Depends on the timescale.

I have the feeling that smaller players are about as likely to get past GPT-n family in the next 2-3 years as I am to turn a Farnsworth Fusor into a useful power source.

Major technical challenges that might be solvable by a lone wolf, in the former case to reduce the data/training requirements and in the latter to stop ions wastefully hitting a grid.

But in 10 years the costs should be down about 99%, which turns the AI training costs from "major investment by mega corp or super-rich" into "lottery winner might buy one".

This tech is capital-intensive even when you know how to do it.

I heard estimates in tens of $M. That's rather available.

Isn't that quite a lot of other-than-personnel cost for a software startup? And how many iterations do you throw away before you get one that generates income?

I did not necessarily mean 10 people startups. There are quite a few companies smaller than OpenAI, but much larger than 10 people.