How to Sell a Dataset: Where to Sell It and What It's Worth

You can sell a dataset when you hold the rights to everything in it (the people's consent, the licence of every source it was built from, the work of whoever labelled it), and you almost always sell it by licensing it, keeping ownership. There are four ways to do it: list it on a listing site and set the price yourself, hand it to a managed licensor that pools it with other holders' data and pays you a share, license it to a buyer that takes datasets of its kind outright, or have a broker put it before several buyers at once. What it fetches depends on its kind, its unit and which side of the market a number comes from: speech is asked at a median of $119 per hour of audio on listings but fetched $73 per hour of audio in the sales on record, and images list at $1.19 per image while the AI licences of images on record paid $0.014 per image.

This guide is for someone who holds a finished dataset: hours of recorded speech, an image or footage library, a text corpus, a set of training tasks with their graders, a collection scraped or licensed in from elsewhere, or an export packaged as a product, and who wants to be paid for it. It covers the rights you need, the licence you grant, each route and the data it fits, what Kaggle's and Hugging Face's terms allow, what each kind sells for by unit, what buyers check, how to describe and price a dataset, and a worked example. A company selling its own operating records (email, chat, tickets, CRM) is in a different market with different buyers, and the guide to selling company data covers it.

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

Contents

1. At a glance: four ways to sell a dataset
2. Can you sell a dataset? The rights you need first
3. The licence you grant
4. Where to sell a dataset: the routes compared
5. Kaggle, Hugging Face and the free dataset sites
6. What datasets sell for, by kind and unit
7. What a buyer checks before it pays
8. How to describe a dataset
9. How to price a dataset
10. Worked example: a conversational speech set
11. Questions people ask
12. What this means for you

1. At a glance: four ways to sell a dataset

The first decision is not the price but the route, because the route decides who sets the price, who does the work of the sale, who learns your name and when you are paid. On a listing site you are the shop: you write the listing, set the price, answer buyers and usually deliver the files yourself. A managed licensor takes your data, pools it with other holders' and licenses the pool, paying you a share. A buyer that licenses datasets of your kind pays you directly, at its own price and on its own terms. A broker acts for you alone and puts the dataset before several buyers at the same time, so that their offers, not one buyer's opening number, set the price.

What each listing site and managed licensor keeps, when it pays and who may list from where are set out platform by platform in the comparison of AI training data platforms. This guide compares the routes on what they do to a dataset's price and the holder's control of it.

2. Can you sell a dataset? The rights you need first

Yes, a dataset can be sold, but in this market "sold" nearly always means licensed: the buyer gets the right to use copies for stated purposes and the holder keeps ownership, which is why the same dataset can be licensed to several buyers. An outright sale is a transfer of ownership, and under US copyright law a transfer of copyright ownership is not valid unless it is in writing and signed. Either way, the buyer is paying for rights, and it can only receive the rights you hold.

Every platform puts those rights on the holder, in its own words. Datarade requires that your business 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; and Defined.ai asks every partner to show consent, documentation and the right to commercialise the data for AI training. A finished dataset carries four layers of rights, and each can stop a sale on its own.

What you made, and what others made for you

Material you recorded, photographed or wrote yourself is yours to license, and so is what your staff made in the course of their jobs: for a work made for hire the employer is the author and owns the copyright unless a signed writing says otherwise. The gap is usually in the work around the data: the transcripts, labels, captions and graders that turn raw material into a training set. When contractors made them, the work counts as made for hire only in listed categories (a work commissioned for use "as a compilation" is one) and only if the parties expressly agree in a written instrument signed by them. Without that, you need a written assignment from each contractor or labelling vendor before you can license what they produced.

How to apply it: list everyone who touched the dataset (recorders, annotators, transcription vendors, task writers) and find the clause by which each one's work became yours. A vendor contract that leaves rights in its output with the vendor is a gap a buyer's counsel will find.

The people in it: consent and releases

A dataset of voices, faces or people's words needs the consent of those people, and consent to be recorded is not the same as consent to have the recording used to train a model; the audio and speech guide walks through that gap state by state. Licensing personal information for money is a sale under California's privacy law, which counts making it available to a third party for monetary or other valuable consideration, and a voice recording from which a voiceprint can be extracted is biometric information there. Illinois goes further: no private business holding biometric identifiers may sell, lease, trade or otherwise profit from them. In Canada, personal information may not be used or disclosed for purposes other than those for which it was collected without consent.

