I don't quite get it if that is bad thing for Nvidia and AMD. How would them be able to optimize their GPUs with a model?

This is getting played in (parts of) the US press like it's DeepSeek trying to hide the fact that it trained its model on "illegal" hardware but in fact it's really likely about trying to give Huawei an early advantage.

I don't get this.

Models have become a commodity, hardware not so much; why would Nvidia care? It's the models who are competing primarily, not hardware manufacturers.

Also, models are software, so easy to switch out. AI hardware otoh has a approx. 3-5 year useful life, I don't see how this affects hardware already running, or getting built.