Just like existence of MariaDB does not prevent Snowflake from being worth $50B, just being good enough on laptop is not enough to replace the need for the cutting edge.

If this happens we will just get more things done with the same amount of compute (see: Blinn's law). The demand for GPUs does not really come from algorithmic compute requirements but from social expectation of progress in the field of AI. People will use all the compute they can get doing research using the budget they are given. What matters is how this budget is set.

Training is very expensive and requires GPUs. What you read about is running trained model on consumer devices (even phones!).

Let's assume for a moment you could sort of trade parallel computation for vast space, fast search and retrieval. So in this hypotheticals computational theory, you could build a lookup machine from CPUs and ssds.. squeezing the parallel cores into one CPU, by squeezing the shaders running into a million hashes.. And before you know it your simulating a micro verse trying desperately to find out how to avoid climate change. What if God hates recursion?

The current giants have shown that "it" can be done. From now on we can reasonnably hope for massive progress in efficency at the low end - as well as massive capability improvements at the high end. That goes both ways, probably.