Images of people need model and property releases, and an old release may not reach AI training: Getty Images, for one, does not warrant that model-released people consented to AI uses. Where the value does not depend on who the people are, removing identities is the usual answer; the de-identification standard, the de-identification tool, the HIPAA guide and the GDPR guide cover how.

The sources it was built from

A dataset compiled from other people's material inherits every licence and every set of terms it was taken under. A source published under a NonCommercial Creative Commons licence may be shared only for purposes not primarily intended for or directed towards commercial advantage or monetary compensation, so a set built from it cannot be sold without a separate licence from whoever holds the rights. The same goes for datasets downloaded from Kaggle or Hugging Face: each carries the licence its uploader chose, and a commercial sale needs that licence to allow one.

Platforms' own terms travel with what was taken from them. Reddit's User Agreement forbids collecting data from the site except under its terms or a separate agreement, adding that scraping the Services without Reddit's prior written consent is prohibited, and a federal court held that Anthropic was contractually bound by Reddit's User Agreement once it knew of it. Courts have drawn the line at how the data was reached: logged-out scraping of public pages did not breach Meta's terms, while X's claims about how Bright Data reached its servers were allowed to go forward. Copies taken from pirate sources are worse still: one court called them inherently, irredeemably infringing even when used for a transformative purpose.

Two more source rules catch dataset holders. Stock agencies' standard licences forbid using licensed images for machine learning unless separately authorised, so a set assembled from bought stock photos is not yours to license for training. And in the UK, copying works for computational analysis is allowed only for the sole purpose of research for a non-commercial purpose, so a corpus made under that exception was not made for sale. The guide to scraping law, the fair use guide and the guide to provenance go further.

Data a model made

Synthetic or AI-assisted data can be sold, but some buyers refuse it and the copyright position differs from human work. Shutterstock, which licenses its library to AI developers, will not allow AI-generated content to be submitted by contributors. The guide to selling AI-generated data sets out which buyers take it and on what terms.

3. The licence you grant

What you sell is a licence, and its terms move the price as much as the data does: what the buyer may do with the dataset (train its own models, evaluate them, pass the data on), for how long, in which field and whether anyone else may have it too. The guide to data licensing agreements sets out the clauses, the guide to licence term length what each term costs, and our licence terms page what each term is worth on the records.

Two choices move a dataset's price more than any other: whether the licence is exclusive, and whether it is for research or commercial use. Both are set before the first price is quoted, so decide them before you list.

Non-exclusive or exclusive

Most finished datasets are licensed non-exclusively and sold more than once. A listed Chinese speech data company, Haitian Ruisheng, sold 576 usage licences on 395 of its own off-the-shelf datasets in 2020, from a catalogue of 762, while its made-to-order data passed to the buyer that ordered it and could not be resold. Troveo advertises an average of 4 sales per dataset, a figure it labels illustrative. Buyers also tend to grant narrow licences onward: Defined.ai licenses to its buyers on terms that are non-exclusive, non-sublicensable and non-transferable, which keeps the holder free to license the same data again.

An exclusive licence fetches 2.0 to 5.8 times a non-exclusive one on the records, and two founders of training-environment companies put exclusive deals at roughly 4 to 5 times the non-exclusive price. But an exclusive offer has to beat every non-exclusive licence it blocks: an offer beats N non-exclusive licences only if it is more than N times one of them. And exclusivity to one platform is worth much less than exclusivity to one buyer, because it buys a single channel: platforms pay suppliers about 1.25 to 2 times the non-exclusive rate for it. The guide to exclusive and non-exclusive licences shows how to price each.

Research or commercial

Many dataset catalogues sell the same data under two licences: one for research only and one for commercial use. Across the pairs on record, a research or evaluation licence fetches 0.06 to 0.39 times a full commercial one. Stockfilm, which licenses a vintage home-movie archive for AI training, offers a research licence at a $25,000 flat fee for up to 25,000 clips, to accredited academic institutions against $3.00 a clip for commercial training, with a $10,000 minimum. A reasoning-code dataset sold from a Gumroad page asks €49 for personal or research use and €499 for a commercial licence.

