OpenAI could use less hardware to get similar performance if they used the Chinese version, but they already have enough hardware to run their model.

Theoretically the best move for them would be to train their own, larger model using the same technique (as to still fully utilize their hardware) but this is easier said than done.