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".