If large models were primarily used for distillation to make smaller models better, that could be a more efficient use of resources. Smaller providers could choose to fund a larger model and benefit from it.