AI Training Data Prices in 2026: What Data Sells For

What nine kinds of AI training data actually sell for in 2026, how each price gets set, and how to estimate your own

In August 2026 Google won the entire corporate archive of bankrupt Spirit Airlines with a $10 million bid, against two other qualified bidders - Chapter 11 docket summary. A year earlier, a YouTube creator who licensed a thousand hours of old videos earned $33,000, a little over $33 an hour of footage, or about 55 cents a minute - Semafor. Both are sales of training data to AI companies. One prices a whole company's history at eight figures; the other prices an hour of video below what a parking space costs in most cities.

That gap is the problem anyone holding data runs into. There is no price list. A marketplace that sells training data says so in plain words: "There is no price list for AI training data" - Troveo. Buyers rarely publish what they pay, holders sign confidentiality clauses, and the few figures that do surface are scattered across court dockets, earnings calls, job boards and vendor price guides. A company founder asking "what is our Slack worth?" or a photographer asking "what is my archive worth?" usually gets either silence or a number from someone who wants to buy it.

This guide collects the public figures instead. It covers the nine kinds of data that sell for AI training today: company records, source code, video, audio and speech, images, text and archives, datasets and RL environments, physical-world data for robots, and expert human work. For each one you get a table of public prices with the date and the source, the direction demand is moving and the evidence behind that call, and what raises or lowers a price. Around those sections sit the mechanics: how prices actually get negotiated, why quoted prices run far above realised ones, where each dollar goes once a lab pays, the largest disclosed deals, and how to estimate a range for your own data without fooling yourself.

A word on who is writing. AIData.Broker publishes the price board, the deal ledger and the buyer board that this guide draws on, and it also acts as the agent of holders who license or sell their data. That makes us a party in this market, not a neutral observer, so every figure below links to where it was originally published. Check any number you plan to rely on.

This guide is research from public sources, not legal, tax or investment advice: take your own advice before you sign anything.

Contents

1. The price map: nine kinds of data at a glance
2. Why there is no price list
3. How prices actually get set
4. Company records
5. Source code
6. Video and footage
7. Audio and speech
8. Images
9. Text and archives
10. Datasets, tasks and RL environments
11. Physical-world data for robots
12. Expert and human work
13. Quoted prices versus realised prices
14. Follow the dollar: unit economics
15. The biggest disclosed deals
16. How to estimate your own data's range, and its limits
17. Where prices are heading: a two-tier market
18. FAQ
19. Conclusion: a decision framework

1. The price map: nine kinds of data at a glance

Before going kind by kind, it helps to see the whole spread on one page. The table below gives the range of public prices for each kind of data, the unit the market prices it in, and the direction demand moved through 2026. The ranges are deliberately wide, because they are: the same kind of data sells for ten or a hundred times more when it is linked, scarce, cleanly licensed or sold exclusively. Each range is built from the dated, sourced rows in the sections that follow, and every row also lives on the public price board.

Two things stand out. First, the unit changes from kind to kind: company records sell per company, code per repository or pull request, video per minute, speech and robot data per hour, images per image, books per title, tasks per task. A single "price per terabyte" does not exist anywhere in this market, and anyone quoting one is converting from something else. Second, demand is not moving in one direction. Some kinds are getting more expensive as AI labs shift from reading the internet to practising real work, while others, the ones that were first to get a price list, are already cheapening.

Read the table as a map of anchors, not a menu. A number in it tells you what someone, somewhere, paid or offered for a particular thing at a particular date. It does not tell you what your data will fetch, because your data is not their data. What the table does tell you is the order of magnitude you are dealing with and the unit a buyer will want to talk in, which is the first thing to get right in any conversation with one.

For the kinds priced by the hour, it is possible to put them on one axis. The chart below shows the low and high ends of the public ranges per hour of data, from raw first-person video collected in India to custom-recorded speech. The point it makes is that the format of the data does not set the price: video appears both near the bottom (a creator's realised rate) and near the top (licensed film), and speech spans from a catalogue price to a custom commission at nearly three times that.

Sources: raw egocentric, creator, realised, egocentric to spec, speech, catalogue, teleoperation, creator, quoted, film, marketplace, speech, custom.

The rest of this guide explains why those bars sit where they do. The short version: price follows what a buyer cannot get any other way, multiplied by how sure the buyer can be that it is allowed to use it.

