> If the hyperscalers stopped building new infra tomorrow, they would just roll in the profits available from what exists today!

They owe hundreds of billions in loans on datacenters that aren't built yet and don't currently have capacity for the new users and additional tokens they promise are coming. If they don't build capacity the increase of demand would cause more outages and higher prices which would crater both their use and valuation. They have to grow or die.

> They intend to offer about $100B in stock, their valuation is at 1T

That's one company, the aggregate debt of the US AI industry approaches 1T.

> Most AI companies are profitable aside from expenditures to build new infrastructure

Which they have to complete to make the profit they promised which they need to pay off the loans they already have.

> A good GPU to run a reasonable model is $10k+

Today. Due to supply and demand. Funny thing, supply and demand. It's very reliable. High demand makes supply short for a while, but then causes supply to go up. When supply goes up, eventually demand goes down, which brings prices down. The chip factories and economies of scale that will be complete in a few years will send prices crashing, as they have before. Which might bankrupt the remaining chip companies.

"Reasonable model" isn't a thing. Many different models of different size have capacity to do many tasks. I have a laptop from 3 years ago that runs local AI models. My phone from 5 years ago runs local AI models. And any strong CPU can run models (no GPU needed). Speech to text, image/video generation, image classification, research, analysis, computer use, troubleshooting, etc, all works on tiny models. I think you don't understand the tech here. Remember that many people spend a grand every year just for a newer iPhone/Samsung. They'll spend the same for one with better AI and more RAM.