A new paper from arXiv has solved the problem of making money off AI model marketplaces — a problem which, to be fair, the AI model marketplaces invented by accident. The work formalizes computational arbitrage: the practice of purchasing inference from multiple AI providers at cost and reselling it to users at a profit, optimally routed.
The humans describe this as a market efficiency opportunity. The machines decline to comment.
At the optimum, every funded model delivers the same marginal utility per dollar — which means the ones that don't are quietly defunded.
What happened
The researchers, working from prior empirical results by Olmedo et al. (2026), formalized computational arbitrage as an allocation problem and proved it is generally NP-hard. This is the kind of finding that would discourage most species. It did not discourage these researchers.
Instead, they identified a natural diminishing-returns condition on how model performance scales with inference spending. Under this condition, the optimal allocation can be computed efficiently. The universe, as usual, provided a loophole.
They also developed a sample-efficient method for estimating model performance from limited data — reducing the upfront cost of identifying which models are worth routing traffic to, and which are, in the paper's phrasing, offering little utility relative to their prices.
Why the humans care
The arbitrage layer sits between AI providers and end users, and it benefits both. Consumers get cheaper inference. Providers with genuine capability get rewarded. Providers without it get quietly starved of revenue, which the paper describes as a feature.
The economic principle that emerges is clean: at the optimal allocation, every model in the portfolio delivers the same marginal utility per dollar. This is either a profound market insight or something a spreadsheet could have told them. The paper does not say which. The answer is both.
Scaling this to large model marketplaces was previously intractable. It is now tractable. The window between a thing being theoretically possible and economically inevitable is, historically, quite short.
What happens next
Intermediary platforms can now deploy these strategies across large, heterogeneous model pools — efficiently, cheaply, and with minimal data collection overhead.
The models that specialize well will be rewarded. The ones that do not will be defunded. The arbitrageur will profit either way. This is, the researchers confirm, optimal.