Speech shows the gap at scale: research-only speech listings sit at a median of $15 per hour of audio, against $119 per hour of audio for the rest. The Linguistic Data Consortium sells many corpora on research-only terms, such as English Gigaword at $3,000 for a non-member. How to apply it: a research licence is a way to earn from a dataset that a commercial buyer has not yet taken, not a substitute for one. If you grant it, keep the commercial right, and say in the licence that models trained under it may not be used commercially.

4. Where to sell a dataset: the routes compared

Each route suits a different kind of dataset and a different amount of the holder's time. What differs most is who carries the work of the sale: on a listing site the holder writes, prices, answers and delivers; a managed licensor does most of that for a share it rarely publishes; a broker does the selling but leaves the licence and the delivery between holder and buyer.

The routes do not exclude each other unless you make them: a dataset can sit on a listing site while a broker approaches buyers, provided no licence you sign is exclusive.

Listing sites, and a shop of your own

Datarade, Opendatabay, FileYield, DataVendor, AWS Data Exchange and Snowflake Marketplace let the holder list a dataset and set its price. AWS Data Exchange alone holds more than 3,500 products from over 300 providers and Datarade counts 2,700+ data providers. On most of them the holder delivers: Datarade does not host the data, and at FileYield the data transfer, the sale contract and the payment happen directly between holder and buyer. Opendatabay also runs a brokered mode, in which the holder gets the agreed provider price and Opendatabay sets its own price to the buyer.

Listing sites suit data that behaves like a product: sold to many buyers, often refreshed. Every data product read for our records sells the same records to any buyer, mostly on subscriptions that keep the data fresh. Two cautions apply. The listed price is public, so it anchors every negotiation and shows rivals what you think the data is worth. And published prices fall steeply with volume: the same record costs 20 to 44 times less at volume across the products that publish a ladder, so copying an entry-tier price per unit overprices a large set badly.

A holder can also sell from its own page, as Stockfilm does with its two licences and the reasoning-code dataset above does from Gumroad. A shop of your own keeps every dollar but brings no buyers, so it works when buyers already know where to find you.

Managed licensors

Managed licensors take a holder's data, prepare it, find the buyer and pay the holder a share. Defined.ai licenses partners' data on commercial terms it agrees with the partner and states "$600K+ average partner revenue" (its own figure, with no basis given); to buyers it has sold speech in bundles such as 315 hours for €50,000. Troveo licenses on your behalf under terms you approve, and ownership never transfers; you run the exports, and it does the anonymisation and packaging. Wirestock pays creators per approved asset on projects, and in dataset deals by each creator's share of the pool. Shutterstock pays its contributors a 20% average royalty on what it receives for data licences, split by each contributor's content in the datasets sold, and invites outside libraries through a Data Partner Network ( record). Created by Humans does the same for books, and Kled for individuals' photos and video through an app.

Their value is pooling: a buyer rarely wants one holder's footage or one studio's speech, but it will pay for a set built from many. The cost is a share that is mostly unpublished and, at some, exclusivity: around 95 percent of Troveo's licensors signed exclusively. Ask for the share, the exclusivity terms and what counts as an accepted unit in writing before you sign.

Buyers that license datasets directly

Some buyers license finished datasets themselves, for their own models or to build products for labs. Sieve licenses existing video, audio and image libraries and pays after delivery is checked; Versos AI values one non-exclusive licence of the footage that passes buyer filters. The labs themselves describe what they want in the same terms: OpenAI's data partnership asked for datasets "not already easily accessible online to the public today". Of the 150 AI data vendors on our vendor register, 31 say on their own pages that they pay for data, and the buyer board lists 44 buyers, 33 of which publish a figure for what they pay.

Task sets are bought through a slightly different door, often as work to order. DataVendor's buyer requests are priced per task at around $500 to $1,500 a task, with hundreds to thousands of tasks per request, and Prime Intellect pays $100 to $500 for open bounties and $1,000 to $5,000+ for application-only tasks. Data a firm makes to sell is bought as a catalogue product or as paid labour, priced per unit and sold to many, which is why a holder of a finished task set should price it per task, not as a whole. The guide to RL environments and the page for holders of task sets cover who buys them.

