How Big Is the AI Training Data Market? What the Records Show

No one can state the size of the AI training data market from public evidence, and the published estimates that try to measure it disagree with each other by an order of magnitude. What the records do show is a floor and a shape. The large data vendors' newest reported annual revenue or run rates add up to about $8.1B, led by Mercor at a reported gross run rate of over $2 billion. Data companies have raised $4.2B in disclosed funding rounds. The single largest transaction, Meta's investment in Scale AI, was reported at about $14.3 billion for a 49% stake. Licensing deals for content, the part most people picture, are far smaller and mostly undisclosed: only 19 of 111 disclosed deals carry a dollar amount.

This guide is for founders and finance leads deciding whether their data is a real asset, investors and reporters sizing the market, and buyers benchmarking what they spend. It explains why the headline estimates disagree, then builds the picture from the ground up, using only revenues, deals and funding that companies or credible reports stated: what the data vendors earn, what licensed content earns its owners, what investors put in, the large strategic transactions, and how the market splits into segments. It ends with what is unknown, because the unknowns are large.

AIData.Broker maintains the market tracker, deal ledger, vendor register and company pages this guide draws on, and acts as the agent of US holders who license or sell their data. Every number below links to its record and source. This guide is research from public sources, not legal, tax or investment advice.

Contents

1. Why the published estimates disagree
2. A bottom-up picture from the records
3. What the data vendors earn
4. Licensed content: what disclosed deals add up to
5. The money raised by data companies
6. The large strategic transactions
7. The market by segment
8. What is unknown
9. Questions people ask
10. What this means for you

Why the published estimates disagree

Search for the size of this market and the first page is full of market research summaries, each with a headline figure, a growth rate and a forecast year, and the headlines sit an order of magnitude apart. We do not cite those figures as facts, for a simple reason: none of them can be checked. The underlying reports are sold, their methods are summarised in a paragraph, and the summaries are copied from site to site until the original is impossible to find. But the reasons they disagree are worth understanding, because they show what "the market" can mean.

They measure different things

The biggest source of disagreement is definition. Some reports count only annotation software: the tools companies use to label images and text. Others count data services: the people paid to label, rank, write and check data, which is where most of the money is. Some include licensed content (news, books, images, user posts); most leave it out because it is hard to see. Few count company records sold for training, or RL environments (simulated workplaces where AI agents practise tasks), because those markets barely existed when the reports' methods were designed. A report on annotation software and a report on human data services can both be titled "AI training data market" and differ tenfold without either being wrong.

They are built top-down

Most estimates start from a larger number (total AI spending, or the revenue of the IT services industry) and assume a share for data. Small changes in the assumed share move the result by billions. Even public companies reach for the broad number when they describe their opportunity: Reddit's IPO prospectus framed its data licensing opportunity with "estimates from International Data Corporation's ('IDC') Artificial Intelligence Tracker" for "the broader AI market," not a figure for data licensing ( Reddit final prospectus, SEC). A share of a forecast is a forecast.

The biggest companies are private

The largest vendors in this market (Mercor, Surge AI, Scale AI, Handshake AI, micro1, Turing) do not publish audited accounts. Their revenues reach the public through funding announcements, leaks to the press and their own blog posts, often as run rates (the latest month's revenue multiplied by twelve) and often as gross figures that include the money passed straight through to contractors. A report that cannot see these companies' revenue has to guess it, and a report that takes the leaked run rates at face value counts money that is not revenue in the usual sense.

The buyers do not say what they spend

The other side of the market, the AI labs, almost never discloses its data spending. OpenAI, Anthropic, Google DeepMind and Meta do not break out what they pay vendors or licensors, and what does surface comes from reports of internal plans. The Information reported that Anthropic's leaders discussed spending more than $1 billion on RL environments over a year ( the record); founders told a reporter that the six largest Chinese AI labs spend about $500 million a year with American data-labelling companies ( the record). And forecasts are forecasts: a former OpenAI researcher predicts leading labs could spend more than $100 billion a year on data by the end of the decade ( the record). Each of these is reported, not stated by the company, and none is a measurement of today's market.

A bottom-up picture from the records

Instead of a top-down share, this guide adds up what can be seen. The method is simple and its limits are stated in each section. We take only figures that a company published (a filing, its own announcement) or that a named publication reported, each held as a record with the source's own words. We keep stated and reported apart. We never add up figures that describe different things (a gross run rate and a licence fee are not the same kind of money) without saying so.

The table summarises the building blocks. Each is explained in the sections that follow.

None of these rows is the market. Funding is not revenue, an acquisition price is not annual spending, and vendors' gross revenue includes what they pay the people who make the data. Together they bound it: the data market is at least as large as what vendors are paid, and the stakes around it are as large as what investors and acquirers have put in.

