Why Selling AI Models Gets Worse Over Time

Jigar Doshi lays out the economics of frontier model decay with hard numbers: GPT-3 held its lead for 30 months, GPT-4.5 lasted five. Margins are collapsing in parallel - GPT-3 recovered 3.6x its costs, while GPT-5 projections sit at 1.1x. Inference already costs OpenAI 26 times what training does, and Cursor's billion-dollar ARR goes almost entirely back to API fees. The argument is that labs need to shift from selling compute access to monetizing what AI discovers - drug candidates, proprietary datasets, synthetic biology outputs. OpenAI's partnership with Ginkgo Bioworks on 36,000 protein synthesis reactions points in exactly that direction. Isomorphic Labs and DeepMind are making similar bets, treating models as research instruments rather than products.