Know a company with years of records? Refer it and earn up to $100K
For dataset owners

Sell a Dataset You Built: Licensing It to AI Companies

Sell a dataset you created to AI companies for training: what buyers check first, how to send a safe sample, exclusivity and how datasets are priced.

Short answerChecked

To sell a dataset you built to AI companies, you license it: the buyer gets written rights to train on a defined copy, and you keep ownership. Buyers check where every record came from, who holds the rights, how any personal data was collected, and a sample before they talk price. Most datasets offered are never licensed. First-party data with clean rights and good documentation does best; scraped or resold collections rarely pass review.

House keys on a lease folder and an inspection checklist
Key points
  • Selling a dataset to AI companies almost always means licensing it: you keep ownership and the license sets the uses, term and fee.
  • Buyers review provenance, rights, personal data, documentation and a sample before price, so write those answers down first.
  • Send a small, marked sample under a short evaluation license and keep the full set in your systems until a license is signed.
  • Non-exclusive licenses let you license the same dataset to several buyers; exclusivity should be narrow, short and paid for.
  • Public prices are rare and come in different units: per title, per hour, per record or by subscription.

What do buyers check before you sell a dataset to an AI company?

Buyers ask the same handful of questions whether you are selling datasets on a cloud marketplace or licensing one straight to an AI developer. Each row below comes from a buyer's, marketplace's or standards body's own published page, read on 20 November 2026. Answer them in writing before anyone asks for a price.

What a buyer checksWhat the source saysSourceDate checked
How the data was createdProvenance means knowing where data comes from, how it was created and whether it can legally be used, recorded in 22 fields grouped as Source, Provenance and UseData & Trust Alliance, Data Provenance Standards v1.0.020 Nov 2026
Rights chain and license to useThe same standard has fields for the license to use, the intended use and whether proprietary data is presentData & Trust Alliance20 Nov 2026
Personal data and noticeNo data that identifies a person unless it is publicly available information, and listings must say if data was aggregated or anonymizedAWS Data Exchange publishing guidelines20 Nov 2026
Third-party contentThe court found training on books was not itself illegal, but acquiring them from pirate websites was wrongful; the $1.5 billion settlement, about $3,000 per work, got final approval on 20 July 2026AP News; the Authors Guild20 Nov 2026
DocumentationA dataset card records license, language and size; a data dictionary lists each column and its meaningHugging Face Hub docs; AWS Data Exchange docs20 Nov 2026
SampleA small part of the product meant to show what the whole set is like, up to 10 samples per data set at 50 MB eachAWS Data Exchange docs20 Nov 2026
Usage terms and exclusivityThe provider sets what subscribers may and may not do in a written subscription agreement, and can make private offers to selected customersAWS Data Exchange docs20 Nov 2026

We work only with data you created in running your business, or hold under a written licence that lets you license it for AI training. We do not take scraped web data, purchased contact or consumer lists, people-search or location data, or personal data collected without a notice that covers this use. That is Origin's own rule, and it matches what the sources above ask for. A dataset you built yourself has a provenance story you can tell in full. A collection gathered from other people's sites or bought from a list vendor usually does not.

What does it mean to license a dataset for AI training?

Licensing a dataset for AI training means you give a buyer written permission to use a defined copy of your data to train, test or tune models, for a set term and fee, while you keep ownership. You are not handing over the asset. The license decides what the buyer may do, which is why two buyers can pay very different amounts for the same rows.

In practice the license names the dataset version, the permitted uses (training, evaluation, retrieval), whether the buyer may pass it to contractors or related companies, how long it runs, and what happens to copies when it ends. A model already trained on the data is often kept by the buyer, so ask what the deletion clause covers. The guide to what data licensing is explains the general terms, and the AI licensing guide shows how content deals are shaped. This page stays with datasets you made yourself.

How do I package a dataset so AI labs will buy it?

Package it so a reviewer can say yes without a call. People who review datasets for machine learning spend most of their time on the paperwork, not on the rows. A clean package has five parts.

  • A dataset card. On Hugging Face a dataset card is the README that documents the data, and its metadata records the license, language and size (Hugging Face). Write one even if you never post the data there. See the dataset card glossary entry.
  • A provenance sheet. Fill the fields of the Data Provenance Standards that apply to you: who issued the data, the range of dates it was generated, the method, the format, where consent documentation sits, the license to use and the intended use. IBM tested these standards in its clearance process for datasets used to train foundation models.
  • A data dictionary. One row per column or field, with its meaning and units. AWS Data Exchange asks providers to describe which columns the data set holds and what they mean (AWS).
  • A rights file. For each source inside the set, note who created it, under what employment or contract terms, and any third-party license. Mark anything you cannot clear and leave it out.
  • A privacy note. Say what personal data the set ever held, the notice it was collected under, and how it was removed or masked before anyone else sees it.

Format matters less than consistency. Use open formats a buyer can load without your software (CSV, Parquet, JSON Lines, WAV, MP4), keep one schema across files, and give every record a stable ID so a buyer can report problems row by row. Then count what you have: records, hours or files, the date range, and the share that is labeled. The data provenance glossary entry defines the terms reviewers use.

