I think OpenAI's limitations are quickly verifiable results. 10,000 agents running in parallel aren't exactly talking to each other. They have a goal and one eventually hits the goal which can be automatically confirmed. This is why math is currently attacked, lean is easy to confirm A = B.

They aren't gonna create room temperature superconductors because they'll need a lab that can quickly, and verifiably, test superconductors at 10,000 per minute and that's just a physical limitation no matter how many datacenters you throw at it.