Where to sell data: AI labs vs. data marketplaces vs. data brokers

01

Where to sell data: the four channels open to a company

What it means for a company to sell data

Every company leaves records behind as it works. Its staff write emails, chat messages and support tickets, and its software logs orders, prices and customers. Some of those records are worth money to an outside business that can't produce them itself. Selling data means handing a buyer either ownership of those records or the right to use them, in exchange for payment.

A license is a contract that lets a buyer use your data for agreed purposes while you keep owning it. A sale transfers ownership, so the records stop being yours. Three of the four channels below work through licenses; only an outright database sale hands the database itself to the buyer.

Our guide to data licensing vs. selling covers the terms in depth.

The four channels, defined

AI labs and data vendors

AI labs are companies that build and train AI models, such as OpenAI, Anthropic and Google DeepMind. Data vendors are companies that buy records from businesses, anonymize them (replace names and other identifying details) and package them as training material for the labs. A business owner usually signs with a data vendor, not with the lab itself.

Data marketplaces

Online storefronts where a company lists a dataset as a product and other businesses pay a subscription to receive it. AWS Data Exchange, Snowflake Marketplace, Databricks Marketplace and Datarade are the main ones for business data.

Data brokers

Companies that collect personal information about people from many sources, often people they never dealt with, and resell it to marketers, to lenders and insurers who score how risky a person is, and to people-search websites.

Outright database sale

A one-time transfer of a whole database and its ownership to a new owner, which mostly happens when a business is acquired, winds down or goes bankrupt.

Each channel buys a different kind of record

The four channels don't compete for the same records. AI labs and data vendors want records of staff doing their jobs inside a company, such as a support ticket that runs from a customer's complaint to the fix. Marketplaces and brokers want facts about the world outside the company, such as the prices it pays, where shoppers go and who is on its mailing list. The diagram below sorts the records each channel buys.

Why AI licensing belongs in the comparison

Most lists of places to sell data stop at data marketplaces and brokers, because until recently nobody paid for a company's internal chat and tickets. That changed when AI labs began training AI agents, software that carries out a whole job on its own, such as reading a customer's complaint, looking up the order and issuing the refund. An agent learns a job by studying records of people doing that job, and those records sit inside companies' email, chat and ticketing software. Our guide on why AI companies want your data explains the mechanism.

AI labs and data vendors became a new kind of buyer, one that pays for records almost every company already keeps. The comparison below sets this channel against the three older ones on what each buys, what it pays and how much work it asks of the seller.

02

Which channel pays the most for the least work?

Pay and work pull in opposite directions everywhere except AI licensing

Every channel can be measured on two things: how much it pays the seller, and how much work the seller must do to get paid. On a data marketplace, more money means more work, because the seller has to build the dataset and support its buyers. In a database sale, a bigger price usually brings a longer legal review.

AI licensing breaks that pattern, because the data vendor anonymizes, packages and delivers the records itself. The map below places each channel on both measures.

One-time payment or recurring income

The four channels also pay on different schedules. A license to AI labs and data vendors is usually a single payment for a snapshot of past records, and some vendors pay again for new records as staff keep producing them. A marketplace listing earns subscription income for as long as buyers stay subscribed and you keep the data current.

The data value estimate gives a range for your own records, and our guide to what your data is worth explains what raises or lowers the price.

03

Selling data to AI labs and data vendors

Who the buyer is

The end buyer is an AI lab that is training a model to do office work, such as answering customers or processing invoices. Most labs don't buy from small and mid-sized companies directly; they buy finished training material from data vendors. The vendor is the company a business owner actually signs with.

An AI lab sometimes buys directly when one company's archive is large enough to be worth its own deal. In August 2026 Google won the bankruptcy auction of Spirit Airlines, a US budget airline, for the airline's internal data with a $10 million bid, beating a $7.5 million offer from the data vendor Mercor, as Decrypt reported. micro1 then filed a late $12.5 million bid, and the bankruptcy court had not approved a sale as of October 2026. MIT Technology Review put the archive at 100 million emails and 500 million Teams messages.

What AI labs and data vendors buy

AI labs and data vendors buy records that show work being done from start to finish. A support ticket that runs from the customer's first message to the fix is worth more than a folder of loose documents, because it shows each step and how it turned out. The usual sources are team chat, email, shared drives, ticketing systems such as Jira or Zendesk (where staff track each customer request or task), CRMs such as Salesforce or HubSpot (software that logs every contact with each customer), and written procedures.

Data vendors turn those records into training material in two ways. Some records become worked examples, which show a model a task together with the answer an employee actually gave, so the model can check its own attempt against it. Worked examples serve the model as an answer key.

Other records fill a training environment, a realistic copy of a company's software loaded with anonymized records, where an AI agent practices a task thousands of times and is scored on each attempt. It's similar to how trainee pilots practice in a flight simulator.

How the deal works, step by step

The data vendor handles anonymization, the step that replaces names, emails and other identifying details with the same made-up stand-in every time they appear, before anyone uses the records. Our guide on how data is anonymized shows a record before and after.