What the data vendors earn

Most of the money in AI training data goes to vendors: companies that recruit experts and crowds to write, label, rank and check data, build environments and evaluations, and sell the result to AI labs. Our vendor register lists 150 of them, of which 83 sell expert and human data, 89 sell RL environments and 122 sell evaluations (many sell more than one). A handful have grown very fast, by their own or reported numbers.

The chart shows the newest headline revenue figure for each of the large private vendors in our tracker, the same figures behind the $8.1B total. The periods differ, most figures are run rates rather than a year's revenue, and several are gross of what the vendor pays its contractors.

Sources: Mercor, Surge AI, Handshake AI, micro1, Scale AI, Prolific, Snorkel AI, Mecka, Turing, Proximal, Invisible Technologies, AfterQuery, Prime Intellect.

How to read these figures

Three cautions apply to every bar. Basis: Mercor's, Surge's, Handshake's, Scale's and Prolific's figures were reported by the press; Mecka's, Proximal's, Invisible's, AfterQuery's and Prime Intellect's are the companies' own claims; micro1's, Snorkel's and Turing's were stated by the companies in announcements. None is audited. Period: Surge's and Scale's figures are for 2024, Turing's for early 2025, the rest for 2026, and Mecka's is a projection for the end of 2026. A company's revenue a year ago and another's run rate last month do not add up to one year of anything. Gross versus net: a vendor that bills a lab for an expert's hour and pays the expert most of it reports the whole bill as gross revenue. Mercor pays out 60% to 70% of its topline revenue to contractors, by one report ( the record), and its gross margin was 27% in 2025, rising to 33% in the second quarter of 2026 ( the record). Our pricing guide puts the share of a lab's dollar Mercor keeps at 30% to 40%.

That is why the figure behind the chart is described, in its own derivation, as "an order of magnitude, not a market size" ( the figure). It is still the most informative single number available: the large vendors that sell to AI labs are, together, a business measured in single-digit billions of dollars a year of gross billings, and growing fast.

How fast they grew

The growth is the story the run rates tell best. Mercor reported $75 million ARR in February 2025 and a gross run rate of over $2 billion in June 2026; its first-half 2026 gross revenue was $614 million, more than 90% from AI foundation model companies ( the record). micro1 said it had $7 million of ARR at the start of 2025 and surpassed $1B annual run rate by October 2026. Handshake AI was forecast to end 2025 at $300 million and reported reaching $1 billion early in 2026. Scale AI's $870 million in 2024 came with guidance of $2 billion for 2025, reported before Meta's investment. These are the companies our guide to AI training data companies profiles one by one.

The listed vendors: smaller, slower, audited

A few data companies are listed on stock exchanges and file audited accounts. Their numbers are smaller and steadier, and they are the only revenue figures in this market that meet an auditor's standard.

The listed figures show the concentration that the private ones hide. Innodata's single largest customer accounted for approximately 58% of its 2025 revenue; Appen's revenue from Google fell to $13.4 million in 2024, against $82.8 million in 2023, after Google ended its contract. Data vendors depend on a few AI labs, which means the market's size swings with those labs' choices.

Licensed content: what disclosed deals add up to

The part of the market most people picture, AI companies paying publishers, platforms and archives for their content, is the least visible. Our deal ledger holds 111 disclosed licences to AI companies, from 44 different licensees. Only 19 carry a dollar total or yearly amount, and 92 carry none. The priced deals are listed below; for the news deals in detail see our guide to AI licensing deals with news publishers, and for Reddit, Reddit's AI data deals.

Why we do not print a total

It is tempting to add the column. We do not, because the rows are different kinds of money: a multi-year contract value, a yearly fee, one year's revenue from many buyers, a pool, a winning bid, and a buyer's spending pledge. Added naively they would double count (the Reddit aggregate overlaps the Google row; Shutterstock's revenue line includes deals listed elsewhere) and mix totals with annual rates. What can be said safely is the order of size. Even taken together, before netting out the overlaps, the disclosed licensing amounts come to well below the vendors' combined run rates above. The licensed content market is real and growing, but it is not where most AI data money goes.

What content owners report

A better measure is what content owners report as their own AI licensing revenue, because it is one company's audited or stated number for a defined period. The list is short.

The pattern matters for anyone trying to size this segment. Archive deals arrive in lumps: Informa called its 2024 AI revenue non-recurring, Wiley's quarterly AI revenue fell by more than half against a year earlier, and Getty attributed a guided revenue decline in 2026 to one-time recognition of 2025 deals ( the lever). Only Reddit, which sells continuing access, reports a line that rises steadily. A market made of one-off archive sales can look large in one year and small the next.

