Who sells AI training data: a census of vendors, what they say they sell, and where their data comes from

Abstract

Firms that sell training data, expert labour, reinforcement-learning (RL) environments and evaluations to AI developers are mostly private and known largely through their own websites. This paper reports a census of 150 such vendors, read on their sites on 15 and 16 October 2026 with every claim tied to a quoted sentence, and links it to three companion tables: 37 organisations that publish a way to sell them data, 227 revenue records and 111 funding records. Evaluations are the most common offering (122 vendors, 81.3%), followed by datasets (104), RL environments (89), expert and human data (83) and annotation (57). Vendors bundle: the median vendor lists three of six kinds, and 79 of the 89 RL-environment sellers also sell evaluations. Most data are made to order: 90 vendors cite experts and contractors, 41 recordings, 30 generation and 15 software replicas, against 17 that license data from its owners and 22 that use the client's own data; 19 do not say. Of the 37 buyers, 24 take company records; among those that publish them, the median floor is $100,000 and the median ceiling $1 million per seller. Revenue disclosure runs from audited filings of four listed vendors to self-reported run rates, and 61 funding rounds with a disclosed amount total $4.30 billion. We describe the method, the classification and the limits of a self-description census.

Keywords: AI training data, data vendors, data annotation, reinforcement learning environments, model evaluation, data markets, company data, market census