His analysis fails to consider trend lines in inference costs. A couple of years ago inference required a supercomputing cluster. Today, you can run quantized frontier tier models on video cards with a $1600 MSRP. At current trends we'll be running frontier models on a typical gaming computer setup in a couple of years. That is going to destroy centralization possibilities.

It also fails to consider the impact of 'good enough.' A lot of the build-out right now is in anticipation of trying to push the frontier level. But there comes a time, a time we're already arguably past, where models being used for a specific task will already be able to complete that task efficiently. Going way beyond the present should enable LLMs to solve ever more esoteric problems, but in economic terms if a company's goal is to harvest corn, and your corn harvester works well enough, they're not going to care whether or not that corn harvester can also solve Fugazi's 7th Great Conjecture in exomathematics.