The money raised by data companies

Funding is not revenue, but it shows what investors believe the market will become. Data companies (vendors, marketplaces, brokers, wind-down services and work platforms) have raised $4.2B across the 60 funding rounds in our tracker whose amount was disclosed. The total leaves out AI labs' own rounds, acquisitions, loans and valuations, and a round with no disclosed amount adds nothing ( the figure). Our market page lists every round; the tracker holds 111 funding rounds and acquisitions in all.

Sources: the funding rounds behind the $4.2B total, summed by the quarter of their announcement. The spike in the second quarter of 2024 is one round, Scale AI's $1B Series F; the 2022 round not shown is Handshake's $200 million Series F, raised before it became an AI data company.

What the rounds show

Leave out Scale's 2024 round and the direction is clear: the third quarter of 2026 raised more than any quarter before it, and the fourth, only a few weeks old on our cut-off date, is already close behind. The largest recent rounds went to the vendors with the fastest revenue growth and to the new segments. Snorkel AI raised $350 million at a $3.5 billion valuation; Mercor raised $350 million at $10 billion and was later reported at a $20 billion valuation; Prime Intellect raised $130M; Labelbox raised more than $100 million at a $1.25 billion valuation; micro1 raised more than $100 million at a reported $4 billion. Robot and physical-world data companies now appear among the large rounds: Mecka raised $25 million, $35 million and $60 million in under a year, and XDOF $70 million.

Two rounds in the latest quarter deserve care. SignSplit announced a $400 million "strategic seed" in its own press release, an unusual size for a seed round that no independent source has confirmed; and Mode's figure is the cumulative total of a Regulation A offering to the public, not a venture round. Both are on record as stated by the companies; both inflate the quarter.

The large strategic transactions

The biggest numbers in this market are not revenues or funding rounds but strategic deals, in which an AI company or a larger services firm bought control of, or a large stake in, a data company. They show what the buyers think the capability is worth to them, which is a different question from what the market spends in a year.

The Scale transaction dwarfs everything else on record, and it is worth reading carefully. Meta bought a minority stake, not the company, and the price reflected talent (Scale's chief executive moved to Meta) and strategic position as much as Scale's data business; TechCrunch described the deal as " a 49% stake in the startup, which produces and labels data". It tells you that one large AI lab valued control of a leading data vendor at tens of billions of dollars. It does not tell you that the data market spends anything like that in a year.

The Disney row runs the other way: the content owner invests in the AI company. Our ledger records "no licence fee disclosed" and a $1 billion equity investment by Disney, so money flows to OpenAI ( the record). Large content deals can be partnerships in which no licence fee changes hands at all.

The market by segment

The records sort the market into five segments that behave differently. They are sized here only by what is on record, in the units the records use, and none of them has an honest single total.

Human and expert data

This is where the money is. Labs pay vendors for people's time and judgment: experts writing problems and solutions, ranking model answers (the work behind RLHF, reinforcement learning from human feedback), building tasks with checkable answers. Prices are per hour or per task, and our price board and rates list them; our guide to AI training data prices and the expert work page explain them. Surge AI's pricing shows how far quality can be priced: it "charges 50% to ten times more than competitors," by one report ( the record).

Licensed content

Publishers, platforms, stock libraries, music labels and academic presses license what they already have. The deals are many and the prices few; the segment's visible revenue is the content owners' own AI lines in the table above. The shift from training to display rights, described in our news guide, means more of this segment is now paid for showing content inside AI products than for training on it.

Company records

The newest segment is the sale of a company's own internal records: the email, chat, tickets, documents and code that show how real work gets done, de-identified and sold or licensed to vendors and labs. It is small in dollars and large in activity. SimpleClosure, which handles startup shutdowns, processed nearly 100 of these deals in a year with payouts of $10,000 to $100,000 per company ( the record) and says its marketplace is on track to return more than $10 million to startups by the end of 2026. Handshake says it has paid $250M out to partners, without a period ( the record). The broker Protege reports transaction volume rising from $30 million in 2025 to at least $100 million in 2026. And the largest single sale, Google's winning bid for Spirit Airlines' archive, was $10M. The buyers' published programmes are compared in our guide to company data buyer programs, and our company records page lists the prices.

Environments and evaluations

RL environments are software: simulated versions of a company's tools in which an AI agent practises tasks and is scored. They are bought per environment and per task, and our guide to RL environments covers what they cost. The segment is the one most likely to be undercounted by older market reports, because it barely existed before 2025. Our datasets and environments page holds the prices.

Physical-world data

Data for robots (first-person video of people working, teleoperation recordings, sensor logs) is priced per hour and is the newest segment to attract large funding. Our physical-world page lists its prices.

What is unknown

A bottom-up picture is only as good as what is visible, and much is not. The gaps below are not footnotes; any one of them could change the market's size by a large factor.

