>copies of SOTA models that only take 20% of the resources

They might be 20% of the price (because they don't have to invest that much in training), but are probably not 20% of the resources (ie. inference), considering they take more tokens to do the same task, and have slower inference speeds.

https://x.com/scaling01/status/2050616057191072161

Even at 2x the tokens (max from that tweet), that makes them 40% of resources. Which is still only 40% of the resources.