical workloads and inference tasks rather than frontier model training .

The Outlook: A Multi-Year Transition

Huawei's open-sourcing of CANN is a meaningful step toward building the developer ecosystem that CUDA enjoys . But most analysts agree the software ecosystem gap — not raw hardware specs — will take years to close . In the short term, China's top AI labs will continue training on Nvidia chips while experimenting with hybrid domestic deployments for less critical workloads, and while Beijing's $295 billion grid plan aims to accelerate the transition .