A simple buyer‑side combinatorial double auction can deliver competitive prices and accurate price discovery for data, model and compute bundles once submarkets are moderately thick; legal or technical No‑Assembly is the dominant remaining friction, but its welfare cost falls quickly as complementarities and market size rise.

Main Finding

When goods are traded only as indivisible bundles and a platform cannot assemble a buyer’s desired bundle from components held by different sellers (“No Assembly”), a practical combinatorial Buyer’s-Bid Double Auction (BBDA) still inherits the large-market discipline of single-good double auctions under plausible conditions. Specifically, in sufficiently thick and stable bundle submarkets (and under a bound on how much a bid on one bundle can reassign traders across submarkets): - buyers’ strategic bid shading vanishes as markets grow,
- clearing prices concentrate at competitive levels and track the common-value component (price discovery), and
- the dominant residual friction at realistic market sizes is the structural loss from No Assembly (not strategic shading).
The result is proved for the two-good case under explicit stability and price-influence conditions; for larger item sets those conditions are stated as maintained assumptions. Multi-agent reinforcement-learning simulations corroborate the theory and show the combined welfare loss is small in moderately thick markets and declines with stronger complementarities.

Key Points

• No Assembly: platforms may be legally/technically unable to combine components from different sellers into a single bundle. This restriction changes allocation possibilities compared with a planner who could assemble bundles.
• Mechanism studied: Buyer’s-Bid Double Auction (BBDA) extended to allow per-bundle scalar bids and asks; the auction selects welfare-maximising trades subject to No Assembly and clears each traded bundle at a single price.
• Information structure: Correlated Private Values (CPV): each trader’s value for a bundle = common component µ(S) + idiosyncratic shock εi(S); traders observe their own values exactly. Complementarities enter via buyer-side superadditive idiosyncratic synergies κi (and symmetric economies of scope for sellers).
• Submarket reduction: conditional on endogenous sorting of traders into bundle-specific submarkets that are thick and regular, each bundle submarket behaves like a single-good BBDA in the large-market limit. This underpins the asymptotic vanishing of shading and the concentration of prices.
• New combinatorial channel: a bid on one bundle can reassign marginal traders across submarkets, producing an indirect price-feedback channel absent in single-good markets. The core technical requirement is a bound on this reassignment (price-influence) so that small bid changes do not produce large price swings.
• Two-good theorem: the incentive reduction (no first-order additional distortion from bidding on bundles beyond the single-good logic) is proved at the derivative level for K = 2 under explicit stability and price-influence conditions. For K > 2 the same reduction is taken as an explicit maintained condition.
• Complementarities help: stronger synergies encourage traders to specialise in offering full bundles, which thickens bundle submarkets and reduces both the structural loss (No Assembly) and strategic loss (shading).
• Simulation findings:
  • Welfare benchmark decomposition: (i) central planner with assembly allowed, (ii) truthful bidding in BBDA (isolate structural No-Assembly loss), (iii) strategic learning equilibrium in BBDA (isolate shading loss).
  • Structural (No-Assembly) loss is typically larger than shading loss at realistic market sizes but both are small and shrink with market thickness and complementarity.
  • With modest market thickness (roughly ~8 traders per side in the simulations) combined losses are small (a few percent of available surplus at moderate size).
  • Sellers’ behaviour remains essentially truthful; buyers exhibit shrinking shading with market growth.
  • Robustness checks: an independent optimisation method reproduces the learning outcomes and no trader could gain materially by deviation from learned strategies.

Data & Methods

• Theoretical analysis:
  • Model: finite sets of buyers and sellers, K indivisible items, each trader can bid/ask on at most one bundle (unit-supply restriction). Values: µ(S) common component + εi(S) private idiosyncratic shock; synergies κi add superadditivity/subadditivity.
  • Mechanism rules: sealed-bid, uniform-price batch BBDA that clears per-bundle trades and enforces bundle-by-bundle balance (No Assembly).
  • Core theoretical device: submarket reduction — conditional on stable, thick bundle submarkets, reduce combinatorial problem to parallel single-good BBDA problems.
  • Asymptotic tools: order-statistic asymptotics and Z-estimator theory for price convergence; LP formulation for allocation (total unimodularity ensures integrality); existence and monotone-equilibrium machinery via Reny and comparative-static results drawing on Milgrom & Shannon.
  • Price-influence and stability conditions: explicit for the two-good case (technical bound on reassignment feedback); for K>2 these are stated as maintained assumptions.
• Simulations:
  • Multi-agent reinforcement learning (MARL) to compute (approximate) Bayesian-Nash equilibria of the combinatorial BBDA when closed-form equilibrium is intractable.
  • Parametric setup: CPV model with synergy parameter κi; in simulations κi drawn i.i.d. (Exponential family parameterised by mean ¯κ).
  • Benchmarks computed: planner-with-assembly optimal surplus; truthful-BBDA surplus; learned-strategy-BBDA surplus. Welfare gaps decompose structural vs strategic losses.
  • Robustness checks: an alternative optimisation-based method reproduces MARL results; unilateral deviation tests show near-equilibrium behaviour.

Implications for AI Economics

• Practical market design for AI assets: For markets trading complementary AI assets (datasets, fine-tuned models, compute bundles), a simple buyer-side combinatorial BBDA is a credible, low-information, implementable mechanism once a modest number of participants are present per bundle submarket.
• Focus design effort on market thickness and legal/technical assembly constraints:
  • Increasing the number of buyers/sellers per bundle (market thickness) and encouraging complementarities (e.g., standardised full-bundle offerings) reduce both structural and strategic inefficiencies.
  • The dominant friction at realistic scales is No Assembly itself (legal/licensing/data-residency/federated-learning constraints) rather than strategic shading. Policy or platform measures that relax No Assembly (allow safe, legal assembly or coordinated licensing) will often yield larger welfare gains than efforts aimed solely at strategic mitigation.
• Pricing and price discovery: BBDA prices per bundle concentrate on competitive levels and reveal the common-value component as markets thicken, so platform-reported bundle prices can be informative signals for users and outsiders.
• Seller-side simplicity: because the auction pins prices on the buyer side, sellers have little strategic leverage over price and can be expected to behave near-truthfully; platform designers can simplify seller reporting rules without large strategic costs.
• Complementarities are not always a problem: contrary to an intuition that synergies make markets harder to organise, stronger complementarities can help by encouraging full-bundle offers that thicken bundle markets and improve outcomes.
• Limitations and caution:
  • The main formal proof covers the two-item case; extensions to larger K rely on maintained assumptions about the boundedness of cross-submarket feedback. Designers should therefore exercise caution when many items and complex bundling patterns exist, and empirical validation or pilot testing is advisable.
  • The informational environment assumes CPV (no signal noise). Introducing noisy signals (winner’s curse / Correlated Interdependent Values) is an important extension and could alter price-discovery dynamics.
  • The model assumes unit supply per trader (natural for exclusive licenses or one-off assets); different supply constraints would require reanalysis.
• Research and policy directions:
  • Empirical testing/pilots of combinatorial BBDA in AI/data marketplaces to validate the required market thickness threshold and the size of No-Assembly losses in practice.
  • Legal and technical work to safely enable assembly (or verified interoperability/licensing) where welfare gains would be large.
  • Theoretical and computational extensions to more than two goods under explicit, verifiable bounds on price-influence and to CIV environments.