AI Training Data Marketplaces Compared: Fees and Payouts

An AI training data marketplace lists data that companies and creators hold and licenses it to AI developers, keeping a share of each sale: from 5% of a sale above $250,000 at Opendatabay and 8 percent at FileYield to 50 percent of a dataset deal at Wirestock, while Protege, Troveo and Defined.ai publish no share at all. What separates them matters more than the headline cut: who sets the price, whether the licence is exclusive, who sees the data before a sale, and when the money arrives, which at some platforms is 4 to 5 months after delivery.

This guide compares fourteen of them on their own published terms: what each lists, what it keeps, who prices the data, what licence it takes, how and when it pays, where a holder may list from, and how much each has actually sold. It is written for the founder, CFO or counsel of a company deciding where to put a dataset, an archive or a codebase, and for anyone weighing a marketplace against a buyer's own programme or a broker. Every term below links to the record of the platform's own page.

Contents

1. The marketplaces at a glance
2. Two kinds: listing sites and managed licensors
3. What each one keeps from a sale
4. Who sets the price, and who sees the data
5. Exclusivity: the clause that moves the price most
6. When and how a marketplace pays
7. What you must do before you list
8. Who can list, and from where
9. Which marketplace fits which data
10. Which marketplaces have sold at scale
11. A marketplace, a buyer's programme or a broker
12. Questions people ask
13. What this means for you

1. The marketplaces at a glance

The table below sets the platforms side by side on the five terms a holder reads first. Each cell is the platform's own wording, linked to the record of the page it comes from; where a cell says "not published", the platform's pages give no figure and the only way to learn it is to ask.

Two cautions before reading it. First, a listing count is not a sales count: a platform with thousands of listings may sell few of them, and the figures in the last column are of different kinds (money paid out, money that changed hands, datasets listed). Second, a share that is not published is not a small share: the platforms that negotiate on a holder's behalf price each deal separately, so the share is a term to settle in writing before signing.

Two names often searched beside these are not in the table as places to sell. Hugging Face holds 1,100,828 public datasets: it is where datasets are published openly, not where they are sold, and NVIDIA agreed to buy it in September 2026. Human Native, a UK marketplace for licensed AI data, was acquired by Cloudflare in January 2026, which tied it to its Pay Per Crawl work for web publishers (section 2).

2. Two kinds: listing sites and managed licensors

The platforms divide by who does the work of a sale. On a listing site the holder writes the listing, sets the price, answers buyers and usually delivers the data itself; the platform provides the shop window, the search and sometimes the payment. On a managed licensor the holder hands over its content, and the platform prepares it, finds the buyer, negotiates the licence and pays the holder its share. The difference decides how much of the holder's time a sale takes, how much control it keeps, and how much of the price it sees.

The listing sites are Datarade, Opendatabay (in its self-serve mode), FileYield, DataVendor, AWS Data Exchange and Snowflake Marketplace. On most of them the holder delivers the data itself: Datarade does not host the data and leaves delivery to the holder, and at FileYield the data transfer, the sale contract and the payment happen directly between holder and buyer. They suit data that is a product: something the holder can sell to many buyers, again and again, often refreshed. Every data product read for our market records sells the same records to any buyer, mostly on subscriptions that keep the data fresh, and none of its makers prices exclusivity.

The managed licensors are Protege, Troveo, Defined.ai, Wirestock and Created by Humans. Troveo licenses on your behalf under terms you approve; ownership never transfers; Protege prepares datasets with the partner, including de-identification and certification; Created by Humans negotiates agreements for your approval. Their value is pooling: a lab rarely wants one creator's footage or one hospital's scans, but it will pay for a set built from many of them, and the platform builds it. Kled works the same way for individuals rather than companies, through a phone app.

A third group licenses published content for AI answers rather than for training. Microsoft's Publisher Content Marketplace and Cloudflare's paid-crawling products pay publishers when an AI product reads their pages: Microsoft says publishers will be paid on delivered value, and Cloudflare's Pay Per Use lets a buyer make an offer that the publisher accepts, declines or counters, with payment following reported use. These suit a publisher whose archive is public and current; they are a different deal from licensing an archive for training, which the guide to news publisher licensing deals covers.

