China's most advanced artificial intelligence models are still being trained on Nvidia chips, according to a report circulating via MSN and aggregated by Bing News, which cites sources at major Chinese AI developers.
The reason, per that reporting, is not a lack of domestic alternatives but the cost of switching. High transition costs are keeping Chinese AI developers reliant on Nvidia hardware even as Beijing pushes the industry toward homegrown silicon.
That framing matters because it reframes a question usually treated as purely political. Export controls and national self-sufficiency campaigns get the headlines, but the report points at something more mundane and arguably more stubborn: moving a large model training run from one chip platform to another is expensive and disruptive.
Training a frontier AI model is a months-long, capital-intensive process, and the software that makes it work is tuned tightly to the hardware underneath. Swapping the underlying chips can mean rewriting and re-tuning that stack, absorbing delays, and risking worse results — costs that a company racing competitors may not want to pay mid-cycle.
The source item does not name specific Chinese firms, quantify the transition costs, or say how long the dependence is expected to last. Those details aren't in the reporting available here, so treat the specifics as open.
What is clear from the report is the direction of the tension: policy pressure pushes one way, engineering economics pushes the other, and for now the engineering is winning.
This matters because the effectiveness of chip export restrictions — and the timeline for China building a genuinely independent AI stack — may depend less on whether domestic chips exist than on whether developers can afford the disruption of actually switching to them.