Skip to content
Outrich Data

AI training data market statistics

Updated 2026-11-16 · 6 min read · By the Outrich Data team

There is no reliable public figure for the total size of the AI training data market, and none is asserted here. What exists is a small set of reported deal values and the ranges that buyers of company data advertise. This page lists those figures with their sources and explains what each one shows.

What this page contains

Most deals for AI training data are private. The parties sign confidentiality terms, and prices are not filed anywhere. What reaches the public is a handful of figures from company disclosures, press reports and analyst compilations, plus whatever buyers choose to print on their own websites.

This page collects those figures and nothing else. Every number below is attributed to the page it was read on, and the same pages are listed under sources. No estimate, average or projection has been added.

Readers looking for a single market-size number will not find one here. The reasons are set out further down. The short version is that the published estimates rest on assumptions that cannot be checked, and repeating one would give it weight it has not earned.

The figures come from two kinds of source. The first is secondary reporting on deals between content owners and AI developers. The second is the buyers' own web pages, read on the date shown in the source list. Neither is an audited record. Both are quoted here as claims made by the named source.

Reported deal values

Neudata reports the Reddit and Google agreement at about $203M total contract value, Shutterstock's deals with several AI labs at $25M to $50M per deal, the Taylor & Francis and Microsoft agreement at $10M upfront, and the Le Monde and Prisa agreements with OpenAI at $1M to $5M.

Monda lists Wiley's deal with an undisclosed buyer at $23M, one time, and X data access at about $42K a month. The table sets these out with the kind of data involved and the date of each source.

PartiesDataReported valueSourceSource date
Reddit and GoogleForum posts and commentsAbout $203M total contract valueNeudata2025-06-10
Shutterstock and several AI labsImages, video and metadata$25M to $50M per dealNeudata2025-06-10
Wiley and an undisclosed buyerAcademic and professional books$23M, one timeMonda2024-10-28
Taylor & Francis and MicrosoftScholarly publishing$10M upfrontNeudata2025-06-10
Le Monde, Prisa and OpenAINews archives$1M to $5MNeudata2025-06-10
X data accessPost firehoseAbout $42K a monthMonda2024-10-28
Closed startups and AI labsA company code baseAbout $25,000MIT Technology Review2026-10-13
Closed startups and AI labsWorkspace data (chat, docs)About $15,000MIT Technology Review2026-10-13

What the deal values show, and what they do not

These figures show that large content owners have been paid substantial sums for access to their archives, and that the structures differ. The Reddit and Google figure is a total contract value. The Wiley figure is a one-time payment. The Taylor & Francis figure is an upfront amount, which implies more may follow. The X figure is a recurring monthly price for access.

They do not show a price per record, per word or per image. The sources report totals, and the volumes covered are mostly not public. Dividing one by a guessed volume produces a number with no basis.

They also do not describe company operational data. Every one of the publisher and platform deals concerns published or user-generated content at very large scale. A mid-sized company's internal tickets and documents are a different asset, sold in different volumes, to meet a different need. And the figures are secondhand. They are what Neudata and Monda report, drawn from earlier disclosures and press coverage, and the underlying contracts have not been published.

Selection is a further limit. Deals become public when a listed company has to disclose them or when a party chooses to announce them. Smaller and less flattering transactions stay private. A list of public deals is therefore tilted toward large, well-known content owners and says little about the ordinary case.

The closed-company figures

The last two rows of the table are of a different kind. MIT Technology Review reported that closed startups were selling a company code base to AI labs for about $25,000, and workspace data such as chat and documents for about $15,000.

These are the only figures in the table that concern internal company data. They show that such transactions happen and that the reported amounts for small, defunct companies are modest next to the publisher deals.

They do not show what an operating company's data would fetch. A closed startup is typically small, its history is short, and it sells under time pressure. The report describes particular cases. It is not a survey, and it does not establish a typical price.

What buyers of company data advertise

Several firms that buy or license company operational data publish a range on their own sites. Scale AI gives $10K to $1M+ per partnership and labels it illustrative. micro1 states $100K to $3M+. Mode states $100K to $5M. Polyshares states $100K to $2M and above. DataPaid states $100K to $2M per company. Mercor does not publish a range. Troveo offers a calculator and describes the result as illustrative only.