3. What each one keeps from a sale

Marketplaces charge in four ways. A percentage of each sale is the most common: Datarade, Opendatabay, FileYield, DataVendor and Wirestock publish theirs. A share that is negotiated and not published is the rule at the managed licensors that price each deal: Protege pays a revenue share when developers license your data; the percentage is not published, and Troveo pays a share of the revenue on every sale; the share is not published. A stated split to the creator is Created by Humans' model, at 80 percent of proceeds to the author. Rewards the platform decides is Kled's: its terms pay for content Kled accepts, in a form Kled decides, and Kled may rule that content earns nothing.

The published percentages differ in shape as well as size, so they are worth reading closely:

• Datarade offers three plans: $0 a year plus 30 percent, $6,000 a year plus 20 percent, or $12,000 a year plus 15 percent. A paid plan pays for itself only once a year's sales are large enough that the points saved exceed the fee, so a holder with one dataset and no sales history starts on the free plan.
• Opendatabay slides with the size of the sale: its cut is 30% up to $100, falling to 5% above $250,000, so it is cheap for a large sale and expensive for a small one. In a brokered sale the terms change: the holder gets the agreed provider price and Opendatabay sets its own price to the buyer, so the holder never learns what the buyer paid.
• FileYield charges both sides: 8 percent from the holder and 6 percent from the buyer, and its own fee will be processed through Stripe once payments exist.
• DataVendor, run by HUD for code and task data, keeps 20% of vendor earnings, charged at withdrawal and never on referral fees, plus Stripe's cross-border transfer fee.
• Wirestock keeps 50 percent of a dataset deal, against 15 percent on ordinary licences, a change it made in May 2023.

The stated creator splits sit at the other end. Created by Humans' 80 percent is reduced, its terms say, by processing fees and the Created by Humans fee, described in your agreement or on the platform. The Authors Guild calls 75 to 85 percent to the author fair, which puts Created by Humans inside the range writers' own body accepts.

Compare offers net of the cut and the payout rules, not on the headline price. Two code platforms quote on opposite bases: DataVendor's listed price is before its 20 percent, while Pangea's partner quotes the holder's side with nothing deducted. A higher listed price at a platform that keeps half can leave the holder with less than a lower price where the holder keeps most. The offer check puts a written offer's net figure beside the market.

A worked example

Take, as an example, a mid-sized company with several years of recorded customer calls in English that it has the right to license, with every caller's consent recorded. It has three marketplace routes.

On a listing site such as Opendatabay or Datarade it writes the listing and sets an asking price. Comparable listings give it a reference: an English podcast set of 1,165 hours asks $5,660 and a German call-centre set of 212 hours asks $1,320. Those are asking prices, not sales. Before it lists, it must remove the callers' personal data, because datasets holding unanonymised personal data are prohibited; after a sale it delivers the data itself and collects payment under its own contract with the buyer.

With a managed licensor such as Defined.ai or Protege it shows consent, documentation and the right to commercialise the data for AI training, and the platform packages the recordings, finds buyers and proposes terms for its approval. It does less work and sees fewer of the numbers: the platform's share is not published, so it must ask for it in writing before it signs.

The third route is to skip marketplaces: a buyer's own programme for company records, or a broker that takes the archive to several buyers at once (section 11). Recorded calls of a company's own customers are also company records, and the audio guide and the speech price guide set out what each route has paid an hour.

4. Who sets the price, and who sees the data

On a listing site the holder sets the price, and the price is public. That helps a buyer more than a holder: it anchors the negotiation at the asking price, and it shows competitors what the holder thinks its data is worth. Published prices also fall steeply with volume: the same record costs 20 to 44 times less at volume across the data products that publish a ladder, so a holder that copies an entry-tier price per record overprices its data badly. And a listed price is a ceiling, not a sale. SimpleClosure's test marketplace for closed startups' assets recovered about $1M across nearly 100 deals, about $10K each, against far higher quoted ranges.

Who learns the buyer's name differs too. On DataVendor a buyer who makes an offer is named to you, so you choose whom to accept; a buyer who buys outright or takes a sample stays anonymous. Its buyers see metadata, redacted file trees and quality evidence, never your raw files before buying, and a sample is a paid, random draw of one item that earns the buyer no discount on the rest. Datarade redacts any sample-preview columns that may contain personal data, and every listing must link its own terms of use, which means the holder writes the licence itself.

