In AI model marketplaces providers differ in cost and capability, creating an opening for computational arbitrage. An intermediary can buy inference from a set of models and resell it at a higher price, earning profit. Besides the arbitrageur’s gain, consumers and providers also receive value, so the practice could reshape the market. Prior work shows feasibility on small model pools, yet large markets demand strategies that are both tractable and cheap to deploy. We formalize arbitrage as an allocation problem and prove it is NP‑hard in general. We then show that under a natural diminishing‑returns condition on how performance scales with inference spending, the optimal allocation can be computed in polynomial time. To cut the upfront cost of finding arbitrage strategies we propose a sample‑efficient method for estimating model performance from limited data, dramatically reducing required samples. Our analysis yields several economic insights: at optimum each funded model provides equal marginal utility per dollar; arbitrage shifts revenue away from providers whose utility is low relative to price; and it rewards complementary specialization. Review