The same pages state who the buyers want to hear from. The table records the model each firm describes, its advertised range and its stated minimum size, in the firms' own terms.

Stated minimum sizes follow a pattern. Where a figure is given, the buyers ask for companies with some tens of employees at least, and one notes that it accepts fewer in some fields. Two state no minimum.

BuyerModel as describedAdvertised rangeStated minimum size
Scale AIData partnership program$10K to $1M+ per partnership (illustrative)40+ employees preferred
micro1Buys operational data$100K to $3M+30+ full-time employees
ModeBuys operational data$100K to $5M30+ employees, fewer in some fields
PolysharesDirect buyer$100K to $2M and aboveNot stated
MercorBuys operational dataNot publishedNot stated
DataPaidDirect buyer$100K to $2M per companyStrongest fit 30+ US employees
TroveoLicensing marketplaceCalculator, illustrative onlyEstimator starts at 20 employees

What advertised ranges show, and what they do not

Advertised ranges show what buyers are willing to say in public to attract sellers. They show that several firms are competing for the same kind of supplier, and that their stated ranges overlap heavily.

They are not records of payments. A range on a marketing page is a claim by the firm that wrote it. None of the pages states how many companies were paid, what the median payment was, or how often a seller landed at the top of the range. Two firms use the word illustrative themselves.

The ranges are also very wide. The top of each range is many times the bottom. A range that wide tells a seller that the result depends on the specifics of its data and its rights, and little more. Outrich Data does not treat any figure on this page as a quote. A written offer follows a rights review of the particular dataset.

Why no total-market figure appears here

Research firms publish estimates of the size of the AI training data market and forecasts of its growth. Those estimates differ widely from one another. They define the market differently, with some counting labeling services and synthetic data and others counting licensing alone. They rest on private deal data and on assumptions that are not disclosed in full.

A figure of that kind cannot be verified from public sources, so it is not repeated on this page. The same applies to claims about average deal size, typical price per unit and the share of companies that have sold data. If a reliable, checkable total becomes available, it belongs here with its source. Until then, the honest statement is that the total is not known.

Readers will see such totals quoted elsewhere with confidence. A useful test is to ask what the number counts, where the underlying deal data came from, and whether another firm's estimate for the same period is close to it. Where those questions have no clear answer, the figure is better left out of a board paper or a valuation.

How to use these figures

The figures are useful as context and misleading as benchmarks. A few rules keep them in proportion.

  • Cite the source with the number. A deal value without its source and date is a rumor.
  • Keep the structure attached. Total contract value, upfront payment, one-time fee and monthly access are not the same thing.
  • Do not convert totals into unit prices. The volumes are not public.
  • Do not compare publisher archives with internal company records. They are different assets.
  • Treat advertised ranges as marketing claims by the firm that made them.
  • Do not quote a total market size unless you can show how it was calculated.
  • For your own data, rely on a written offer made after a rights review, not on a table.

Common questions

How big is the AI training data market?
No reliable, verifiable total is available, and none is asserted here. Published estimates differ widely and rest on undisclosed assumptions. What can be cited are individual reported deals and the ranges buyers advertise, each with its source.
What is the largest reported AI data licensing deal on this page?
Among the figures collected here, the largest is the Reddit and Google agreement, which Neudata reports at about $203M total contract value. It concerns forum posts and comments, not internal company data.
What do buyers pay for company operational data?
Firms that buy it advertise ranges on their own sites, from $10K at the lowest stated figure to $5M at the highest. Those are marketing claims, not payment records. What a given company receives depends on its data and its rights.
Why are the figures for closed startups so much lower?
MIT Technology Review reported about $25,000 for a code base and about $15,000 for workspace data. Closed startups are typically small with short histories and sell under time pressure. The report describes particular cases, not a typical price.

Sources

  1. Neudata2025-06-10
  2. Monda2024-10-28
  3. MIT Technology Review2026-10-13
  4. Scale AI2026-11-16
  5. micro12026-11-16
  6. Mode2026-11-16
  7. Polyshares2026-11-16
  8. Mercor2026-11-16
  9. DataPaid2026-11-16
  10. Troveo2026-11-16

Find out if your data qualifies.

Eleven questions, about five minutes, no call required. You get an answer on screen and a written follow-up.

esc