At a managed licensor the platform prices, and the holder approves. That gives the holder a professional's price at the cost of seeing the deal only through the platform: Troveo says a buyer will usually test a sample before committing, and buyers receive structured, anonymized data with a manifest of exactly what is included, which Troveo prepares. How a buyer reaches its first number is the subject of the guide to how buyers calculate an offer; the guide to proving quality without handing data over covers what to show before any file moves.

5. Exclusivity: the clause that moves the price most

An exclusive licence promises that nobody else will get the same data, and buyers pay for that promise: our valuation shows an exclusive licence at 2.5 to 5 times a non-exclusive one, and two founders of training-environment companies told Epoch AI that exclusive deals cost roughly 4 to 5 times non-exclusive ones. On a marketplace the question is sharper, because there are two different exclusivities: exclusivity to a buyer, which the buyer pays for, and exclusivity to the platform, which only stops the holder selling elsewhere.

The second is worth much less. Where exclusivity means sole distribution through one platform, platforms pay suppliers about 1.25 to 2 times the non-exclusive rate. A holder that signs its data to one marketplace exclusively gives up every other channel for that premium, so the terms of the exclusivity matter as much as the share.

The platforms' terms range from none to total:

• Kled takes the most. Its licence is transferable, worldwide, perpetual and irrevocable, and Kled may sublicense it through several tiers; it is exclusive to Kled; the contributor promises not to let anyone else use the content to develop, train or test third-party AI models; and if consent is withdrawn, you must refund what you were paid.
• Troveo reports that around 95 percent of its more than 7,000 licensors signed exclusively, and its terms make exclusivity time-limited and earned. Its own guide names a term, a scope and a performance trigger as the three limits that make an exclusive deal safer.
• Protege explains in its own case study why a platform wants it: holding exclusive AI-training rights to a partner's catalogue let it price the data where buyers could not negotiate with anyone else.
• Defined.ai and Protege both give their buyers narrow licences: Defined.ai's is non-exclusive, non-sublicensable and non-transferable, so buyers cannot pass your data on, and Protege's is non-exclusive, non-transferable and non-sublicensable, limited to training their own models. A narrow licence to the buyer keeps the holder free to license the same data again.
• Datarade leaves it to the holder: you keep ownership and decide how your data is licensed.

Non-exclusive data can earn more than once. Troveo advertises an average of 4 sales per dataset, a figure it labels illustrative. The guide to exclusive and non-exclusive licences and the guide to licence term length set out how to price and bound each.

6. When and how a marketplace pays

Almost no marketplace pays at signing. The money moves when a buyer pays the platform, and then on the platform's own schedule, so a holder should read the payout terms as closely as the share. The terms each platform publishes:

Troveo's terms show one more thing to check: it pays on accepted hours, a subset of the hours it submits to a buyer, so the hours a holder delivers and the hours it is paid for can differ. Ask every platform what counts as accepted, who decides, and how long that takes.

Where the payment runs directly between holder and buyer, as at Datarade and FileYield, the holder carries the buyer's credit risk and the collection itself, under the contract it wrote. The guide to when AI data buyers pay sets out what starts the payment clock at each buyer and what to ask for.

7. What you must do before you list

Every marketplace puts the rights on the holder. Datarade requires that your business must hold the necessary rights, licences or consent to distribute the data; at DataVendor submitting a listing confirms you own the supply and may license it; Protege says its partners typically license content they own outright or control under global all-media or similarly broad rights. A listing is a promise about rights, and the promise is the holder's, so the rights work comes first. The rights check asks the questions a buyer's lawyers will, and the guide to what buyers check before they buy lists them.

Removing personal data is the holder's job on the listing sites and the platform's on some managed ones. On DataVendor removal is your job: take out secrets, credentials, customer information and anything you have no right to share before you connect or upload, and its listing checks cover credentials and personal data in the listing text. Kled requires that the content holds no personal data of anyone under 18. Troveo, by contrast, connects to none of your systems: you run the exports, and it does the anonymisation and packaging. Our de-identification standard and the de-identification tool do this on the holder's own computer.

What sells is what a buyer cannot get elsewhere. DataVendor will not list public repositories, since a buyer can already clone them, and the labs say the same of everything: they pay for what is not on the open web. A listing should say what is in the data, how much, over which years and what is excluded; the datasheet builder produces that page, and the guide to preparing data for sale covers exports that keep their history.

