New auction model proves AI asset markets price fairly without bundle assembly

Zachariadis and Yang's new paper tackles a critical question for emerging AI asset marketplaces: how do you run an efficient double auction when goods are only valuable in bundles—like combining a dataset with a fine-tuned model—but the platform legally or technically cannot assemble those bundles from different sellers? The authors formalize this 'No Assembly' constraint in a combinatorial buyer's-bid double auction. They show that under explicit stability and price-influence conditions, each individual bundle submarket behaves like a classic single-good double auction: bid shading vanishes, clearing prices concentrate on competitive levels, and prices accurately discover the common value. The key incentive proof holds rigorously for two goods, with larger item sets left as a maintained condition.

The paper's real-world message comes from multi-agent reinforcement-learning simulations. They decompose welfare loss and find that No Assembly—not strategic behavior—is the binding friction in finite markets. And that loss is small in moderately thick markets, shrinking further as complementarity between goods increases. With 67 pages, 8 figures, and 6 tables, this is a deep theoretical contribution with immediate practical relevance for any company building a marketplace for training data, model weights, or fine-tuned AI assets.