Why auction design matters (ft. Nobel economist Paul Milgrom)

Paul Milgrom and Bob Wilson won the Nobel Prize in Economics and co-designed the FCC Spectrum Auctions, which generated over 100 billion dollars in revenue for the US government and enabled technologies including 5G wireless. Milgrom entered auction theory opportunistically, writing a term paper on auctions specifically to attract Wilson as his PhD advisor at Stanford, and his foundational ideas consistently emerged from observed field problems rather than abstract inquiry.

The Glosten-Milgrom paper overturned the prior consensus that bid-ask spreads reflect inventory holding costs, showing instead that spreads are driven by informational asymmetry because buyers signal positive information and sellers signal negative information to the market specialist. Empirical work confirmed that a large portion of the spread is attributable to information rather than inventory costs. The model also showed that transaction prices form a martingale because the information content of each trade is already embedded in the transaction price, and it produced results about how markets break down when asymmetric information becomes too severe. The paper has recently attracted a new audience of computer scientists and engineers working in decentralized finance.

The Milgrom-Weber paper introduced affiliated values, which became the standard framework for reasoning about auctions with correlated values. Affiliation captures the normal practical case where each bidder's assessment of value is informative about the actual value of the good, and it was mathematically necessary to verify that first-order conditions yielded true equilibria. Crypto auctions are considered a domain where affiliated values are especially relevant, as bidders draw inferences from others' bids about ecosystem value. Myerson's revelation principle was described by Milgrom as shocking to economists at the time, though Milgrom noted that recent work including from his own students has shown that focusing solely on direct mechanisms causes some properties to be lost.

Pacific Bell asked Milgrom in the early 1990s to review an FCC spectrum auction proposal he found deeply flawed, one that would have allowed a single buyer to acquire all spectrum nationwide. He and Wilson proposed a simultaneous multiple-round ascending auction, made seven trips to Washington to convince the FCC, and demonstrated the design using linked Excel spreadsheets on a floppy disk handed directly to FCC staff. The FCC hired Charlie Plott at Caltech to run laboratory experiments confirming the design worked and identifying what would go wrong with alternatives. An activity rule was added to prevent deliberate obstruction and ensure the auction converged in reasonable time. The first live test used seven paging spectrum licenses, raised approximately 100 million dollars, and produced identical prices for essentially identical licenses, validating the design. The simultaneous ascending auction spread worldwide after FCC adoption.

The FCC Incentive Auction, completed roughly eight to nine years before the recording, is described as one of the biggest market design triumphs of the 21st century. By around 2010 all the most valuable lower frequencies had been allocated, broadcast viewership had shifted to internet and cable, and digital encoding made old analog spectrum blocks more than necessary for high-definition signals. The auction cleared channels 38 to 51 of all television broadcasts and repacked that spectrum for wireless data use, ultimately clearing 70 megahertz for modern data services. The reallocation problem is structurally a graph coloring problem covering approximately 2,200 television stations and 130,000 constraints, and it is NP-complete, meaning no algorithm could solve it at the required scale. The software was expected to solve roughly 99 percent of instances, with one percent timing out, and Kevin Leyton-Brown estimated that improving timeout behavior saved the government approximately two billion dollars. Auctionomics received a technical Emmy Award for its work on the auction.

The auction linked a reverse auction buying broadcast stations to a forward auction selling wireless broadband licenses. Designers chose a descending clock auction for the reverse side to make participation obviously strategy-proof for small station owners. Prices in the reverse auction could only fall and clearing targets could only decrease, while reduced clearing increased competition in the forward auction and pushed forward prices up. The auction ended when forward revenues were sufficient to cover reverse payments. Designers expected religious and public broadcasters would refuse to participate, but KQED sold a third station for 95 million dollars after finding it could serve its full audience with two, and a religious broadcaster in Chicago received approximately half a billion dollars plus a VHF station, the highest price paid in the auction.

Milgrom identified a key financing problem facing data center builders: banks are reluctant to lend against assets facing rapid technological obsolescence, with construction costs reaching 20 to 30 billion dollars and a faster competing chip capable of quickly eroding collateral value. A futures market on compute rental prices would allow that obsolescence risk to be hedged more broadly, unlocking more debt financing for data center construction and in turn supporting greater chip sales. Milgrom noted that many assumptions in economics are arbitrary conveniences that are simply wrong, a realization reinforced by engagement with computer science, and that game theory has been revitalized by ideas from computer science after economists largely exhausted the economic ideas within it. On career advice, he told students to pursue rigorous technical training, hold themselves to high standards of evidence and logic, and prioritize flexibility over betting on any specific knowledge domain.