The small print in TFA states the many limitations of their method, i.e. that it actually provides no evidence about which GPUs were really used for training, and much less about who was their owner and where they were located.

The method just strongly suggests that the initial values of the weights, before starting training, were generated with the NVIDIA cuRAND library on a Blackwell GPU (or on an AMD GPU, where the NVIDIA library is emulated for compatibility).

The training might have been done with thousands of GPUs of other types. It is plausible that the GPUs used for training or at least a part of them may have been of the same type with the one used for initializing the model, but it is not proven in any way.

A scenario that matches what is reported in the article, but in which there is only a single Blackwell GPU involved, is if the LLM was intialized on the workstation of some lead developer, who had the privilege to use the best GPU they had, and then it was transferred to a server cluster for training, where other kinds of GPUs were used.

Now that this has been published, it would be very simple in the future for any Chinese AI company to avoid this "fingerprinting", by just using a different PRNG, instead of cuRAND.

Yes it's simple to avoid this fingerprinting now that people know about it, but nobody knew about it before. I don't believe people are casually transferring snapshots between chip architectures without a really good reason, either. It's hard enough to coordinate these training runs without involving multiple chip architectures in a single run.