AI Training Data Companies
Source a training set from the companies that license, label, collect or generate it. We index 50 of them by type, with the prices they publish, and introduce your team to the ones that fit.
Existing archive
news_archive_2015_2024
42,180 articles · text
license: AI training, yesWritten by experts
finance_sft_demos
5,000 prompt/answer pairs
rubric: 2 reviewersGenerated from a schema
claims_tabular_synth
250,000 rows · tabular
seed: masked recordsBefore anything ships
- Each part has a named source
- License names AI training
- Seller may sublicense
- Personal data removed, method stated
Training set, v1
3 sources
One manifest, one datasheet, one license file per component.
Invented example. The datasets, counts and names are made up to show the shape of a sourcing project.
The index
Six kinds of company sell training data
Licensed data, annotation, expert and RLHF networks, synthetic generators, marketplaces and evaluation. Most buyers use two or three of them on one model. The directory lists each company alphabetically within its type, never ranked.
Pricing
Four ways vendors charge, and the rare public numbers
Most companies in the directory quote per project and publish nothing. These are starting points from the few that do, plus interview and press figures where no vendor prints a rate.
How each model behaves on a real project, or put your volumes into the cost estimator.
Per task or item
A rate per labeled object, entity, ranking or finished task
$0.02a box at Label Your Data; keypoints from $0.015Per hour
Annotator, expert or participant time
$8/hrProlific's participant floor ($12 suggested), plus its feePer graded task
One RL task with its grader, built to order
$200–$2,000the usual range in Epoch AI's interviews; exclusivity about 4 to 5 times moreLicense fee
The right to train on existing data, for a term
$1–$4/minfor video footage, as Bloomberg reported
Published starting rates and reported figures, not quotes.
Choosing a route
Match the gap in your model to the cheapest route that closes it
Start from what is missing, not from a vendor list. The full comparison of the three routes covers rights and risks.
You need breadth: many domains, styles or years
Licensed data or a marketplace
Watch for: Whether the seller may sublicense for training
You hold raw images, audio or text without labels
Annotation platform or managed service
Watch for: Guideline drift across annotators; run a pilot
The model must learn expert judgment
Expert and RLHF networks
Watch for: Expert vetting and how disagreements are settled
An agent must learn to use software
RL environment builders
Watch for: Whether tasks have reliable graders
Real data is sensitive, rare or slow to collect
Synthetic data generators
Watch for: Test on real held-out data
Ask about provenance before you ask about price
Get the answers in writing. A missing answer usually means the work has not been done. The provenance checklist has all 21 questions.
Who created each part of the set?
Contributors, a licensed publisher, a public source or a generator model, stated per component, not once for the whole delivery.
Does the license name AI training?
"Internal use" or "analytics" is not the same permission.
Can the vendor sublicense it?
A reseller needs rights from the original holder that allow it.
What survives the term?
A survival clause for your trained model matters more than the expiry date.
What is warranted, and who pays if a claim arrives?
Warranty and indemnity decide who carries a rights claim. Have your own lawyer read both.
How was personal data removed?
Ask for the method, not a promise. Recorded voices need extra care: some states, Illinois among them, treat voiceprints as biometric identifiers.
Which kind of training data company do you need?
Four questions, in the order most buyers rule each type in or out.
- Step 1
Check whether the data already exists
Licensors sell permission, not labor: price turns on size, uniqueness, exclusivity and term. Labelbox Prism licenses company data, exclusive by default. On marketplaces, AWS Data Exchange charges providers 3% on public offers.
Licensed data: Appen, Centific, Copyright Clearance Center, Labelbox Prism and others. Marketplaces: AWS Data Exchange, Databricks Marketplace, Datarade, Defined.ai and others.
- Step 2
Commission what people must make
Labeling platforms give your team the tool; managed services bring the annotators too. Expert networks write demonstrations, rank outputs and build graded tasks, and RL environment builders recreate software such as CRMs and ticket queues for agents to practice in.
Annotation: CloudFactory, Dataloop, Encord, HumanSignal (Label Studio) and others. Expert and RLHF: Bespoke Labs, Halluminate, Handshake AI, Innodata and others.
- Step 3
Fill the gaps with a generator
Synthetic data is strongest where real data is scarce or cannot leave a secure system, and weakest where the generator shares the blind spots you want fixed. Test on real held-out data, never on synthetic alone.
Synthetic: Aindo, K2view, MOSTLY AI (Syntho), Parallel Domain and others.
- Step 4
Prove the data improved the model
Evaluation platforms run test sets against model versions and trace failures in production; red-teaming specialists probe for unsafe behavior. This is the one corner with list prices.
