AI Training Data Companies: Who Does What in 2026

AI training data companies fall into six groups: human and expert data platforms such as Scale AI, Surge, Mercor and Handshake; licensed content marketplaces such as Protege, Troveo and Defined.ai; buyers and brokers of company records; builders of RL environments and evaluations; annotation and data-tooling companies; and data brokers that sell datasets off the shelf. Our vendor register reads 150 of these companies on their own sites: 122 sell evaluations, 89 sell RL environments and 83 sell expert and human data, and most sell more than one of these. The largest of them now report revenues and run rates that add up to $8.1B, and 29 say they pay for data from outside, which makes them buyers as well as vendors.

This guide maps the market by category. For each group it explains what the companies buy and what they sell, how they are paid, the largest players with the revenue, funding or terms on record, and what a holder of data should know before dealing with them. It is written for the founder, CFO or counsel of a company that holds data and is being approached by these firms, and for anyone who wants a map of who does what. The live registers it draws on are the vendor register (who sells), the buyer board (who publishes a way to buy) and the company pages, where each organisation's deals, funding and financials are gathered.

AIData.Broker acts as the holder's agent: the holder itself licenses or sells its data to the buyer, and we do not take ownership of it. This guide is research from public sources, not legal, tax or investment advice.

Contents

1. The map at a glance
2. Human and expert data platforms
3. Licensed content marketplaces
4. Company-records buyers and brokers
5. RL environment and evaluation builders
6. Annotation and data tooling
7. Data brokers
8. How the categories are converging
9. Questions people ask
10. What this means for you

The map at a glance

The categories below are defined by how a company makes its money, not by what it calls itself. That matters because the same name can sit in several rows: Scale AI sells expert data, evaluations and annotation tools and also licenses companies' records; Labelbox began as an annotation tool, runs an expert network and now buys operational data. Read the table as a map of business models, then look up any one company on its company page to see which models it runs.

The column a holder should read first is "What they buy". Four of the six groups now pay someone outside for data: experts for their work, rights holders for their content, companies for their records, or builders for their environments. That is where a holder's data can be sold.

1. Human and expert data platforms

These are the companies that pay people to make data that does not exist until someone makes it. A lab that wants its model to reason through a tax question, write production code or refuse a dangerous request needs thousands of worked examples and judgments from people who know the subject. The platforms recruit those people (engineers, doctors, lawyers, accountants, writers in dozens of languages), design the tasks, check the work and deliver it. The technique most associated with them is RLHF (reinforcement learning from human feedback, where people judge which of a model's answers is better), but the work now runs well beyond that, into writing full tasks with automatic checks, red-teaming and grading long agent runs.

This is the largest group by revenue on record. The scale of the workforce is large: Handshake was paying ~$60M a month to over 30,000 experts by late 2025, Mercor shows $4M+ in daily payouts, and Labelbox reports an expert network of more than 4 million people. Our expert work page shows what each kind of task pays the people who do it.

The largest players

How they are paid, and what they keep

Labs pay these platforms by the expert hour or by the task, and the platforms pay their experts less. The gap is documented in two places. For one expert hour, a lab pays Mercor $160 to $187 an hour and Mercor keeps $48 to $75 an hour; Appen, which reports its gross margin after paying its crowd, kept 40.3 percent in 2025, and Innodata spent about 60 percent of 2025 revenue on direct costs, keeping about 40 percent. Experts themselves earn $55 to $101 per hour of work across advertised roles, and Mercor states an average contracted rate of $112/hr.

The money comes from very few places. One customer accounted for about 58 percent of Innodata's 2025 revenue, and Handshake's job postings describe its largest lab accounts as running at $100M+ a year each. A small number of labs set the price for the whole group, which is why its revenue swings with each lab's training plans.

What a holder should know

This group matters to holders for two reasons. First, its largest members have become buyers of company records: micro1, Handshake, Mercor, Turing, Scale, Labelbox and Appen all publish programs that pay companies for their operating records, and their business is to resell what they buy, as data or as training environments. micro1's resold datasets were reported at gross margins as high as 80% to 90%. Their published terms differ a great deal; the comparison of company data programs sets them side by side.