2. Why there is no price list

Ordinary commodities get price lists because the goods are interchangeable and there are many buyers and sellers. A barrel of oil is a barrel of oil. Training data has none of those properties. Each dataset is different from every other one, there are only a handful of serious buyers (the frontier labs, a few dozen smaller labs, and the vendors that supply them), and the value of a dataset to a buyer depends on what that buyer's model already knows. A corpus that teaches one lab's model something new may teach another lab's model nothing, because the second lab already bought something similar last quarter.

Start from what a lab is actually buying. It is not buying bytes. It is buying an improvement in a model, on some task it cares about, at a moment when that improvement is worth something. That has three consequences for price. It is buyer-specific: the same archive is worth more to the lab that is weakest at the skill it teaches. It decays: once a model has learned a capability, more data teaching it is worth little, which is why one industry analysis calls an RL environment "a one-time delivery" that loses half its value within months - Next Signal Prediction. And it is capped by substitutes: if a lab can generate the data with its own models, or hire contractors to recreate it, it will not pay much more than that would cost.

On top of that, almost nothing is disclosed. Of the 111 licensing deals with AI companies in the AIData.Broker deal ledger (July 2023 to October 2026, each linked to its original publication), only 19 carry a dollar total or an annual amount. The rest say "terms not disclosed", or give a figure in words such as "tens of millions of euros". And 40 of the 111 are not even for training: they license content to be shown or cited inside an AI product's answers, which is a different product with a different price. The chart below shows how few disclosed deals carry money figures, half-year by half-year.

Sources: every disclosed deal and those with a dollar amount, counted by the half-year of their date.

The spikes are not demand spikes. The first half of 2024 is when stock libraries, Reddit and the big news publishers signed their first AI deals; the second half of 2025 is a run of announcements by Meta with news publishers and by the major music labels with AI music companies, almost none with a price attached. The trend line that matters is the lower one, and it never rises above six in a half-year. Most of this market trades in the dark.

What replaces a price list, then, is a set of anchors. A holder who knows the anchors for their kind of data, and knows which way their own data deviates from them, can negotiate. A holder who does not will accept the first number offered, which in this market is usually the bottom of someone's published range. The next section explains how the people on the buying side actually turn anchors into a price.

3. How prices actually get set

Practitioners who sell to AI labs describe a consistent stack of arguments, from the strongest to the weakest. Knowing the stack matters for both sides, because the price lands wherever the strongest argument that either party can support puts it. A holder who can show measured improvement in a model is negotiating on value; a holder who can only show their costs is negotiating on a floor.

Here is the stack, strongest first, with the evidence that each one is real and not just theory.

Measured lift is the strongest argument because it is the only one tied to what the buyer is actually paying for. The big vendors now run a training job on an open model with the data before pitching a lab, then show the before-and-after scores. Labs do the same thing on their side after delivery. One practitioner essay describes it bluntly: purchase orders are tied to deliverables and "Nothing renews by default", and a single lab contact works with seven vendors that "are tasked with producing the same type of datasets" - Thoroughly Intrigued. A lab that measures each purchase against held-out tests will pay again for what moved the scores and drop what did not.

Compute parity sounds abstract but is the reason RL task prices sit where they do. If a lab is going to spend roughly $2,400 of compute training on a task over its life, a task bought for $50 that turns out to be broken wastes far more money than it saved. That logic pushes task prices up toward the hundreds of dollars and beyond, and it is why the cheapest tasks are not the best deal for a buyer.

Comparables are what most people mean when they ask about price, and they are what the tables in this guide provide. They are useful for one thing above all: they tell you the order of magnitude and the unit. They are weak on their own, because the comparable dataset may have been linked, exclusive or unusually clean, and yours may not be.

Cost plus is the floor, and for anything made by people (expert tasks, recordings, commissioned speech) it is where pricing starts. Mercor's chief executive put the arithmetic in one sentence: a task that takes ten hours of a person paid " $150 an hour" carries " a $1,500 cost basis", before margin - Sequoia. For data that already exists, such as an archive, cost plus barely applies: the cost of producing it was paid years ago for other reasons, which is exactly why archive holders should never price on what the data cost them.

The multipliers that move a price most

Once an anchor sets the base, a few multipliers move it a long way. They matter more than size, and a holder who understands them can often double or halve a price without changing a single file. The evidence for each is uneven: one has a published number, the others are documented by what buyers say they pay for and what they walked away from.

