PoUW blockchains could subsidize AI inference without weakening security

A new paper by Cornell Tech's Rafael Pass, published on arXiv, provides a rigorous economic model of Proof-of-Useful-Work (PoUW) blockchains—systems where mining computation also performs valuable work like machine learning inference. The common criticism of PoUW is that “useful” work could reward attackers, making 51% attacks cheaper and undermining security. Pass develops a competitive equilibrium model with three allocation choices: pure mining, pure useful work, or duplex work (both simultaneously with overhead).

The model yields a closed-form characterization of equilibrium prices and allocations driven by a single parameter—the token-inference ratio—that measures token adoption relative to the inference market. Three distinct regimes emerge: 'Bitconia' behaves like classical Bitcoin PoW; 'Fortessia' sees duplex replace pure mining, increasing security without changing useful output; and 'Duplexia' where token rewards effectively subsidize inference, lowering prices and expanding supply. Crucially, the economic cost of a majority attack remains tied to the block reward, not the useful work value. In Duplexia, the blockchain generates socially useful computation that wouldn't exist otherwise, with expansion monotonically increasing in token adoption and technological efficiency.