Is the compute for running an LLM cheap enough to scale at the moment? LLMs seem to be a great generalist solution but could specifically targeted NLP solutions

Is the compute for running an LLM cheap enough to scale at the moment? LLMs seem to be a great generalist solution but could specifically targeted NLP solutions still outperform in terms of speed/cost when you are processing high volumes of inputs?