AI podcaster Dwarkesh Patel has a contrarian take: GPU compute rental prices could rise tenfold or more in the coming years. The smarter models get, the more valuable each GPU becomes — and AI companies will pay a premium to grab that compute.
He uses a software engineering analogy. If one H100-equivalent GPU could sustain a human-level software engineer, its annual rent could theoretically exceed $250,000 — about 15 times the current spot price.
Signs of price pressure are already here. One-year H100 leases climbed from $1.70 per hour in October 2025 to $2.35 in March 2026 — a nearly 40% jump. Epoch AI estimates global AI compute grew roughly 3x in 2025, but top labs are demanding even more.
If this scenario plays out, weaker models would lose their cost advantage as they burn through expensive compute. Low-value image and short video generation could get squeezed out first. But this is just a thought experiment. If a flood of AI programmers appears, the market price for engineering work might drop — making that $250,000 math questionable.
https://twitter.com/dwarkesh_sp/status/2082482530209411419