How the price is set

A data vendor prices records by how much of each job they show, how hard they are to find elsewhere and how well records in different tools connect, more than by how many files there are. A support ticket that links to the customer's emails and the final invoice is worth more than the ticket alone.

micro1 publishes its pay tiers on its data partnerships page, and they sit inside the full $100K to $2M+ range. Polyshares, another data vendor, says its typical deals run $100K to $2M and above.

$100K–$500K

micro1's tier for qualified company datasets.

$500K–$2M

micro1's tier for large datasets, or for companies where several teams keep contributing records.

$2M+

micro1's tier for operating records so rare that no other company could supply them.

micro1 announced its $1B commitment to buying enterprise data in October 2026. Data vendors also pay companies per task, on top of or instead of buying records. The Information reported that the owner of a 14-person HVAC company, a business that installs and repairs heating and air conditioning, is paid about $1,000 per task to grade how AI models handle his firm's routine work, on pace for more than $150,000 a quarter.

The data vendors that buy company data

Several data vendors now run programs that buy records directly from companies. They differ in who they accept and how they pay, so it is worth reading each one's terms.

AI licensing fits companies with years of records across several tools, whether still operating or shut down. Our qualification guide covers the criteria one by one.

04

Data marketplaces: how to sell a dataset on AWS, Snowflake, Databricks and Datarade

What a data marketplace is

A data marketplace is a website where companies list datasets for sale and other businesses buy subscriptions to them. It works much like an app store, with datasets in place of apps. The seller, called the provider, lists each dataset with a description, a sample, a price and license terms. The marketplace helps buyers find the listing, bills them, delivers the data and keeps a fee on each sale.

The buyers are mostly data teams at other companies, the analysts and engineers who feed outside data into their own dashboards and forecasting models. Those teams expect each update on schedule, in a documented format, for as long as they pay.

How the main data marketplace companies work

The large B2B (business-to-business) data marketplaces are built into the cloud platforms, such as AWS and Snowflake, where buyers already store their data. That placement is their main advantage, because a subscriber can search and analyze the purchased data inside its own account, without downloading or moving files.

What sells on a data marketplace

Marketplaces sell data as a product: a dataset that is useful to many buyers at once, updated on a schedule. Typical examples are prices, transaction totals, shipping rates, foot traffic (how many people visit each store), weather and firmographics. A logistics company that records the rates it pays to ship freight on thousands of routes owns exactly this kind of dataset, because every shipper wants to know the going rate on a route.

Personal data is the main thing marketplaces restrict. AWS, for example, requires an extra review, its Extended Provider Program, before a provider may list sensitive or non-public personal information.

The work behind a marketplace listing

A marketplace listing is a small business of its own. The provider cleans the data, writes documentation and a data dictionary (a list explaining what each column means), sets license terms, publishes updates on schedule and answers buyer questions. AWS lists a defined customer support process and a way to keep data regularly updated among its provider requirements.

Revenue depends entirely on whether buyers find the listing and keep paying. That makes a marketplace the right channel for a dataset that buyers need updated again and again. It is the wrong channel for a one-time archive of internal Slack messages and support tickets, because the data teams that shop on marketplaces aren't looking for them.

What is the best data marketplace for businesses?

The best data marketplace for a business is the one its buyers already use. If your likely customers run on AWS, list on AWS Data Exchange; if they run on Snowflake or Databricks, list there; if they search across categories, add a Datarade profile. Many providers list on more than one. Our guide to data monetization covers the wider strategy of building and selling data products.

05

Selling data to data brokers: what they buy and the legal risk

What a data broker is

A data broker is a business that collects personal information about people from many sources and sells it to other companies, usually about people who never dealt with the broker directly. California's privacy agency, CalPrivacy, defines a data broker in those terms. Brokers sell to marketers, lenders, insurers, other brokers and people-search sites, websites that look up a person's address and phone number.

A company that sells to a broker is usually selling information about its own customers, such as their names, emails, phone numbers, addresses and what each person bought. Brokers typically price that by the record or per thousand records, so a list of tens of thousands of customers fetches little next to the other channels.

The consent problem

Customers gave a company their details in order to buy something, not so the company could resell them. A company's privacy policy is a promise about what it will do with that information, and selling contact data against that promise can be treated by regulators as a deceptive practice. The first check, before anything else, is what the privacy policy said at the time each customer signed up.

What the CCPA requires when you sell personal data

The California Consumer Privacy Act (CCPA) gives California residents the right to know what a business collects, to have it deleted and to opt out of its sale. It applies to for-profit businesses that have more than $25 million in annual revenue, buy, sell or share the personal information of 100,000 or more California residents or households, or earn half their revenue from selling it. Several other states have passed similar laws.

Which states make data brokers register

Several US states require data brokers to register each year and pay a fee. California has gone furthest: its Delete Act created a single deletion tool, DROP, that since January 1, 2026 lets residents ask every registered broker to delete their data in one request.