8. Who can list, and from where

Most marketplaces say little about where a holder may be based, and the few rules that exist are uneven. AWS Data Exchange is the strictest: a provider must be a citizen or resident of, or a business organized in, one of 13 listed places, while AWS Marketplace's general page lists 16 places for paid products and an older FAQ names only the United States and EU member states. A provider outside the US files a W-8 and a VAT or GST number, and needs US bank details, which it can get through a virtual US bank account from Hyperwallet.

The others are open with limits. Datarade's provider form lists every country in the world, but its publishing policy forbids selling data to countries under international sanctions. Opendatabay is a UK-registered marketplace that says it serves users worldwide, under UK law. Kled's terms say the service is run from the United States, a notice rather than a ban, and contributors who earn $600 or more in a year must file a W-9 or a W-8BEN before cashing out. FileYield and Wirestock leave taxes to the user and to the contributor. Protege builds datasets from more than 150 providers worldwide.

A holder outside the US should also expect US withholding questions on licence payments; the guide to US withholding tax on data licences explains them, and the guide to buyers that take holders outside the US lists who does.

9. Which marketplace fits which data

The platforms are specialised far more than their homepages suggest. Each pools the kind of data its buyers already ask for, so the first question is not which platform is biggest but which one sells your kind of data. The table maps the common holdings to the platforms that take them and to the guide that covers each in depth.

Two rows need a word. Business data products sell in volume for very little each: scraped and compiled records sell non-exclusively by the record for fractions of a cent to a few dimes, so they fit a listing site with a subscription model and do not fit a one-time sale. An AWS listing can still be large: SafeGraph's US places data is listed at $150,000 for 12 months.

Company records are the reverse: few marketplaces take them, because each archive is unique, carries customers' and staff's personal data, and is priced by how many training tasks it yields rather than by the record. DataVendor delivers them as one ZIP and asks first for a short inventory: asset type, origin, who owns it, your right to license it. Troveo's licensing guide asks for a list of your systems and the years each covers, your customer contracts and employee agreements. Most buyers of company records, though, are data vendors with programmes of their own, and the comparison of those programmes sets their terms side by side.

10. Which marketplaces have sold at scale

A marketplace earns its share only if buyers come to it. The evidence of that is uneven, and it is worth sorting by kind before trusting any of it. The strongest is money that moved, stated by the platform with a buyer or an investor behind it. Next is a platform's own claim about its partners' earnings or its users. Weakest is a count of listings, which says how much is on the shelf, not how much has sold.

By that measure the record shows:

• Money that moved. Protege's transaction volume, what buyers paid through it, was $30 million in 2025, with at least $100 million expected in 2026; it raised $25 million and $30 million. Troveo has paid more than $50 million to more than 7,000 licensors for more than 7 million hours. Wirestock reports a $40 million annual run rate and $15 million paid to contributors.
• The platform's own claims. Defined.ai states $600K+ average partner revenue, its own figure with no basis given. Kled claims 500,000+ contributors and announced a $12 million licence without naming the buyer.
• Listings. AWS Data Exchange holds more than 3,500 products from over 300 providers, Snowflake more than 3,400 live listings from over 820 providers, Datarade more than 2,700 providers and Defined.ai 819 datasets. Kled's catalogue held 127 lots on offer on 27 October 2026, read daily on our index of data on offer.
• Backing. Human Native was bought by Cloudflare; Mozilla Data Collective raised $5 million from Mozilla.
• Nothing on record. FileYield publishes no figures, and Opendatabay publishes listings but no sales.

Before trusting a newer platform with an archive, ask it for the evidence in the first group: deals closed, money paid to holders, and a holder who will speak to it. Every firm in this guide has a page in our register of companies and our vendor register, with what is recorded about it; the guide to AI training data companies places the marketplaces among the rest of the market.

11. A marketplace, a buyer's programme or a broker

A marketplace is one of three ways to reach AI buyers, and it is the right one for some data and the wrong one for other data. The choice turns on three things: whether the data is a product many buyers want or a one-off archive, how much of the sale the holder can run itself, and whether it wants one buyer's price or several buyers bidding.

