The problem with GenAI is that the inference cost grows linearly with usage, and the constant creation of new models means the training costs are also not recouped.

probably even slightly more than linear at this point as they use up so much electricity and other resources that it becomes more and more expensive to serve more users. User 1 is easy, that's just the existing data center. User 1 million, well, you need a new datacenter and drive up hardware prices, but it's fine. User 10 million, you need to build/rent gas power plants, figure out where to get water from, RAM and everything costs 5 times as much as it used to, bribe local politicians to build despite backlash etc.