The largest single lever is exclusivity. Two founders of environment-building companies independently told Epoch AI that exclusive deals are " roughly 4-5x more expensive than non-exclusive ones" - Epoch AI. That is the only published figure for the premium, and it comes from environments, not archives, but the logic carries: a buyer pays several times more to be the only one with something. It also cuts the other way. A holder who licenses non-exclusively can sell the same data to several buyers, and three non-exclusive sales can be worth more than one exclusive one. Whether to sell once at the higher price or several times at the lower one is the main strategic decision in any sale.

The next lever is control of de-identification, and the Spirit Airlines auction priced it precisely. Mercor bid $7.0 million and then $10.0 million for the archive, "each conditioned on completing deidentification with its own tools rather than through a third party", and neither was selected; Google won at $10 million with a third-party scrub it paid for - Chapter 11 docket summary. The estate's own filing said such non-price factors "may be outcome-determinative". In plain terms: at the same headline price, a buyer who agreed to an independent scrub beat a buyer who wanted to do it in-house.

The remaining levers are linkage, scarcity and rights. Linked records (a support ticket tied to the code change that fixed it and the discussion around it) let a buyer rebuild a whole piece of work, and they price well above loose files. Scarce data (an uncommon language, an unusual trade, a niche programming language) cannot be recreated by hiring contractors, so it holds value. Clean rights decide whether a buyer's legal team will sign at all. Here is what moves each kind in practice:

• Exclusivity: roughly 4 to 5 times the non-exclusive price, decaying as similar data reaches rivals
• De-identification control: an independent scrub the buyer pays for won the Spirit auction
• Linkage: work traceable from request to result beats loose files of the same size
• Scarcity: what a vendor cannot recreate by hiring people or generating it
• Rights and consent: a complete rights file is what gets a deal through legal review

Notice what is missing from that list: volume. Raw size is the most overrated input in this market. A wind-down service that analysed codebases found that " roughly 37% of raw code disappeared" once third-party code, generated code and internal duplication were removed - SimpleClosure. Buyers strip that out before they value anything, so terabytes are a poor proxy for price. In practice, before talking to any buyer, a holder should answer four questions: what is my anchor, can I show lift, am I selling exclusive or not, and who will do the de-identification. The kind-by-kind sections below give the anchors.

4. Company records

Company records are the working history of a business: its procedures and runbooks, its tickets and project histories, its documents, its email and chat. Two years ago almost nobody bought them. In 2026 they are the fastest-moving category in the market, and the reason is structural. AI labs no longer get much from reading more of the internet; to make models that can do jobs, they train them by having them practise tasks inside simulated workplaces, and a simulated workplace is only as realistic as the records that fill it. That is why the biggest buyers of company archives are the vendors that build those training environments, plus a few labs buying direct.

The price range is the widest of any kind of data, from tens of thousands of dollars for a closed startup's archive to eight figures for a whole airline. Several buyers now publish tiers or ranges on their own sites, which makes this one of the better-documented categories, although published ranges are ceilings and invitations more than prices (see section 13). The table below gathers the public figures.

Demand: rising. The clearest signal is money committed in advance. On 9 October 2026, micro1 said it would " spend $1 billion over the next 12 months acquiring and licensing enterprise operational data" - micro1. It also pays a finder's fee of $50,000 per referral for introducing a company - micro1 referral page. Epoch AI's interviews with environment builders and lab staff describe enterprise workflows (expense reports, spreadsheets, CRM updates, ERP) as a major growth area of environment demand - Epoch AI. And the Spirit auction drew three qualified bids from AI companies for the records of an airline that had stopped flying. You can follow the main buyers on their own pages: micro1 and Mercor.

Why do the published ranges span two orders of magnitude? Because size alone barely matters. A closed ten-person startup with two years of loose files sits at the bottom; a specialised operating company with a decade of records that link across systems sits at the top. Handshake AI's own estimator shows $383,083 to $896,969 for a US company of 100 people with ten years of operation, and labels that indicative rather than an offer - Handshake AI. The SimpleClosure figures run the other way: nearly 100 deals in a year for more than $1 million in total, so most sat near the bottom of its $10K to $100K range - Forbes.