Registration usually falls on the broker that buys and resells the data, not on a company selling its own customers once. A company that sells personal data again and again, as a side business, can start to look like a broker itself under these laws. A lawyer should decide where that line falls. Our guide to whether it's legal to sell company data covers the checks.

A company that has a contact list almost always has years of internal records too. Those records are worth far more licensed to AI labs and data vendors, and because every name is anonymized first, no customer's identity is put up for sale.

06

How to sell a database outright

When an outright sale makes sense

An outright sale makes sense when the company, or a part of it, is changing hands anyway. In an acquisition, the database transfers with the business. In a wind-down or bankruptcy, it is sold as one asset among many, usually to pay creditors.

People who search "sell my database" while their business is still running usually want to be paid for it, not to give it up. A license does that: the company is paid and keeps the database and the right to license it again.

Privacy promises travel with the data

The Federal Trade Commission (FTC), the US agency that enforces consumer protection, treats a company's privacy policy as binding on whoever buys its customer data. In the RadioShack bankruptcy in 2015, the FTC asked the court to protect customers' personal information and set out conditions for any sale, as it had in 2000 with Toysmart, a bankrupt online toy seller.

• Sell the data with other assets, not aloneCustomer data should go along with the business it came from, not as a standalone list.
• Sell to a buyer in substantially the same line of businessCustomers who shopped at an electronics retailer shouldn't end up in an unrelated company's marketing.
• The buyer honors the old privacy policyWhatever the seller promised when it collected the data still applies after the sale.
• New uses need fresh consentBefore using the data in a materially different way, the buyer must notify customers and get their consent.

The FTC took the same line in 2025, when the genetic-testing company 23andMe was in bankruptcy. Its chairman wrote that any buyer of 23andMe's customer data must be bound by the company's privacy promises.

Internal records now sell in shutdowns too

A company that shut down used to sell its customer lists, domain names and equipment. Closing companies now also sell their internal archives of email, chat and code to AI labs and data vendors, usually as licenses or anonymized datasets. A service that helps startups shut down says 47% of its deals now include a sale of code or data, MIT Technology Review reported.

The Spirit Airlines auction shows how an archive is sold in a bankruptcy. The airline's emails and Teams messages are to be stripped of identifying details before the buyer receives them, and the sale needs the bankruptcy court's approval, which was still pending in October 2026 after micro1's late bid. In a bankruptcy the money from each sale goes to the company's creditors, so every dollar the archive brings in pays down what the company owes.

How to sell a database, step by step

1. List what the database holds: tables, systems, years covered and which fields identify people.
2. Read every privacy policy and customer contract that applied when the data was collected.
3. Decide whether a license or a sale fits; if the company keeps operating, a license usually does.
4. Get a value range from more than one buyer, using a data value estimate as a starting point.
5. Have a lawyer review the deal before any sample leaves the building, because handing over even a sample shares confidential records.

07

Which channel fits your company

Two questions decide most cases

The first question is whether the company is still operating, because that decides whether a license or a sale is on the table. The second question is what kind of data it holds: records of its own work, a dataset other businesses would subscribe to, or lists of people. The flow below walks through both.

The quick reference

Using more than one channel

The channels can be combined, because they buy different records. A freight company could list its dataset of shipping rates by route on AWS Data Exchange, and license to a data vendor the email and the tickets its dispatchers, the staff who assign trucks to loads, write each day. The one thing to check is exclusivity: an exclusive license bars you from selling the same records to anyone else, so make sure it covers only the records in that deal.

08

Mistakes companies make when they sell data

• Selling what you could licenseA sale ends your ownership and your chance to license the same records again. An operating company rarely needs to give the database up to be paid for it.
• Sending a sample before legal reviewHanding over a sample shares real records with the buyer. Agree the anonymization rules and the confidentiality terms first.
• Taking the first offerBuyers price the same records differently. A second quote, or a value estimate, tells you whether an offer is fair.
• An exclusive license that's too broadExclusivity should cover the records in the deal, for a set term, not everything the company will ever produce.
• Selling personal data against your own privacy policyWhatever the policy promised when customers signed up still applies to the data, whoever ends up holding it.
• Letting an employee sell company recordsOnly the company can license its data. An employee uploading work files on their own is a breach, not a deal.

09

How to start selling your company's data

Make the list a buyer needs

A buyer prices company records from a short inventory: each tool, the years it covers, roughly how much it holds, and which teams used it. Any dataset that outside companies have asked to buy belongs on the list too, because it may sell on a data marketplace.

Check the rights before you sign

A company has to settle three things before any deal: that it owns the records, that its customer contracts and privacy policies allow the use, and which categories, such as HR or medical records, to leave out. Our guides to the legal checks and data licensing terms walk through each point.

Compare offers on the same terms

Offers from different buyers can only be compared when each is read on the same five points: the price, whether the deal is exclusive, what the buyer may use the data for, how long it keeps raw copies, and when it pays. Real examples of what companies have earned are collected in our guide to data monetization examples.