[qeternity]

This is not the case for LLMs. FP16/BF16 training precision is standard, with FP8 inference very common. But labs are moving to FP8 training and even FP4.