Second, these platforms also reach a company's knowledge through its people. Handshake has paid individuals $6 per accepted page, up to $30,000, for professional documents they own, and the expert networks recruit working professionals in large numbers. A company whose staff do this work on the side should make sure its policies say what employees may and may not bring to it.

2. Licensed content marketplaces

A marketplace does not buy data outright. It signs up holders of content (video libraries, music catalogues, photo archives, books, speech recordings, and now company records), packages what they hold into datasets, licenses those datasets to AI developers, and pays each holder a share of what it collects. Some, like Protege, link data from many providers into one dataset for a buyer; others, like Datarade and Opendatabay, are listing sites where the holder sets its own price and the marketplace takes a commission.

For a holder, a marketplace is a way to reach many buyers without negotiating each deal, at the cost of a share of every sale and, often, of control over who buys and on what terms. The marketplaces are growing quickly: Protege's transaction volume was $30 million in 2025, with at least $100 million expected for 2026 (what buyers paid through it, not its revenue).

The largest players

The big image and video libraries sit beside these marketplaces rather than inside them: Shutterstock and Getty Images license their own collections directly to AI companies, and Shutterstock's data, distribution and services revenue reached $203.3 million in 2025. The guide to how AI companies get their training data covers those direct licensing deals.

What a holder should know

Read the exclusivity. Exclusivity is the largest lever on price, worth 4 to 5 times a non-exclusive licence, and marketplaces often ask for it. Troveo says 95 percent of its 7,000+ licensors signed exclusively; its own terms make exclusivity time-limited and earned. Protege's case study explains why a marketplace wants it: holding exclusive AI-training rights to a broadcaster's catalogue let it maximise the price because buyers could not negotiate with multiple parties. That logic can work for the holder or against it, depending on whether the marketplace or the holder captures the higher price.

Compare offers net of the cut. Take rates vary widely. Datarade charges 30 percent, or 20 to 15 percent with an annual fee; Opendatabay's commission falls from 30% to 5% as a sale grows; Wirestock moved its dataset-deal commission to 50 percent; HUD's DataVendor keeps 20 percent of earnings at withdrawal. A non-exclusive licence through a marketplace can sell more than once (Troveo advertises an average of 4 sales per dataset, labelled illustrative), so the right comparison is expected net proceeds over the licence's life, not the headline share.

3. Company-records buyers and brokers

This is the newest group and the one most likely to email a company directly. It buys or licenses a company's internal records (email, chat, tickets, code with its review history, CRM, documents) and turns them into training data or training environments for AI labs. Our buyer board lists 37 buyers with a public way in across 5 kinds, and 24 of them take a company's records. The group mixes three business models: principals that license or buy for their own account and resell, brokers that connect a holder with a buyer for a fee, and wind-down services that sell a closing company's assets as part of a shutdown.

The large expert-data platforms (micro1, Handshake, Mercor, Turing, Scale, Labelbox, Appen) are now the biggest principals here, and the program comparison covers them in depth. The table below shows the companies whose main business is company records.

What a holder should know

Three things. First, the best-funded buyers act for themselves: a principal's published figure is the start of its negotiation, and public estimators differ about 21 times at the low end for the same company. Second, track records are thin: most of this group is months old, several rest on their own statements, and SimpleClosure has warned founders about new buyers advertising $1 million to $5 million offers with no completed purchase behind them. Third, brokers' fees come from somewhere: a broker that charges both sides, as FileYield does, is paid by the buyer as well as the holder, and a holder should know whose interest each intermediary serves.

The most useful public price signal in this group is the one sale that was run as an auction: Spirit Airlines' archive, where competition took Google's opening bid of $5 million to $10 million. The guide to selling company data explains what that means for a private sale, and company records shows what buyers want.

4. RL environment and evaluation builders

An RL environment is a simulated piece of software (a website, an inbox, a ticket tracker, a spreadsheet) with tasks and an automatic check of whether a model completed them; the model practises in it by trial and error. An evaluation (or benchmark) is a set of tasks used to measure how well a model does something, and red-teaming is the work of trying to make a model misbehave so the lab can fix it. These are the products labs need to train and test AI agents, and this is the most crowded part of the market: our register lists 89 vendors selling RL environments and 122 selling evaluations.

