Priced at Zero: The Data That Cannot Move, What It Is Worth, and What Stands Between It and a Market

Abstract The market for AI training data prices the data that can be copied: a news archive, a set of published papers, a clinical and molecular database built to be licensed. Those deals run from tens to hundreds of millions of dollars, and each was possible because the seller was in the business of selling. The data that is likely worth the most, identified patient histories, a drug company’s failed experiments, a bank’s view of its customers’ payments, an agency’s raw reporting, has no price, because the law, the consent, or the promise under which it was collected means it cannot be sold as a copy, and nothing has existed that could sell it any other way. This paper estimates what that zero conceals and shows that it takes three forms: a missing revenue line in medicine and research, a loss no institution can shrink alone in finance, and a failure that ends in a public inquiry in national security. It reviews what has been tried: pooled de-identified databases, the sharing and clean-room services of the large data platforms, and consortiums that bring the computation to the data. The techniques work. The market has not formed, because every success so far depends on an operator or convener its members already trust. It sets out five obstacles that compound, custody law, the memory of models, the trusted intermediary, the ownership of outputs, and owners who have been asked the wrong question, and maps each to the seven requirements the first paper in this series set out for a layer beneath the work, naming what that layer answers and what it does not. It argues that the layer moves the cost of protecting shared data from legal work repeated relationship by relationship to infrastructure that enforces and proves the owner’s terms run by run, with the automotive industry as the worked example, and that the larger gain is time, because the months spent negotiating permission are months in which nothing is discovered. It answers eight objections, including that the incumbents will close the gap themselves and that the layer is only a new intermediary, and states what would show its thesis wrong. It concludes that consortiums run on trust while markets run on receipts. The party that makes the receipt both sides accept will not own the data. It will own the receipt.