Yeah but that is what most companies were buying these pricy chips for. Training is a one time thing, especially if at some point you have compressed all available quality data (books,papers, etc.) into the system. What should additional training lead to? If the interference side runs on a smartphone (which is now possible in the future) than all these data center investments were for the bin.
And even if you don't trust the deepseek model, i am very sure that OpenAI and Meta will produce a model in relative short timeframe that will be as efficient as the deepseek model, because it is open source and everything of it can be copied.