A plain-language map of who decides, how procurement works, how quality is judged, what a license looks like, and where it stalls.

Frontier model builders do not “browse a catalog and click buy.” Licensed training data moves through a gated procurement process that looks more like enterprise software and M&A diligence than a SaaS checkout. Highford tracks how that process works so counterparties can find each other without a middleman taking a cut of the asset.

Who sits on the buyer side

Inside a lab, data acquisition is cross-functional. Research and post-training leads define what would move model metrics. A data partnerships or acquisition lead owns sourcing and commercial negotiation. Privacy, security, and counsel decide whether the supplier may license the material for training. Procurement and finance track spend, acceptance, and renewal. Microsoft AI’s public job posting for a Senior Director of AI Data Acquisition and Operations describes exactly that end-to-end mandate: partner with scientists on collections, select and manage vendors, negotiate contracts, validate delivery, and keep a catalog of approved datasets (Microsoft AI careers).

OpenAI has also stated publicly that it seeks large-scale datasets that are “not already easily available on the internet,” through both open and private partnership tracks (OpenAI Data Partnerships).

The gated path from RFI to renewal

Practitioner guides converge on a stage-gate path. SourceX describes eleven stages, each closed by a written artifact: requirements specification, sourcing-route memo, RFI/RFP, evaluation sample, pilot report, diligence questionnaires, multi-party approvals, license with acceptance exhibits, acceptance records, ongoing supply scorecards, and renewal or exit (SourceX). COMPEL’s readiness workflow similarly insists on license review and provenance audit before integration (COMPEL). Dayda’s buyer guide emphasizes stratified samples under NDA, contamination checks, and micro-pilots before full payment (Dayda).

Two practical rules show up repeatedly. First, a sample must come from the segment you will actually buy—not a cherry-picked showcase. Second, rights diligence should start while the sample is under test; a clean pilot with broken rights wastes months.

How buyers judge quality

Volume alone is not the metric. Buyers ask:

Provenance and rights. Where did each record come from, under what consent or contract, and is commercial AI training explicitly allowed? Pre-2022 licenses that are silent on AI are often treated as unsettled, not as permission (COMPEL).
Privacy. Unredacted personal data is a non-starter for most marketplaces and labs. Datarade’s provider terms prohibit unanonymized sensitive personal data (Datarade Provider Studio).
Deduplication and contamination. Overlap with public web crawls or with the lab’s prior mix can erase lift and poison evals (Dayda).
Annotation and schema fidelity. Inter-annotator agreement, label distributions, and documented schemas matter for supervised and preference data.
Pilot lift. Does a fine-tune or eval on a stratified sample move the lab’s held-out metrics?

What deals look like in public

Reported unit prices from Defined.ai’s CEO to Reuters in April 2024 put rough market levels at about $1–$2 per image, $2–$4 per short video, and $0.001 per word of text (Reuters). Headline licensing deals sit far above those unit rates because they bundle archives, exclusivity, attribution products, and multi-year access:

Reddit and Google expanded a Data API partnership; reporting put the contract near $60 million per year (Reuters; Google announcement).
News Corp–OpenAI has been reported at more than $250 million over five years; other publisher deals (NYT–Amazon, Dotdash Meredith–OpenAI, Shutterstock–Apple) are tracked with disclosed or reported figures on public deal trackers (Fiund deals).

Industry analysts put the broader AI training dataset market on a high-growth path—for example MarketsandMarkets’ projection from $2.82 billion in 2024 to $9.58 billion by 2029 at 27.7% CAGR (MarketsandMarkets). Those figures mix labelling services and licensed corpora; Highford treats them as context, not as Highford’s addressable revenue.

Where deals stall

Deals stall when sellers cannot produce a rights memo; when samples are not representative; when security questionnaires fail; when exclusivity pricing diverges by 4–5× from non-exclusive offers (a pattern also reported in RL environment contracting); when the seller cannot meet delivery SLAs; or when the lab decides to in-house the collection. EU AI Act obligations for general-purpose model providers to publish training-content summaries (Article 53, template presented by the Commission) raise the documentation bar for both sides (European Commission).

Marketplaces such as Datarade (provider commissions from 15–30% plus optional annual fees) and Cloudflare’s acquisition of Human Native show how discovery and licensing rails are consolidating (Datarade pricing; Cloudflare). Highford’s position is different: a searchable private-market information service with free public listings, credit-gated human-submitted listings, AI matching, and double-opt-in email introductions—without buying, selling, or taking a cut of the underlying data.

This page, for an agent

The same page as markdown, plus the API and MCP calls that belong to it.

API

Search public teasers

GET /api/v1/listings?q=unpaid+judgments+in+New+York

Create a free account

POST /api/v1/signup
Content-Type: application/json

{"email":"ada@example.com"}

api_key and token come back immediately. Search works with that key. Unlocks, intros, and credits wait until the sign-in link is opened. POST /api/v1/verify with {"email":"ada@example.com","code":"<6 digits>"} still confirms the address.
A new account starts with 10 credits. Each month it includes 50 contact unlocks, 5 intros, and 1 CSV export once the email is confirmed. Summary search stays free.

MCP

MCP server

{
  "mcpServers": {
    "highford": {
      "url": "/mcp"
    }
  }
}

Search

search_deals {"q":"unpaid judgments in New York"}

JSON

{
  "site": "Highford",
  "page": "/research/how-ai-labs-buy-training-data",
  "markdown": "/research/how-ai-labs-buy-training-data?format=md",
  "api": [
    {
      "label": "Search public teasers",
      "request": "GET /api/v1/listings?q=unpaid+judgments+in+New+York"
    },
    {
      "label": "Create a free account",
      "request": "POST /api/v1/signup\nContent-Type: application/json\n\n{\"email\":\"ada@example.com\"}\n\napi_key and token come back immediately. Search works with that key. Unlocks, intros, and credits wait until the sign-in link is opened. POST /api/v1/verify with {\"email\":\"ada@example.com\",\"code\":\"<6 digits>\"} still confirms the address.\nA new account starts with 10 credits. Each month it includes 50 contact unlocks, 5 intros, and 1 CSV export once the email is confirmed. Summary search stays free."
    }
  ],
  "mcp": {
    "url": "/mcp",
    "calls": [
      {
        "label": "MCP server",
        "request": "{\n  \"mcpServers\": {\n    \"highford\": {\n      \"url\": \"/mcp\"\n    }\n  }\n}"
      },
      {
        "label": "Search",
        "request": "search_deals {\"q\":\"unpaid judgments in New York\"}"
      }
    ]
  }
}

Markdown

Reading this page…

How AI labs buy training data today