DeepSeek has shown that significantly less costly training is possible when incentivized. Even for SOTA models.