What AI training data sells for: a sourced index of public prices, 2023 to 2026

Abstract

Training data for machine learning is bought and sold without an exchange or a posted price list. This paper assembles the public record of its prices: 1,022 observations for nine kinds of data, read from 735 source documents between 11 and 17 October 2026, 1,009 of them dated 2023 or later. Each observation keeps the figure as published, the side of the market where the money moves (paid to makers, offered to holders, asking price, made to order, completed sale, cost to produce) and a conversion to a common unit by stated arithmetic, with ranges entered as geometric midpoints. Medians in each kind's main unit run from $0.525 per image to $316,228 per company archive, and within a main unit the upper quartile is 2.0 to 37.8 times the lower. The side of the market moves prices as much as the kind of data does. For agent-training tasks, the median price buyers pay for collection to order is 9.5 times the median paid to the people who make them. For company records, the median published offer to holders is 11.2 times the median of the four completed sales with a stated amount. Completed sales are 73 of the 1,022 observations, and only 11 observations come from court filings. We describe the disclosure, promotion, concentration and survivorship biases of a public-record compilation and state the method so that every figure can be recomputed. JEL codes: D40, D83, L14, L86, O33

Keywords: training data, data markets, price index, data licensing, artificial intelligence, market transparency, posted prices