In a sense it doesn't, in that if DeepSeek can do this, making OpenAI-type capabilities available for Llama-type infrastructure costs, then if you apply OpenAI scale infrastructure again to a much more efficient training/evaluation system, everything multiplies back up. I think that's where they'll have to head: using their infrastructure moat (such as it is) to apply these efficiency learnings to allow much more capable models at the top end. Yes, they can't sleep-walk into it, but I don't think that was ever the game.