Environment builders are paid by contract. Epoch AI found contracts often six to seven figures per quarter, a replica of a website at about $20K, and about $300K for a Slack-grade clone, and tasks at $200 to $2,000 each. Exclusivity is priced explicitly at roughly 4 to 5 times a non-exclusive deal.

The largest players

The large expert-data platforms build environments too: Mercor's chief executive described tasks priced from $50 to $10,000, and Snorkel, SuperAnnotate and Centific list environments among what they sell. The guide to RL environments covers how they are built and priced.

What a holder should know

Environment builders are a new kind of buyer for a holder. A replica of software can be built from scratch, but the most realistic environments copy a real company's work: Ooak anonymises a company's records into a code-named digital twin, Collinear builds simulated workplaces from companies' code and business-tool history, and The General Data Company builds environments from industrial companies' operating records. A holder whose records show a complete workflow (tickets linked to code, deals linked to emails) has what these builders need.

The holder should also know that environments age fast. One investor's analysis put an environment's useful life at a half-life of about six months. That churn keeps demand for fresh material high, and it is one reason a holder can ask for a fixed term rather than a perpetual licence: the builder will want fresh material again, and the holder can price the renewal.

5. Annotation and data tooling

Annotation is the original data-labelling business: drawing boxes around objects in images, transcribing audio, tagging the entities in a sentence, classifying documents. It is done with software (the tools) and by people (the workforce), and many companies sell both. It still underpins computer vision, robotics and speech, even as text models have moved to the expert work of category 1. The tools range from open-source (CVAT) to enterprise platforms (Labelbox, SuperAnnotate, Encord, Dataloop), and the services from managed workforces (Sama, CloudFactory, TaskUs) to self-serve marketplaces.

Prices in this group are per item and small. Appen's self-service platform starts around $0.01 to $0.10 per simple annotation and Scale's typical per-task range runs $0.05 to $1.00+, as reported by a procurement benchmark; Roboflow quotes $0.10 per bounding box and $0.20 per polygon; CVAT's services start at $0.10/object.

The players, and the consolidation

The group is consolidating into larger platforms. Dell bought Dataloop, Uber bought Segments.ai to grow its data-labelling business, and Handshake acqui-hired key Cleanlab employees. Our register counts 11 vendors on it that have been acquired.

What a holder should know

Annotation companies are mostly sellers of labour and software, not buyers of data. The exceptions are the ones that have added data programs: SuperAnnotate pays on delivery and again on each relicense, and Labelbox's Prism publishes datasets starting around $50K with licences exclusive by default. For a holder, the relevance is indirect: if your data needs labelling before a buyer will take it (images without tags, recordings without transcripts), these are the companies that do it, and the cost is per item.

6. Data brokers

A data broker, in this market, sells datasets it collected or aggregated, off the shelf, to anyone who pays. Some scrape the public web and sell the result in bulk: Bright Data sells a Wikipedia articles dataset from $250 for 100K records and Quora data at up to $0.0025 per record. Others build catalogues of recorded data and sell licences to it repeatedly: Nexdata lists speech datasets on Datarade starting at $20,000 a purchase, and on Opendatabay at up to $526,300 for 1,500 hours.

The broker's economics rest on selling the same dataset many times. One catalogue seller, Haitian Ruisheng, sold 576 licences on 395 of its own datasets in a single year. Off-the-shelf data is correspondingly cheap: one vendor prices catalogue call-centre audio 5 to 60 times below custom collection. The catalogues we read every day hold 511 lots on offer, 165 of them companies' records offered through brokers; the lot index lists them.

What a holder should know

A broker's dataset competes on price, not scarcity. Scraped public data and catalogue recordings are sold to many buyers, so their prices fall toward the cost of collection. A holder's private records are the opposite case: one copy, not on the web, licensed to few buyers. If a broker offers to list your data in its catalogue, ask whether the licence is exclusive, how many times the broker may sell it, and what share reaches you.

