This is a dynamic that I don't think has existed for a while. In the 90's there was a real need for the average person or business to require a faster computer

This is a dynamic that I don't think has existed for a while. In the 90's there was a real need for the average person or business to require a faster computer. For quite a while (~2005-2020), computers seemed to be approximately sufficiently powerful for the majority of tasks outside specialized/niche cases. Now, suddenly, LLMs and other ML models have crossed a threshold of everyday usefulness such that it is worth throwing more compute/memory at it. More resources may or may not give us better results with these resource-hungry models, yet it's surely the case that more resources allows more individual LLM agents to be deployed. If the marginal utility of an incremental agent is nonzero, then there'll be demand for more compute.