A listing site fits data that is a product: priced per unit, sold many times, refreshed, with the holder able to write its own licence, deliver the files and collect payment. A managed licensor fits creators and media libraries whose value appears only when pooled with others', and holders who would rather approve terms than negotiate them, at the price of an unpublished share and often exclusivity. A buyer's own programme, such as those of micro1, Mercor, Turing and Handshake, buys company records and code directly, at that buyer's price and on its terms; the comparison of buyer programmes reads each. A broker acts for the holder alone and puts the data before several buyers at once, so that their offers set the price; the page on going direct or through a broker sets out what each route gains and gives up.

AIData.Broker is a broker of the last kind, for holders based in the United States, the United Kingdom and Canada.

Questions people ask

What is an AI training data marketplace?

A platform that lists data held by companies or creators and licenses it to AI developers, keeping a share of each sale. Some are listing sites where the holder sets the price and delivers the data, such as Datarade and DataVendor (HUD); others negotiate and package on the holder's behalf, such as Protege and Troveo.

Which is the best marketplace to sell AI training data?

The one that already sells your kind of data. Video libraries go to Troveo, Protege and Wirestock, speech to Defined.ai and Protege, code to DataVendor, books to Created by Humans, and business data products to Datarade, AWS Data Exchange and Snowflake. Company records fit few marketplaces and sell mostly through buyers' own programmes or a broker.

How much do AI data marketplaces take?

From 5% of a sale above $250,000 at Opendatabay (its cut is 30% up to $100) and 8 percent at FileYield, which also charges the buyer 6 percent, to 50 percent of a dataset deal at Wirestock. Datarade keeps 30%, or 20% to 15% on a paid plan and DataVendor 20 percent at withdrawal. Protege, Troveo and Defined.ai do not publish their share.

Do I keep ownership of data I list on a marketplace?

At most of them, yes: Datarade says you keep ownership, Protege that you keep ownership and license limited rights, and Troveo that ownership never transfers. Kled is the exception in practice: it takes a transferable, worldwide, perpetual and irrevocable licence that is exclusive to Kled.

Can I list the same dataset on more than one marketplace?

Only if no licence you have signed is exclusive. Kled's contributors promise not to let anyone else use the content to train third-party AI models, and around 95 percent of Troveo's licensors signed exclusively. Exclusivity to one platform earns about 1.25 to 2 times the non-exclusive rate, so weigh it against every other channel it closes.

When does an AI data marketplace pay?

After a buyer pays, on the platform's schedule. DataVendor's earnings settle 14 days after payment, AWS Data Exchange pays monthly, only after the subscriber has paid, and Troveo's payout can come 4 to 5 months after delivery. At Datarade and FileYield payment runs directly between holder and buyer.

What happened to Human Native?

Human Native, a UK marketplace for licensed AI training data, was acquired by Cloudflare in January 2026, and Cloudflare tied it to its paid-crawling work for publishers. Cloudflare has since added Pay Per Use, where a buyer's offer is accepted, declined or countered by the publisher and payment follows use.

Is Hugging Face a marketplace for training data?

Not for selling it. Hugging Face is the largest hub for open models and datasets, with 1,100,828 public datasets, published openly rather than sold. Data placed there publicly is data a buyer no longer has to pay for, which is why DataVendor will not list public repositories.

What this means for you

A marketplace is a good channel for data that is already a product and a poor one for an archive that is one of a kind. Whichever you use, the terms that decide what you end up with are the share, the exclusivity and the payout rules, and those are on each platform's own pages or nowhere.

1. Know what your data is worth first. The price board and market rates hold every public price behind this guide, and the valuation puts a range on your data from them, so you can tell a fair listing price or a fair share from a poor one.
2. Pick by data kind, not by size. Use the table in section 9, then read each candidate's terms on its page in our register of companies.
3. Ask for the unpublished numbers in writing. Where the share, the payout timing or what counts as accepted is not published, get it in the agreement before you sign.
4. Bound any exclusivity. A term, a scope and a trigger that releases it; never a perpetual, irrevocable licence for a share the platform decides.
5. Settle the rights and remove personal data before listing. The rights check and the de-identification tool do both on your side.
6. Compare every channel. Buyer match finds the buyers whose published rules fit what you hold, the buyer board lists their terms, and the offer check reads any offer you receive against the market.

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. The holder pays us a fee fixed in the mandate and paid only if a sale closes, and we run sales only for holders based in the United States, the United Kingdom and Canada.

This guide is research from public sources, not legal, tax or investment advice.