But that lets companies that aren't colluding to catch up technologically and maybe sell their services for a lower cost. How is that anti-competitive?
This protects them from continuous model distillation by inferior models. So they only release a new model when a 3rd competitor catches up to their public production models' capabilities.
And the public gets a new model in 6 months or a year instead of every few months. But the new model would be a (1) gigantic leap in undisputed capabilities, (2) catastrophic in terms of job losses/replacements worldwide and they've had a year to prepare for the politics of it, and (3) the public would have access to the "tiny/medium/..." models while the big money/gov clients get the real super juice hidden behind the gates.
There is little to no political will for preventative measures. Conversely, there is an overabundance of political will to react to disasters.
Thus, there are preventionists whose plan B strategy is to write legislation and wait for a disaster to leverage newfound will to get it passed. Better late than never.
All that said, there's no political will to prepare for the politics of AI. It doesn't exist. The best we can accomplish is to be ready with a plan for when disaster hits.
They can make a lot of money by cutting down on R&D and inference costs until competitors can catch up.
They can also buy competitors that catch up.
And if they can get the government to regulate in the meantime, it’s party time for the monopolists!