What raises a company-records price, in the order the evidence supports it:

• Linked work: a ticket tied to the change that closed it and the discussion around it
• A specialised trade: insurance, logistics, manufacturing, law, finance
• Long continuous history: many years, not months
• Records written by the company about its own work, not forwarded third-party material

What lowers it is the mirror image: loose mailboxes with nothing tying them together, a small team with a short history, customer or employee personal data that cannot be cleared, and software terms that forbid export for AI use. The last one is easy to miss. Slack's API terms, for example, were changed in 2025 to restrict bulk export and use in language models, so an archive pulled through the wrong route can be unsellable however rich it is.

Rights are the gating question here, more than anywhere else. Owning the files is not the same as having the right to license them. Customer contracts, software licences, employee notices and third-party material inside the archive all need checking before a buyer sees a sample. Spirit's sale shows what happens when they are not: unions and software vendors objected that their confidential material sat inside the archive, Google agreed to drop the passenger transaction, refund, inflight purchase and WiFi sales systems from the sale at the same price, and the sale still awaited court approval as of this writing. For a company considering a sale, the practical sequence is: build the rights file first, decide whether the company is selling once or licensing to several buyers, and get the same datasheet in front of every buyer the holder approves at the same time so that offers compete. That sequence is how AIData.Broker runs a sale when it acts for a holder, and the guide to selling a company's data explains the process.

5. Source code

Source code is the most commoditised form of company data, and the prices show it. Public code is already in every model, so a private repository is worth something only to the extent that it teaches what public code does not: a scarce language, an unusual domain, or, most valuable of all, the history of how changes were requested, argued over and shipped. Code on its own, without the tickets and reviews around it, is the cheapest company record there is.

Several buyers publish prices for single repositories, and they cluster in a narrow band of a few thousand to twenty thousand dollars. The exceptions are legacy and scarce codebases, which one buyer prices an order of magnitude higher, and full operating histories, which are priced as company records (section 4).

Demand: split. Ordinary web and app code has become a commodity, with 13 buyers, marketplaces and brokers on the AIData.Broker buyer board now taking repositories, so the price for a typical repository is set by competition among buyers who all want roughly the same thing and can get plenty of it. Scarce code is different. Turing prices the same buyer's legacy codebases (COBOL, .NET, Java) at $10K to $100K+ while capping modern ones at $10K - Turing. The gap between those two figures, on the same page from the same buyer, is the clearest evidence in this guide that scarcity rather than size sets the price.

The last row of the table matters more than it looks. Together AI generated a dataset of 258,000 coding-agent trajectories across 1,655 repositories and reported that " the total cost of generating this large-scale, long-context trajectory dataset was $130k" - Together AI. That works out to about 50 cents a trajectory (our arithmetic). It is a production cost, not a sale price, and the dataset was released free, but it marks the ceiling on anything a model can produce for itself: if synthetic data can stand in for yours, yours is worth about what the synthetic version costs.

What raises a code price follows directly. Reviews and tickets linked to the commits add the most, because they turn a snapshot into a record of how work was done. A scarce language or domain (embedded, scientific, industrial, finance, or the legacy languages above) adds a premium. Years of real production history help, and so does code a buyer can actually build and run. What lowers it: common web or app code with no history; generated code, duplicates and vendored libraries, which buyers strip first (the 37% figure in section 3); and open-source licences or client ownership over parts of it.

Rights trip up code holders more than most. Code written for a client often belongs to the client, depending on the contract. Open-source components keep their own licences. A contractor's work belongs to the company only if the contract says so. For a single repository with none of those problems, our view is that a platform that grades and lists repositories (the ones in the table) is usually the fastest route, and AIData.Broker says so to people who bring one to us. Codebases that come with their tickets and history, a scarce language, or a whole company's records are where competing offers through an agent make a difference. The source code page keeps the current anchors.

6. Video and footage

Video is where the gap between a quoted price and a realised one is widest, and where many creators first met this market. Video models and world models learn how things look and move from footage, and the large public platforms forbid scraping theirs, so footage with clear rights is what remains to buy. In late 2024 and early 2025 AI companies began paying creators for unpublished footage, and the quoted rates got a lot of attention.

The table shows the public figures. Note how different the units are: per minute for creator footage, per clip for short videos, per hour for longer films, and a single multi-year licence for a whole marketplace's library.

Two caveats on that table. The Calliope figure comes from a small trade blog that does not say whether it is the buyer's price or the share reaching rights holders. The Kled deal was announced by the company itself, with the buyer unnamed and terms unverified. Both are included because they are what is public, and both should be weighed lightly.

