Comment by decimalenough

Not surprised. You need massive amounts of compute for training LLMs from scratch, but practical uses of LLMs are increasingly focused on using existing models with minimal or no tweaks, which requires far less computing power.