1. What AI labs spend in total. No lab discloses its data spending in a filing. The best evidence is reports of internal plans, like Anthropic's discussed spending on environments, which are not measurements.
2. What labs do in house. Large labs employ their own annotators and build their own environments. That spending never appears as a vendor's revenue, and no record measures it.
3. Net revenue of the private vendors. The run rates are mostly gross. With payouts of 60% to 70% at one vendor, the vendors' own net revenue is a fraction of the headline.
4. The prices of most licences. 92 of the 111 disclosed deals have no amount. Content deals are confidential by default.
5. Double counting. Vendors buy licensed data and resell it, brokers report transaction volume that includes what they pay holders, and a licence fee can appear both as a publisher's revenue and as part of a vendor's cost of sales. Adding segments without netting overstates the market.
6. Run rates are peaks. A run rate is the best recent month times twelve. Vendors that sell to a few labs can lose a large share of revenue when one lab changes course, as Appen's experience with its Google contract shows ( the record).
7. Markets outside the US. Chinese labs' spending with American vendors was reported at about $500 million a year; their spending with Chinese vendors, and much of Europe's and India's market, is not in our records.
8. Synthetic data. Data generated by models to train other models is a growing substitute for bought data, and its cost sits inside labs' compute budgets, not in any data market figure.

These gaps are why we publish the records rather than a single number. Each figure above links to its source, so anyone can rebuild the picture, and the records, figures and sources pages show what has been measured.

Questions people ask

How big is the AI training data market in 2026?

There is no verifiable single figure. The large data vendors' newest reported revenue or run rates add up to about $8.1B, mostly gross of contractor payouts and over different periods, which is best read as an order of magnitude for the vendor segment. Licensed content and company records add a smaller amount that is mostly undisclosed.

Why do market size estimates for AI training data differ so much?

They define the market differently (annotation software only, or human data services, or licensed content too), they are built top-down from shares of larger forecasts, and they cannot see the private companies' revenues or the labs' spending. Two reports with the same title can measure different things.

Who are the biggest AI training data companies by revenue?

By newest reported figure: Mercor at over $2 billion in gross run rate, Surge AI at $1.2 billion in 2024 revenue, and Handshake AI and micro1 at about $1B each in run rate, followed by Scale AI at $870 million in 2024. Most of these are reported or self-reported; see our guide to AI training data companies.

How much do AI companies pay for licensed content?

Few prices are disclosed: only 19 of 111 deals carry an amount. The largest reported yearly figures are about $50 million a year for News Corp's deal with OpenAI and up to $50 million a year from Meta; the largest stated contract total is Reddit's $203.0 million across its January 2024 licences.

How much money has been invested in AI training data companies?

Disclosed funding rounds in our tracker total $4.2B, excluding AI labs' own rounds, acquisitions and loans. The largest single transaction is Meta's reported $14.3 billion investment for 49% of Scale AI, which is counted separately as a strategic investment.

Is the AI training data market growing?

Every visible measure says yes. The largest vendors' run rates multiplied within eighteen months (Mercor from $75 million to over $2 billion), funding per quarter rose through 2026, and new segments (company records, environments, robot data) appeared. Licensed archive revenue is the exception: it arrives in one-off lumps that can fall year on year.

Can a company sell its own data into this market?

Yes. Company records are a small but fast-growing segment, with buyer programmes from vendors such as micro1, Mercor, Handshake and Turing and published price ranges per company. Our guide to selling company data to AI labs and the valuation explain how it works.

What this means for you

For a holder, the size of the market matters less than its shape. Most of the money goes to vendors who pay people to make data to order; a smaller and lumpier stream goes to owners of content and records that cannot be made to order. That second stream is where a company's own data sits, and its value comes from being something a vendor cannot produce: real records of real work, with clean rights.

1. Find your segment. Is what you hold content (published or user-made), company records, expert work or something buyers would pay to have made? The what labs buy page and the buyer board map the demand.
2. Price it in the market's units, not as a share of a market forecast: per company, per hour, per title, per task. The price board, rates and our guide to how much company data is worth give the anchors, and the valuation applies them to you.
3. Expect few buyers and confidential prices. The deal ledger shows how rarely prices surface; a competitive process is how a holder finds the real one. See negotiate for the levers.
4. Watch the vendors. The companies in the chart above are also buyers of data; their pages and the vendor register show who pays for what.

AIData.Broker acts only for holders: the holder licenses or sells its own data to the buyer, and pays us a fee fixed in the mandate only if a sale closes; we run sales for US holders. Read direct or through us to decide whether you need an agent. For more context, see how AI companies get training data, Reddit's AI data deals and AI licensing deals with news publishers. How we source every number explains the method behind this guide.