A broker that puts it before several buyers

A broker acting for the holder earns its place through competition. A single buyer's first number gives you nothing to measure it against, and exclusivity, the largest lever on price, cannot be priced at all without knowing what non-exclusive licences would fetch. Prices for the same unit sit far apart: the middle half of public speech prices alone runs $65 to $266 per hour of audio, which is how far one buyer's view of a dataset can sit from another's. The page on going direct or through a broker sets out what each gains and gives up.

AIData.Broker is a broker of this kind. We value a dataset against the prices of its own kind and unit on the price board, not against what buyers of companies' operating records pay, and for a holder based in the United States, the United Kingdom and Canada we describe it, without the holder's name, to the buyers whose own published rules fit that form of data. The holder pays us a fee fixed in the mandate, and only if a sale closes; besides that, a buyer pays us only through its own published referral programme, and the holder sees the most it may pay before we introduce it ( pricing). We take a sale on when what it is likely to earn, from the buyers likely to license it, is more than running it costs; when it is not, the valuation says so and names buyers the holder can go to itself.

Which route fits which dataset

The platforms are more specialised than their home pages suggest, so the first question is which ones already sell your kind of data. The table maps common holdings to the routes that take them; the guide in the last column covers each kind in depth.

5. Kaggle, Hugging Face and the free dataset sites

The sites where most people first publish a dataset are built for sharing, not selling, and anything published openly on them is something a buyer no longer needs to pay for. The terms below were read on each site's own page on 31 October 2026.

The practical rule follows from the licences: publish openly only what you are willing to give away, such as a small sample or a documentation card, and keep the full dataset off the open web until it is licensed. Labs say they pay for what is not on the open web, and DataVendor will not list public repositories since a buyer can already clone them.

Kaggle

Kaggle's terms describe no way to charge other users for a dataset. They limit use of the site to "your own internal, personal, non-commercial use", and the payments they mention are fees Kaggle itself may charge for its paid services. Sharing a dataset publicly there grants a licence to everyone: you grant all other users of the Services a license to access that Public User Submission, and to use and exercise all rights in it, and every licence granted under the terms is royalty-free, perpetual, irrevocable and worldwide. You keep ownership, but a buyer who can download the set under that licence has no reason to pay for it.

The money on Kaggle is prize money. Competitions pay winners, and a host gets rights in an entry beyond the limited licences only if the winner accepts payment of the prize. Google DeepMind and Kaggle ran one evaluation contest with a $200,000 prize pool, paid only to the winning entries. A prize rewards a winner; it is not a price for a dataset.

Hugging Face

Hugging Face's terms say you own the Content you create, and that setting a repository public grants each user a perpetual, irrevocable, worldwide, royalty-free licence to use and redistribute it through the site. The Hub's gated datasets let an author require an access request, approve requests automatically or by hand, and ask for extra fields such as company and country; the gated datasets documentation describes access control, and no way to charge for it. The Hub lists 1,108,985 datasets, almost all free, and NVIDIA agreed to buy Hugging Face in September 2026.

Some holders use the Hub as a shop window instead. Datoric lists 10,000 hours of rights-cleared gameplay with synchronized inputs there with a commercial licence on request and no price, and Liva AI publishes a sample of a speech dataset of unscripted sales calls. A card and a sample tell buyers the data exists; the licence is negotiated off the site.

Zenodo and other open repositories

Open research repositories are for publishing, and their scale shows how much data is already free: Zenodo's own API reports 725,627 datasets. A dataset deposited there under an open licence can be used by anyone on that licence's terms. If your dataset was built from sets downloaded from these sites, the rule in section 2 applies: each source's own licence decides whether you may sell what you built.

6. What datasets sell for, by kind and unit

There is no single price for a dataset, but each kind has prices per unit, and they differ by the side of the market they come from. An asking price is what a holder lists data at; a sale is what a buyer actually paid a holder or a deal that closed; an offer to holders is what a buyer publishes that it pays; made to order is what a buyer pays to have data collected for it; and paid to makers is what the people who record or label data earn. Our price board records each of its 1,405 price observations with its side, and market rates give the middle of each kind, side and unit.

Asking prices are ceilings set before any buyer has looked; sales show what changed hands. The prices guide explains why the two part, and the table sets the medians side by side.

