Why smarter AI models could drive up compute prices 10x

Dwarkesh Patel's post is explicitly a two-hour, time-boxed experiment, and he says so at the top: many sub-questions will not be nailed down, but the alternative is no progress on the topics he is curious about. That framing matters for how the argument should be read. It is a chain of arithmetic and market observation rather than a fully specified model, and its value lies in the single comparison it forces on the reader.

That comparison is the anchor of the whole piece. If a human-level software engineer could run on an H100-class GPU, then at current market rates for software engineers, that GPU should rent for over $250,000 a year — roughly fifteen times today's spot price. Everything else in the post is an attempt to work out what has to be true for that gap to close, and what closing it would do to the industry.