These are the costs by the AI companies themselves to train their basic AI models (mostly LLMs), what isn't counted is how much it costs an average company to train the AI for their own systems. Why?
Not only do they realise the model training cost, but it also creates a stark awareness of their technology debt. Especially with regard to their data management, governance and quality of data.
Eventually, these issues will be dealt with. Maybe. Research is showing few companies are finding value from AI products sold by large tech companies. It's a Catch 22. The systems need to be trained, but the systems have to be in good order. Few are.
This may be the real challenge and hurdle to overcome.