Dwarkesh Patel Argues AI Compute Could Get 10-15x Pricier

The pitch of Dwarkesh Patel's latest essay is a back-of-envelope thought experiment that gets uncomfortable when you sit with it. If an H100 could host a human-level software engineer, then at the salary a human engineer commands, the going rate for the GPU-hour implies that same H100 should rent for over $250k a year. His words. That is roughly 15x what H100 spot capacity trades for today.

The reason this is not just a thought experiment is that the trend lines already point in that direction. Spot prices for compute are up 40%+ from the February trough. Google is reportedly paying around $900 million a month for 110K GPUs, a blend of GB200s and GB300s, which works out to roughly twice the current spot rate per hour. Meanwhile Anthropic's revenue has 10x'd year over year, and Patel writes that the company likely ends the year with roughly $100-150B of revenue. Margins on Fable inference reportedly went from 40% in 2025 to above 80% this year, with blended inference margins over 70%. When a buyer is willing to pay 2x spot and a seller is capturing 80% margins on the sale, prices are almost certainly heading up.

The supply side is the pinch. Patel's decomposition of compute growth through 2030 comes out to roughly 3x a year: 1.4x from Moore's Law, 1.2x from new fab construction, and 1.8x from AI taking leading-edge wafer allocation away from other devices. If revenue compounds at 10x while capacity compounds at 3x, the gap has to clear somewhere, and price is the obvious release valve. That is the case for a 10-to-15x repricing of inference over the next few years.

Hold the estimate loosely. This is one writer's model, extrapolating from revenue figures reported by a single lab and a spot market that has been jumpy in both directions. 'Fable inference' is not really defined in the essay, so the margin figure is doing a lot of work without much context around what fraction of the mix it actually represents. And extending the 10x-a-year revenue curve to $1T by end of 2027 assumes the market absorbs AI services at that scale, which is itself the open bet.

If Patel is even directionally right, the quiet move is already happening: hyperscalers and frontier labs are locking down multi-year GPU capacity at today's rates, and the buyers most exposed are the startups whose unit economics assume cheap tokens will keep getting cheaper. Today's docket alone had Lambda drawing a $917M leveraged loan to lease Nvidia chips and Anthropic joining Macquarie and GIC on the Theseus data-center vehicle.