Training new models is already the domain of large actors only, simply due to the GPU requirements, which serve as a massive moat. That ship has sailed. There isn’t a single open source model, today, that wasn’t trained by a corporate entity first, and then only fined tuned by the community later.