AI training data marketplaces, buyers, and brokers

By Cody Houser, Founder · Last updated October 24, 2026

The short answer: there are five ways to sell or license data for AI training: dataset marketplaces like Defined.ai and AWS Data Exchange, data licensing companies like micro1 and Labelbox Prism, brokers, wind-down and bankruptcy sales, and direct deals with AI labs. Packaged datasets fit marketplaces. An operating company’s internal records usually go through a licensing company or a broker.

The five routes at a glance

Route	Best for	What you sell	How pricing works	Effort
Dataset marketplace	Data businesses with a repeatable product	Packaged datasets: audio, images, text, feeds	You set list or subscription prices, or earn licensing fees per sale; the platform may charge fees	High: you package, document, and maintain the listing
Data licensing company	Established companies with deep operational records	A license to historical email, chat, files, CRM, and accounting	Buyer quotes per company after scoping, against its published ranges	Medium: export and review; the buyer handles much of the processing
Broker	Operating businesses that don’t know which buyer fits	The same records, matched to a buyer’s current requirements	The buyer’s proposed agreement sets the price	Low to start: a description of the company and its systems
Wind-down or bankruptcy sale	Closing or closed companies	Workspace data, source code, and records left after shutdown	Negotiated through a closure service, or bid at a court-supervised auction	Medium: tied to the dissolution or court process
Direct lab deal	Platforms and publishers with huge, unique libraries	Ongoing access to a content library, sometimes with display rights	Negotiated privately, often paid per year	High: your own legal team and leverage

1. Dataset marketplaces

A marketplace is a catalog. You list a dataset, buyers find it, and the platform handles discovery and often payment. Real examples:

Defined.ai describes itself as an AI data marketplace where enterprises buy or commission training data across speech, text, image, video, and multimodal formats. Its Data Partner program is open to “any business that generates or aggregates data,” pays through “licensing fees from data marketplace sales,” and includes an assessment of your licensing rights. It advertises $600K+ in average partner revenue; that is its own marketing figure.
AWS Data Exchange. Providers register, publish data products, and set offers on AWS Marketplace, and can turn on subscription verification to review buyers before they get access. Data sets must meet the legal eligibility rules in AWS’s seller terms. AWS charges tiered fulfillment fees on revenue it collects for new subscriptions, and buyers pay the subscription or pay-as-you-go prices the provider sets.
Datarade, founded in Berlin in 2018, is a platform for companies to discover, compare, and connect with data providers. It already listed more than 1,500 providers at its 2019 seed round. It is mainly a discovery and matching channel.

The catch for an operating business: a marketplace listing is public and repeatable. Years of internal email and client files aren’t a product you can describe in a listing, and you wouldn’t want them advertised.

2. Data licensing companies that buy operational records

A newer group of companies licenses the everyday records of real businesses and turns them into training and evaluation data for AI labs. Each publishes its own terms:

micro1 says it looks for established companies with 30+ full-time employees and several years of records, mainly in English. It publishes ranges from $100K–$500K up to $2M+, and says it pays in full within 30 days of accepting the data.
Labelbox Prism licenses work documents and operational data from systems such as Salesforce, Jira, Slack, and SAP. It says compensation starts around $50K for targeted datasets and reaches $1M+, paid as an upfront fee plus a revenue share, and that its licenses are exclusive by default.
Nyne licenses anonymized accounting, CRM, scheduling, and project records. Its current program focuses on companies with at least 20 full-time U.S. W-2 employees and three years in business, and it says partnerships are commonly worth $100K or more.
Protege supplies private, real-world data to AI developers, with a focus on healthcare, video, audio, and motion data, and invites data holders to license existing data assets with rights protections.

These companies differ on eligibility, exclusivity, and how they pay. Compare their published ranges in how much company data is worth.

3. Brokers

A broker doesn’t buy your data or run a public catalog. It assesses what a company holds, compares it with buyers’ current requirements, and introduces the company when there’s a fit. The buyer still proposes the agreement, and you decide whether to sign.

A broker earns its keep when you don’t know which buyer wants your kind of records, or when you’d rather not contact several data companies yourself. Ask how the broker is paid and whether its fee depends on a deal closing. If you introduce companies yourself, see how to become an AI data broker.

