> It looks exactly like the two railroad bubbles. Potential profit is there, in the future. But you have to invest in the infrastructure first. If you can't make a profit and pay back the loans fast enough, the infrastructure goes bye-bye, along with all the funds used to finance it.

That's just not true. Profit is there now. If the hyperscalers stopped building new infra tomorrow, they would just roll in the profits available from what exists today!

Totally unlike the railroad bubbles which built out years and years ahead with no customers and then collapsed.

> "There's Still Plenty Of Capital To Raise" - IPO is not how you raise a trillion dollars. A million, sure. A billion, maybe? But a trillion puts retail investors - mom and pop - at risk. IPO is not how you raise funds for a railroad.

You're totally confused about what an IPO is. You're confusing the valuation of Anthropic with what they will raise. They intend to offer about $100B in stock, their valuation is at 1T.

> "Once AI companies become profitable" - there's the big red flag. This isn't Uber or Lyft that can float along on tens of millions in corporate hand-outs every quarter. They need 100x that much money, and nobody has it to hand out. So you can't wait around for profits. There's no magic cushion to float on.

Most AI companies are profitable aside from expenditures to build new infrastructure to reap more profits. That's a very healthy sign.

> "For starters, the average person isn't going to use a locally hosted model, ever" - They will in 4 years. "normal users don't want to have to deal with the hassle or expense of setting up something" - What hassle? You download a program with a fancy GUI and run it. There are a half dozen of these projects already, that provide ChatGPT on your computer from one binary. All you need is a GPU.

Uhuh. A GPU. In your phone? In your laptop? The average laptop in the US is sold for about $700, desktop for about $1k. A good GPU to run a reasonable model is $10k+. It makes zero financial sense for most people to run their own inference. The issue isn't, and never was, the software, it's the hardware. And one thing we've seen from the past years of growth is that people don't look at the models from 3 years ago and say "Oh well, that's good enough, at least I can run it slowly on my $5k+ desktop". Until that happens this is a fantasy.