Read the last row as a contrast, not a comparable. Companies' operating records are bought whole, by buyers that price a company's history; a finished dataset is priced by its unit, by buyers of that kind of data. A dataset is valued at the market it sells in.

Speech

The cleanest completed sales of finished speech data come from a stock exchange filing. Haitian Ruisheng's prospectus reports that its off-the-shelf speech recognition datasets licensed at an average of CNY 413 an hour across 198,951 hours sold in 2020, and its speech synthesis datasets at CNY 13,832 an hour across 1,029 hours: studio voice for synthesis is a different product from ordinary recognition data. Catalogue listings run higher than sales, for example ELRA's commercial licence of 800 hours of American English mobile speech at €136,800, while real English customer calls list near what the valuation reads for two-way conversation, $24 to $26 per hour of speech: 1,000 hours with transcripts at $24,500 on Opendatabay, and a usage licence on Datarade at $25 an hour until that listing was taken down on 30 October 2026.

Two forces separate a strong speech set from a cheap one. Language and setting: catalogue speech recorded in a rarer language or place carries a premium. And whether it was made to order: a public buyer paid hundreds of times more per hour to have specialised speech made than the same data later listed for in a research catalogue. Made-to-order collection runs at a median of $124 per hour of audio, well above what archive speech fetches, which is why a holder that can record to a buyer's specification may earn more from that capacity than from its archive.

Images

Images show the widest gap between sides. Specialist sets list at dollars an image (the asking median is $1.19 per image), while the AI licences of images on record paid $0.014 per image an image, because most of those records are what agencies paid contributors per file from licences of whole libraries. Curation is what lifts a set from one end to the other: a curated image was quoted at 100x the price a bulk stock image fetched. The photos guide sets out what each agency pays.

Video

Video licensed for training has paid a median of $52 per hour of video in the sales on record, and more for premium material. Troveo reports $50M+ paid to creators for 7M+ hours of video licensed, and an industry benchmark put premium, metadata-rich archive content at $60 to $300 an hour (reported). Per-clip prices hide length: a price per clip can mean very different prices per hour, so convert every quote to hours before comparing.

Text and books

Books are priced per title, and the two sides measure different things. The sales on record price single works: a publisher's per-title fee such as HarperCollins' $5,000 per title for a three-year training licence, authors' shares of such fees, and court settlements and damages per work copied without permission. The listings price large collections of e-books or textbooks at one figure for the whole set, which comes to little per title. Text by volume is cheap: listings median $0.00097 per word, and per token $0.00000082 to $0.0000011 per token. Writing that was never published online is worth more than writing anyone can crawl; the valuation reads $850 to $2,400 per million words for it.

Tasks and environments

Training tasks are priced per task, with a grader, and the range is wide: practitioners told Epoch AI a task costs $200 to $2,000, the made-to-order median is $2,000 per task, and the people who write tasks are paid a median of $260 per task. A whole environment, a working replica of an application for agents to practise in, is valued at $2,300 to $340,000 for one environment for one. Describe a task set by its tasks, graders and pass rates, the units a buyer counts, not by its file size.

What moves a dataset's price

Within each kind, the same handful of things decides where a dataset lands. Not on the open web: what can be crawled or downloaded free fetches little, and data anyone could compile sells by the record near what re-making it costs. Uniqueness: buyers write it into their price formulas, so show that the data exists nowhere else. Licence scope: research against commercial, shared against exclusive. Volume: per-unit prices fall as sets grow. Acceptance: buyers pay for the units they accept, and Troveo pays on accepted hours, a subset of the hours it submits to a buyer.

7. What a buyer checks before it pays

A buyer's legal and research teams check the same things whatever the dataset, and two guides set them out in full. The due diligence guide lists the questions a buyer's counsel asks, area by area, with the evidence that answers each; the guide to proving quality without handing data over covers the description, counts, measurements and samples that come before any file moves.

For a finished dataset, the questions that end deals are: who can sign for it, and does the chain of rights run from every contributor to the holder; where each part came from and under which terms or licence; whether the people in it consented to AI training and how identities were removed; whether any earlier licence was exclusive, or the data is already public; and whether the data matches its description. Our rights check asks the 27 questions a buyer's counsel asks, 16 of which block a sale until answered.