How do I send a sample of my dataset to a buyer without giving it away?

Send a small, representative slice under a short evaluation license, and keep the rest in your systems until a full license is signed. Marketplaces build samples in for this reason. AWS Data Exchange defines a sample as a small part of the product that shows what the whole data set is like, and allows up to 10 samples per data set at 50 MB each (AWS). Snowflake requires a trial for paid listings offered publicly on its marketplace (Snowflake). Hugging Face lets authors gate a dataset so each user must request access, and the authors keep full control of who gets in (Hugging Face).

Patronus AI publishes this kind of term: only two named reviewers may evaluate the sample, and it is deleted if no license follows (our Patronus AI review).

  • Size: enough rows to judge quality and coverage, drawn across the whole date range, and small enough that it has no value on its own.
  • Held back: leave out the records a buyer would most want, such as your rarest cases or the newest year. Describe those in counts instead.
  • Marked: add a few unique canary records, or keep a hash list of the rows sent, so you can recognize the sample if it turns up elsewhere.
  • Licensed: a one or two page evaluation license that allows review only, bars training and onward sharing, sets a deletion date and names who will see it.
  • Logged: one sample version per buyer, so you always know who received what and when.

Can I license the same dataset to several AI companies?

Yes, as long as every license you sign is non-exclusive. A cloud marketplace listing works this way by design: one product serves every subscriber who accepts its terms. It is also how most dataset owners earn more than once from the same work. An exclusive license promises you will not license the data to anyone else, so it closes that door for its whole term.

If a buyer asks for exclusivity, narrow it before you price it: one field of use (for example speech recognition only), one territory, a short term, or one slice of the data. Ask for a higher fee or a minimum guarantee in return. The exclusive versus non-exclusive guide lists the clauses to check. Before you sign anything new, reread your earlier licenses for any promise of exclusivity you already made.

How much is my dataset worth?

Nobody can price a dataset before seeing a sample and the rights file, and most datasets offered to AI buyers are never licensed at all. No AI lab we checked publishes a price list for datasets it licenses from outside suppliers. What is public is a set of pricing units and a few dated figures. Use them to choose a pricing model, not to value your own set.

Pricing unitPublic exampleKind of figureSource and date
Per title, fixed termHarperCollins offered $5,000 per title for a three-year AI license, split 50-50 with the author; one author reported his share as $2,500Reported termsThe Authors Guild, 19 Nov 2024; The Verge, 18 Nov 2024
Per hour of mediaVeritone says it can charge $1 to $1,000 per hour of video, depending on the use, tagging and metadataCompany statement in trade pressThe Media Leader, 29 Jan 2026
Per record, with a capPathos AI pays Tempus $100 per delivered de-identified record, within fees of up to $55 million over five yearsSigned deal, SEC filingTempus Form 8-K, 23 Oct 2026
Subscription or usageSnowflake Marketplace paid listings use either a usage-based plan or a subscription-based planMarketplace pricing modelSnowflake docs, read 20 Nov 2026

Four things move the price more than size does: whether the data can be had any other way, how cleanly you can show the rights, how much structure and labeling it already carries, and whether the buyer gets it alone. The pricing models guide explains flat fee, recurring and revenue share terms, and the AI training data price index lists disclosed deals with their units.

Do AI labs buy web-scraped datasets?

The buyers and rules we checked point away from them. Labs already gather public web data themselves: Mercor, a company data buyer, writes that AI development has relied heavily on publicly available information, including websites, books and code. What labs lack is data they cannot collect, so a scraped set offers little they could not get, and it carries the source sites' rights and people's personal data with it.

The risk sits in how data was acquired, not only in what it contains. In the authors' case against Anthropic, the court found that training on books was not itself illegal but that acquiring millions of them from pirate websites was wrongful, and the $1.5 billion settlement, about $3,000 per work (AP), received final approval on 20 July 2026 (the Authors Guild). For personal data, the UK Information Commissioner's Office calls web scraping for generative AI training a high-risk, invisible processing activity, and expects developers to show why other ways of collecting data are not suitable (ICO). A buyer that has read those lines asks every seller where each record came from.

That is why the screening rule near the top of this page excludes scraped data, purchased lists and data collected without a notice that covers this use. If your dataset mixes your own records with scraped material, the part you created can often be separated and offered alone.

What AI training datasets are in short supply right now?

No public measure of shortage exists, so the best evidence is what buyers say they want, in their own words. Mercor writes that demand is growing for proprietary enterprise data. Scale AI asks for real operational workflows, not demos or synthetic data, and for domain depth that makes data hard to replicate. Both point at the same gap: records of real work that never reached the public web.

  • Work records with outcomes: tickets, cases, jobs or projects that show what was done and how it ended.
  • Expert-made material: annotations, reviews and decisions by trained people in a narrow field.
  • Sensor, audio and video recorded for a purpose, with notes on what was captured and when.
  • Long histories from one system, so a model can see how work changed over several years.