Federal rules restrict some buyers. The US Justice Department's Data Security Program forbids a US person from knowingly engaging in a covered data transaction involving data brokerage with a country of concern or a covered person, and a covered person includes an entity with its principal place of business in a country of concern. A licence of bulk sensitive personal data to such a buyer can be prohibited outright, whoever brokers it. Check the buyer's home country, and its owners, before any sale that includes personal data.

How the categories are converging

The six groups are moving toward one another, and the direction tells a holder where demand is going. Expert-data platforms now build environments and buy company records. Annotation tools have added expert networks and data programs. Marketplaces that began with video now take business records. Environment builders license real companies' data to make their simulations realistic. The common thread is that every group has found that the data labs want most is data that already exists inside companies and cannot be made to order cheaply.

The figures on our register show it. Of the 150 vendors we read, 17 say they build from data licensed from the people and companies that hold it, and 29 say they pay for data from outside. Data companies have raised $4.2B in disclosed funding, across 111 funding rounds and acquisitions we record. And 24 buyers now publish a way to buy company records, many of them launched in 2026.

Questions people ask

What are AI training data companies?

They are the firms that supply the data AI models learn from: platforms that pay experts to write tasks and judge answers, marketplaces that license content on behalf of rights holders, buyers of company records, builders of simulated environments and evaluations, annotation companies and data brokers. Our vendor register lists 150 of them, each in its own words.

Who are the biggest AI training data companies?

By revenue on record: Mercor at over $2 billion in gross run rate (reported), Surge at $1.2 billion in 2024 revenue (reported), Handshake at more than $1 billion gross annualised and micro1 past $1B in run rate (both self-reported), and Scale AI at $870 million in 2024 (reported), in which Meta took a stake for about $14.3 billion. The figures are of different kinds and dates, so read them as orders of magnitude.

How do AI data companies make money?

Most charge labs for expert work by the hour or task and pay the experts less: Mercor keeps 30% to 40% of what a lab pays for an expert hour. Marketplaces take a share of each licence they sell, brokers take a commission, and principals that buy company records resell them, as data or as training environments, at margins reported as high as 80% to 90% for resold datasets.

Do AI data companies buy data from businesses?

Yes, increasingly. 29 vendors on our register say they pay for data from outside, and 24 buyers publish programs for companies' records, including micro1, Handshake, Mercor, Turing and Scale. The comparison of company data programs sets their prices, licences and payment terms side by side.

What is the difference between a data vendor, a marketplace and a data broker?

A vendor makes or buys data and sells it in its own name, usually made to order for a lab. A marketplace licenses data on behalf of the holders it represents and pays them a share. A broker sells datasets off the shelf, often collected from public sources, or connects a holder with a buyer for a fee. One company can be all three.

Are AI training data companies legitimate?

The large ones are, on independent evidence: listed companies such as Appen and Innodata file accounts with their exchanges and regulators, and others have funding from named investors. Many newer buyers of company records rest on their own statements, and SimpleClosure has warned about offers with no completed purchase behind them. Check each one's company page and the track record on the buyer board.

How do I sell data to an AI training data company?

Start by knowing which kind of company fits what you hold: content to a marketplace or a lab's licensing team, company records to a records program, workflow-rich records to an environment builder. Then compare several before you sign; the guide to selling company data to AI labs walks through the process.

What this means for you

The training data market is no longer one industry of labelling shops. It is six business models converging on the same scarce input: data that already exists inside companies. For a holder, that means more possible buyers than a year ago, and more variation in what each will pay and ask for.

1. Identify which groups want what you hold. Use the buyer match tool for company records and the what labs buy hub for other kinds of data.
2. Check every counterparty. Read its company page for funding, deals and financials, and its row on the vendor register or buyer board for what it says it does and what others confirm.
3. Know your value before anyone quotes you. Value your data against published prices, not one firm's estimator.
4. Read the terms, not just the price. Exclusivity, sublicensing, take rates and payment triggers differ more than prices do; put any offer through the offer check.
5. Ask several at once. The groups compete for the same data, and a holder that shows one datasheet to every fitting buyer sees how differently they value it.

If you would rather have an agent run that comparison for you, our terms are on the pricing page. For how this market fits into the wider supply of training data, read how AI companies get their training data.