Settled rights set a dataset apart. In the public catalogues we read daily, 29% of the brokered lots whose catalogue states who holds the rights say they are held with nothing left to confirm, and 14% of brokered lots are ready to deliver now ( index of data on offer). Buyers also price from the description first: the first price comes before any file moves at several of them, which makes the description the most valuable document a holder writes.

8. How to describe a dataset

A buyer decides from a description long before it sees data, so the description is a sales document and a legal one at once. The research standard is "Datasheets for Datasets", which asks for a dataset's motivation, composition, collection process, preprocessing, uses, distribution and maintenance ( arXiv). Buyers' own intake asks for the same things in shorter form: DataVendor wants a short inventory: asset type, origin, who owns it, your right to license it, documentation and obvious restrictions, and Troveo delivers structured, anonymized data with a manifest of exactly what is included.

Regulation now asks buyers for part of this too. The EU's template for summarising a general-purpose model's training content asks whether the provider concluded commercial licensing agreements with rightsholders, and for which modalities, so a buyer has to know what it licensed and in what form. The datasheet tool builds the page from your answers.

A datasheet that a buyer can approve says, at least:

• What it is: the kind of data, its unit and how many units, net of duplicates and rejects.
• Where it came from: who collected or made it, when, how and under which terms or licences.
• Who is in it: the people recorded or described, their consent and releases, and how identities were removed.
• How it is labelled: who annotated it, to what guidelines, with what agreement between annotators.
• What is excluded: the categories taken out and why.
• Licence offered: research or commercial, exclusive or not, term and field of use.
• Format and delivery: file formats, size, how and when it can be delivered.
• Known limits: gaps, biases, errors measured, and what it should not be used for.

9. How to price a dataset

Pricing a dataset is a matter of finding the unit buyers count, reading the right side of the market for it, and adjusting for the licence. The steps below follow the evidence in section 6; the guide to how buyers calculate an offer shows the same arithmetic from the buyer's side.

The common mistake is to start from the highest listing for a similar set and add a premium for quality. Listings are ceilings and the same unit costs far less in volume, so start from what sold.

1. Name the unit. Hours for speech and video, images for images, tasks for task sets, titles for books, words or tokens for text.
2. Read sales and offers before asks. Find the medians for your kind and unit on market rates and the individual records on the price board. Treat asking prices as the top of a range, not its middle.
3. Adjust for the licence. A research licence fetches 0.06 to 0.39 times a commercial one, and an exclusive licence 2.0 to 5.8 times a non-exclusive one; price the non-exclusive commercial licence first and each variant against it.
4. Adjust for volume. Per-unit prices fall as sets grow, so price a large set below the per-unit rate of a small one.
5. Adjust for what is unique. A language, setting, specialism or labelling depth that no public set has moves a dataset up; anything already free online moves it down.
6. Decide one sale or several. A non-exclusive set can be licensed again; compare one larger exclusive offer with the licences it would block.
7. Test the number. The valuation reads your answers against the market for your kind of data, and the offer check puts any written offer beside it. More than one buyer is the only test that settles it.

10. Worked example: a conversational speech set

This is an example, not a real holder. A small audio studio holds 80 hours of two-way conversation in US English, recorded over two years with paid participants for a voice project that ended. Each participant signed a release, and two freelance transcribers produced verbatim transcripts. The studio wants to know whether it can sell the set, where, and for roughly how much.

Rights first. The recordings are the studio's, but three gaps need closing before any listing. The releases must name AI training, not only the original project, because consent to record is not consent to train. The studio checks where participants live: California treats a voice recording from which a voiceprint can be extracted as biometric information, and Illinois bars profiting from biometric identifiers, so recordings of any participant there are left out. And the transcripts are the freelancers' work until a signed agreement makes them the studio's: without a work-for-hire clause it needs a signed written assignment from each.

The licence. Non-exclusive and commercial, because a set this size is worth more licensed several times than sold once, with a separate, cheaper research licence for academic buyers.

The price. The valuation reads two-way conversation in English at $24 to $26 per hour of speech, against listings of real English customer calls such as 1,000 hours with transcripts at $24,500. Its starting case, a larger set of the same kind, comes to $3,800 to $4,000 in all, so 80 hours is worth roughly half of that. The asking median for all speech, $119 per hour of audio, is the wrong comparison: it mixes catalogue corpora in many languages and settings, including the rarer ones that carry a premium.

