I was reading a synopsis of the amount of expected AI revenue (@90B/yr) vs the capital outlay to reach that point (@1T) on assets that depreciate far faster than other capital assets, usually in 18 - 36 months. There is no way for this to be a sustainable business with that sort of economics.
We're seeing inference costs skyrocketing because the expected reductions in tokens with greater LLM efficiency and better models needed isn't happening at anything close to the curve that was expected. IE: AGI isn't happening, it's not going to happen, and unless it does, the models aren't going to get exponentially better as is needed to make this all work.
So the AI companies make money selling rocks to each other in the desert, and someone is going to be left holding the bag. Guess who that's going to be? Privatize the profits and socialize the losses.
Unsurprisingly, a guy that should probably have retired about 20 years ago is getting smoke blown up his ass by his advisors about a technology he hasn't the foggiest clue about.
That's ML involving tightly constrained iterative processes and neural networks, and was happening long before LLMs. This whole bubble is LLM bullshit and assuming a word probability machine is capable of "discovering" anything is silly.