4. Wind-down and bankruptcy channels

When a company closes, its data can become a separate asset.

Closure services. SimpleClosure, which handles startup shutdowns, runs AssetHub to “sell or license source code, workspace data, and other startup assets to vetted buyers.” Forbes reported, via Gizmodo, that it processed nearly 100 such deals in a year, paying $10,000 to $100,000 per company.
Bankruptcy auctions. Google agreed to pay $10 million for part of Spirit Airlines’ business data in its bankruptcy, with Mercor the runner-up at $7.5 million. The sale needed a bankruptcy judge’s approval.

These channels only open at the end of a business. If you’re closing, read the closing-business data guide before you wipe any systems.

5. Direct deals with AI labs

Big labs sign licensing deals directly, but the public examples are platforms and publishers: Google with Reddit (about $60 million a year, per a Reuters source), OpenAI with News Corp (reported at more than $250 million over five years), and Microsoft with Taylor & Francis (more than $10 million in the first year).

Those sellers brought massive, well-known libraries and their own deal teams. A mid-sized company calling a lab cold has little leverage and no obvious contact, which is why most operating businesses use one of the other routes.

Questions to ask any buyer or platform

Who is the end user of the data, and what uses does the license permit?
Is the license exclusive or non-exclusive, and for how long?
How is the price set, and is your number an estimate or a written offer?
When are you paid: at signing, on delivery, on acceptance, or over time?
What does “acceptance” mean, and can it reduce the payment?
Who removes personal and confidential information, and to what standard?
Can you exclude categories such as client files, HR records, or privileged material?
How long is raw data kept, and how is deletion confirmed?
What fees does the platform or intermediary take, and who pays them?

The AI data licensing agreement checklist covers the clauses behind each answer.

Red flags

A firm price before anyone has seen a description of your records. Nyne itself says an estimate is “not a written commitment to pay.”
Requests for credentials, raw exports, or sample files before an agreement and a rights review.
No answer on who the end user is or what the data may be used for.
Vague de-identification with no named standard and no written confirmation that it was done.
No retention or deletion terms.
An intermediary presented as the buyer when it’s actually reselling, or the reverse.

Why a broker suits operating businesses

If your business is still running, you probably want a quiet, controlled process: no public listing, no cold calls to labs, and no files leaving the company before you’ve seen terms. That is the case for a broker.

Hyperdrive Data is a broker. It helps established businesses assess whether their email, file storage, internal chat, CRM, and accounting records fit current AI buyer requirements, and connects them with buyers. The first conversation is a description of your company and its systems, with no files, credentials, or personal information. The buyer’s proposed agreement sets the terms, you decide whether to proceed, and you keep your business. Check whether your company fits, or read how to sell data to AI companies for the full process.

Questions people ask

What is an AI training data marketplace?

An online platform where data providers list datasets and AI developers browse, sample, and buy or license them. Defined.ai and AWS Data Exchange are examples. They suit packaged, repeatable datasets that can be described in a listing.

Where can I sell my company’s data for AI training?

Five routes exist: dataset marketplaces, data licensing companies that buy operational records, brokers who match companies with buyers, wind-down and bankruptcy sales, and direct deals with AI labs. Internal business records usually go through a licensing company or a broker, because they need assessment and rights review before anyone can price them.

Which companies buy business data for AI?

Several data companies publish programs for operational records, including micro1, Labelbox Prism, and Nyne. Each sets its own eligibility rules and price ranges on its own site. Large AI labs have licensed content directly from publishers and platforms such as Reddit and News Corp.

Is a data marketplace or a broker better?

A marketplace is better for a clean, packaged dataset you want to sell repeatedly. A broker is better for years of email, files, chat, CRM, and accounting records from an operating business, which need assessment and a matched buyer rather than a public listing.

Do I have to be closing my business to sell its data?

No. Wind-down sales get the headlines, but data licensing companies and brokers work with operating businesses that keep running.

Sources

This guide is general information, not legal, tax, or financial advice. Review any proposed agreement with your own advisers.

Related guides

Check whether your company is a fit. Start with a description of your records; no files needed.