Generic sets that anyone can assemble, such as public price lists or tables copied from government statistics, sit at the other end. They may still sell to analysts through a dataset marketplace, but AI buyers rarely pay for what they could rebuild in a week. The guide to who buys company data lists buyers by record type.

How do I find buyers for a niche dataset?

Start from who would use the data, then pick the route. A niche dataset usually has a short list of real buyers: AI developers building for that field, the data programs that supply them, and analysts who buy through cloud marketplaces. Four routes are open on 20 November 2026, and each has its own rules.

  • Cloud data marketplaces. AWS Data Exchange, Snowflake Marketplace and Databricks Marketplace let a provider publish a product to subscribers on that cloud, under written rules on eligibility and personal data. The data marketplaces comparison covers fees and listing steps.
  • AI data marketplaces and company data programs. Defined.ai describes its data partners as any business that generates or aggregates data, earning licensing fees from marketplace sales. Scale AI, Mercor and Handshake AI run programs that license company records to AI developers.
  • AI developers directly. Few publish a door for outside datasets. OpenAI's Data Partnerships page, read on 20 November 2026, says the program is not accepting submissions at the moment.
  • An independent data partner, who works on your side on rights, packaging and terms, and puts the dataset in front of more than one buyer.

We help you license your data. Origin Data Partners is an independent data partner that helps business owners value and license their company records: the company pays Origin nothing, and a buyer may pay Origin when a deal completes. Origin is not a marketplace, and the license is always signed between you and the buyer. If you built the dataset yourself, the two-minute estimate is the first step, and nothing leaves your systems until you decide.

Not the owner? If a client or employer holds a dataset like this, advisers can introduce the company.

How we checked this

We read every source on this page on 20 November 2026: the marketplaces' own provider documentation (AWS Data Exchange, Snowflake, Hugging Face), a published industry standard (the Data & Trust Alliance), a regulator's guidance (the UK ICO), the buyers' own pages (Mercor, Scale AI, Handshake AI, Defined.ai, OpenAI), a company's own deal announcement, an SEC filing (Tempus), and the Authors Guild and news reports (AP, The Verge, The Media Leader) for dated figures. We searched for later news on each deal and case, which is how we found the final approval of the Anthropic settlement. Each fact is quoted word for word in our evidence file.

What we could not verify: the terms of private dataset deals, which are almost never published; whether any buyer named here would license a given dataset; and two widely cited news reports on per-minute video prices, whose pages blocked automated reading, so we left them out. Company statements and reported offers are labeled as such in the price table.

What to write down before you contact any buyer

Write one page: what the dataset is, how and when you made it, how many records it holds, which rights you hold in each part, and what personal data it ever contained. Add your dataset card and data dictionary. That page answers most of what a buyer asks first, and none of the data has to move to produce it.

Origin Data Partners works with companies that own data and helps get it bought. We are independent of the services on this page, and a buyer may pay us when a deal closes (how Origin is paid). See what your data is worth.

Common questions

I have a dataset. How do I sell it to AI companies?

Confirm you created it or hold a written license that allows AI training. Then write a dataset card, a provenance sheet and a data dictionary, cut a small evaluation sample, and decide whether you will license it non-exclusively. Choose a route: a cloud marketplace, an AI data program, a direct approach to a developer, or an independent data partner. Expect provenance and sample review before any talk of price.

What is a dataset card, and do AI buyers expect one?

A dataset card is a short document that travels with a dataset and says what it contains, how it was made, its license, language and size, and its known limits. Hugging Face renders one as the README of each dataset repository. AI buyers do not share one template, but most review questions map onto a card, so writing it first saves a round of email.

Can I sell a dataset built from my customers' records?

Only if the notice your customers saw, and your contracts with them, allow it, and usually only after personal data is removed. Read your privacy notice and customer terms first. AWS Data Exchange, for example, bars data that identifies a person unless it is publicly available information. Many owners license a de-identified version and keep anything that names a customer out of scope.

Is selling datasets on a data marketplace the same as licensing one to an AI lab?

Not quite. A cloud data marketplace lists a product that any approved subscriber on that cloud can buy on standard terms, often by subscription or usage. A license to an AI developer is usually negotiated one to one, with use limits, deletion terms and sometimes exclusivity. Both are licenses, but the marketplace route trades negotiation for reach.

Can I sell a dataset that is already public on Kaggle or Hugging Face?

You can license your own copy, but a buyer can usually download the public version for free under its open license, so there is little left to sell. Exclusivity is impossible once a set is public. What can still have value is what you never published: newer records, the full version, labels, or the raw material behind the public sample.

Which file formats do AI buyers prefer for a dataset?

Open formats a buyer can load without your software: CSV or Parquet for tables, JSON Lines for records and text, and standard audio and video containers for media. Consistency matters more than the choice: one schema across files, stable record IDs, documented units and a data dictionary. Convert proprietary exports before you send a sample, and note the conversion in the dataset card.

Sources and further reading

What is your company’s data worth?

AI companies pay $20K to $5M for company records. Six questions, no files.