Evaluation: Braintrust, Confident AI (DeepEval), Galileo, Giskard and others.
Disclosure: any company in this list could end up paying us for an introduction that turns into work. Entries, their sequence and our descriptions are written the same way regardless.
What public data deals have been worth
Most licensing deals are private. These figures reached a filing, an earnings release or a named outlet. They show the range, not a price list; the post on deal prices reads them in context.
- $203.0 million
Reddit
Data licensing contracts signed from January 2024, two to three year terms
Reddit S-1, via TechCrunch - $104 million
Shutterstock
AI licensing revenue in 2023
Bloomberg - $23M, then $40M
Wiley
AI licensing revenue, fiscal 2024 then fiscal 2025
Wiley earnings release - About $10 million
Informa (Taylor & Francis)
Initial access fee from Microsoft, plus recurring payments
Informa market update - $10 million bid
Spirit Airlines business data
Google's bid for de-identified staff records; Mercor bid $7.5 million; court approval pending
Reuters - $1 to $4 a minute
Unused video footage
Paid by AI companies for footage
Bloomberg - About $150,000
HVAC company records
Paid by data labeling startups for operating data
The Information - About $25,000
A wound-down startup's code
One founder's codebase after the company closed
MIT Technology Review
As each filing or outlet reported them. Not offers.
Every company in the index
Alphabetical within each type. Each name opens its directory entry, with what its own site publishes about prices and terms.
Disclosure: a company in this index may pay us if we introduce a buyer and work follows. Inclusion, alphabetical order and wording are unaffected. Unaffiliated and unendorsed: none of these companies owns, sponsors or approves this index.
Where the figures on this page come from
- Label Your Data pricing page
- Prolific payment model help article
- AWS Marketplace listing fees
- Labelbox Prism page
- Epoch AI, FAQ on RL environments (interview figures, not contracts)
- Reddit: Reddit S-1, via TechCrunch
- Shutterstock: Bloomberg
- Wiley: Wiley earnings release
- Informa (Taylor & Francis): Informa market update
- Spirit Airlines business data: Reuters
- Unused video footage: Bloomberg
- HVAC company records: The Information
- A wound-down startup's code: MIT Technology Review
Questions buyers ask first
What does an AI training data company do?
It supplies the examples a model learns from. Depending on the company, that means licensing existing data, collecting new data to order, labeling data you hold, recruiting experts to write or rank examples, generating synthetic data, or testing a model against evaluation sets.
Who are the largest AI training data companies?
Scale AI, Surge AI, Mercor, Turing, micro1, TELUS Digital and Appen all sell to frontier AI labs, but few publish revenue, headcount or prices, so a size ranking would be guesswork. The directory reports what each one says about itself.
How much does AI training data cost?
Public starting rates span from two cents for a labeled box at Label Your Data up to hundreds or low thousands of dollars for one graded RL task, going by Epoch AI's interviews. Put your own volume into the cost estimator to see the totals.
Where can I buy AI training datasets?
Marketplaces such as AWS Data Exchange, Datarade, Snowflake Marketplace, Databricks Marketplace and Defined.ai list datasets at prices their providers set. Licensors such as Labelbox Prism, Copyright Clearance Center and the RSL Collective license existing data directly, and TELUS Digital sells off-the-shelf sets. See the marketplaces and licensed data sections.
Is it better to license data or commission it?
License when the value is breadth that already exists; commission when you need behavior no existing dataset shows. Most teams do both. The primer on the three routes compares them side by side.
Is it legal to train AI on data you bought?
That depends on what the license grants, not on the fact that you paid. Check that it names AI training ("internal use" or "analytics" is a different permission), that the seller may sublicense to you, how personal data was removed, and what happens to your model when the term ends. Have your own lawyer read the warranty and indemnity. The provenance checklist lists 21 questions.
Can I sell my data to an AI training data company?
Some take data from outside. Labelbox Prism licenses operating data from companies, exclusive by default; Defined.ai runs a marketplace supplied by contributors; on AWS Data Exchange and Datarade, providers post datasets and set the price themselves. This site works for teams buying data, so we do not broker sales.
Get a quote
ZIP code first, then an email for our answer, then a sketch of the dataset. A person here compares the sketch with the vendors in the index and writes back naming the ones whose own terms fit.
Introductions only: we are not a training data vendor, and the vendor we connect you with may pay us a fee.
Received. Nobody outside this site has it.
- Someone here sets your sketch against the index: data types, pricing basis, published minimums.
- You get an email listing the vendors that match, each with the reason and a link to its own site.
- Approve any of them by reply. Until you do, the brief goes nowhere, and the address you gave is used for nothing else.