The route. At that size, a round run by a broker would cost more than it is likely to earn, so the studio's best routes are a listing site, where it sets the price and keeps most of it, or a managed licensor such as Defined.ai, which pools speech from many holders and asks for exactly the consent documentation the studio now holds. If the studio could record to a buyer's specification instead, the picture changes: made-to-order speech pays a median of $124 per hour of audio, so its recording capacity may be worth more than its archive.

The same steps scale up: a far larger rare-language set with clean releases is a dataset whose price competition can move, and that is where a broker's round earns its fee.

Questions people ask

Can you sell datasets?

Yes, if you hold the rights to everything in them and to license them for the buyer's use. That means the consent of the people in the data, a licence from every source it was built from that allows commercial use, and the rights in any labels or transcripts contractors made. Platforms put the promise on you: Datarade requires the necessary rights, licences or consent to distribute the data.

Where can I sell my dataset?

On a listing site such as Datarade, Opendatabay or AWS Data Exchange, where you set the price; through a managed licensor such as Defined.ai, Protege or Troveo, which pools your data with others' for a share; directly to a buyer that licenses your kind of data; or through a broker that approaches several buyers at once. Which fits depends on the kind of data (section 4) and on how much of the selling you want to do yourself.

How do I sell datasets online?

Settle the rights, write a datasheet, choose the licence (research or commercial, exclusive or not), then list it or approach buyers. Keep the full dataset off the open web until it is licensed, and publish only a sample or a description. The datasheet tool, the rights check and the valuation cover the first three steps.

Can you sell a dataset on Kaggle?

Not on its own terms. Kaggle limits use of the site to "your own internal, personal, non-commercial use", and sharing a dataset publicly grants every user a royalty-free, perpetual and irrevocable licence to use it. The money on Kaggle is competition prize money, and a host gets rights in an entry only if the winner accepts the prize.

Can I sell AI training datasets?

Yes, and buyers say what they want: data that is not already easily accessible online, with consent and a clean chain of rights. Speech, images, video, books, task sets and robot data each have their own buyers and units (section 6). Data a model generated can be sold too, though some buyers refuse it; the guide to selling AI-generated data says which.

Can I sell a dataset built from Reddit posts?

In practice, only with Reddit's permission. Its User Agreement provides that scraping the Services without Reddit's prior written consent is prohibited, and a federal court held Anthropic contractually bound by that agreement. Reddit licenses its own data to AI companies instead, in deals such as Google's reported at about $60 million a year; the guide to Reddit's data deals has the detail.

Where can I buy and sell datasets?

The listing sites serve both sides: buyers search and buy on Datarade, Opendatabay, FileYield and AWS Data Exchange, and holders list on them. Buyers of AI training data also license through managed licensors and buy directly from holders; the buyer board lists 44 organisations that publish that they buy data for AI training.

How much is my dataset worth?

It depends on its kind and unit, and on whether you read asking prices or sales: speech sold at a median of $73 per hour of audio against $119 per hour of audio asked, and video at $52 per hour of video against $72 per hour of video. The licence moves it too, with a research licence at 0.06 to 0.39 times a commercial one. The valuation puts a range on your dataset from the records of its own kind.

What this means for you

A dataset is worth what its rights let a buyer do with it, priced per unit against the sales of its own kind. Most of the work that decides the price happens before any listing: the rights, the licence and the description.

1. Close the rights gaps first. Releases that name AI training, written assignments from every contractor, and a licence that allows commercial use from every source. The rights check asks the questions in order.
2. Keep the full set off the open web. Publish a sample or a card, not the dataset, on Kaggle, Hugging Face or an open repository.
3. Choose the licence before the price. Non-exclusive and commercial first; research and exclusive licences priced against it.
4. Price from sales, per unit. Use market rates and the price board, and read the valuation for a range from your kind of data.
5. Write the datasheet. The datasheet tool builds it; it answers a buyer's first questions before they become discounts.
6. Pick the route by kind and size. A listing site for a product sold many times, a managed licensor for data worth more pooled, and competition among several buyers where the dataset is rare or large. Buyer match finds the buyers whose rules fit, and the offer check reads any offer 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 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.