China's AI Large Models Still Rely on Nvidia Chips as Domestic Alternatives Face Software Ecosystem Constraints
Miles Bennett
China's most advanced AI models are still trained primarily on Nvidia chips, with the core barrier to domestic substitution lying not in hardware but in software migration costs — a hurdle that may prove harder to clear than building the chips themselves.
How far has domestic chip substitution actually come?
Multiple insiders at Chinese LLM developers told the South China Morning Post that Nvidia chips remain the mainstay for training the most advanced AI models.
This means → despite steady hardware progress, the actual "lead-actor swap" in production pipelines has not happened.
One person familiar with the industry put it plainly: training large models on Nvidia chips is still the norm among Chinese AI developers.
What is really blocking the switch?
The core obstacle is not chip performance — it is the software ecosystem. Nvidia's CUDA — a programming platform that lets developers tap GPU computing power — is the industry standard; the vast majority of existing code is built on it.
Huawei's alternative, CANN (Compute Architecture for Neural Networks), requires developers to rewrite and re-optimize large amounts of code; CUDA code cannot run directly on Huawei's Ascend chips.
In plain terms = you can swap the chip, but the millions of lines of code written around the old chip do not move with it — that is the real migration cost.
How steep is the migration cost?
James Wang, an AI developer at a Shanghai university-affiliated research lab, estimates that migrating his team's workflow to Huawei Ascend chips would increase time and costs by at least 50%.
This means → for resource-constrained teams, "going domestic" is not a hardware swap — it is a full rebuild of the training pipeline.
This reflects a deeper issue: switching chip architectures creates an engineering bottleneck, not merely a procurement decision.
What to watch next?
Whether migration costs fall as the domestic software ecosystem matures is the key test of whether homegrown chips can meaningfully replace Nvidia.
In plain terms = the hardware catch-up is the first half; matching the software ecosystem is the final — and that final has not yet begun.
Content is for reference only, not financial advice.