Demand: split. Generic footage is plentiful. One creator marketplace alone has licensed more than 7 million hours of video with clean rights, and buyers can choose. That is why the one creator who published his realised earnings, $33,000 for a thousand hours, came in at about 55 cents a minute, a fraction of the $1 to $4 quoted range - Semafor. A talent-agency executive quoted in the original Bloomberg Law report expected the window to close: "I don't think that window is going to last that long" - Bloomberg Law. At the other end, footage that carries something a model cannot get elsewhere, such as gameplay recorded with the player's inputs, first-person video of real tasks, or high-spec cinematic work, is priced above it, and the main creator marketplace says scarce categories such as gameplay and first-person video price above generic footage - Troveo.

What raises a video price: high resolution, drone or cinematic work (Bloomberg Law reported a premium for 4K, drone and 3D animation); gameplay or screen capture recorded with the inputs; signed releases from everyone who appears; and clean metadata on what, where and when. What lowers it: people in frame who never agreed to appear; music, logos or other people's work in the shot; and footage already published where anyone could scrape it, since a buyer is partly paying for the fact that rivals do not have it.

The rights test for video is about people as well as ownership. You must own the footage, and anyone identifiable in it must have agreed to this use. Licensed music in the edit stays the music rights holder's. Creator contracts with AI buyers have also included use restrictions (no digital replicas of the creator, no recreating their scenes), which are terms worth reading closely because they protect the holder. In practice, small libraries are usually best listed directly on a licensing marketplace that pays per minute; large or specialist libraries (thousands of hours, or a scarce category) are where competing offers make a difference. The video page tracks both.

7. Audio and speech

Speech is one of the few kinds of data with something close to a catalogue market. Vendors sell recorded speech off the shelf by the hour, and they publish price guides. Underneath that catalogue sits a custom market, where a buyer specifies the languages, speakers and scenarios and pays to have new speech recorded with consent on file. The two price bands barely overlap.

Why do labs pay for speech at all when so much audio exists? Because most of it is the wrong kind. Voice models were trained heavily on people reading aloud, which is not how anyone talks. Real two-way conversation, with interruptions, hesitation, accents and languages the large datasets miss, is what they lack, and it is hard to collect lawfully because every speaker has to agree.

The last row is not training data in the strict sense: it is royalties for licensing a voice for use, not for teaching a model. It is included because it shows what voice creators collectively earn when a platform pays per use, about $2,100 per creator on average by our arithmetic, and because holders often ask about it.

Demand: rising, for the right kind of speech. The best single data point is the Odia figure: an hour of speech in that language " used to cost about $3-$4, now it's $40", a tenfold rise as large tech companies sought Indian-language data - Context. That is the scarcity premium at work. Custom collection at $150 to $400+ an hour points the same way. Plain read speech in English moves the other way: there is a lot of it, and catalogue prices for it are soft. One more signal: frontier labs are building their own consented supply, for example by asking their own users to opt in to sharing voice conversations, which caps what an outside holder of ordinary speech can charge.

What raises a speech price: two-way conversation instead of read speech; a less common language or a strong regional accent; accurate transcripts and speaker labels; and written consent from every speaker. What lowers it: any speaker who never agreed to this use (which makes the recording unsellable, not merely cheaper); poor audio or heavy background music; and English read speech, which is plentiful.

Consent is not a detail for audio. A voice is treated as biometric data in some jurisdictions, and class actions filed in 2026 accuse large companies of extracting voiceprints without consent. A recording made "for quality and training purposes" in a call centre does not cover selling it to an AI lab, and the other party on a call is the classic trap. For an archive of consented interviews or podcasts, the audio and speech page shows the current anchors; for speech that does not exist yet, the economics are those of commissioned work (section 12), priced by the accepted hour.

8. Images

Images were the first kind of training data to get a visible price list, and they are now the first to cheapen. In 2024 the per-image prices were reported in detail: a couple of cents in bulk, a few cents to a dollar from a photo host in talks, a dollar or two for curated images with metadata. Large stock libraries signed multi-year licences with big tech buyers worth tens of millions of dollars each. By 2026 that business has stalled.

The structural reason is substitution. Image models already exist and are very good, so the marginal value of yet more ordinary photographs is low, and generating synthetic images is cheap. What still sells is what models do badly: specialist subjects, images with rich metadata, and libraries with releases that a buyer can defend in court.

The same Defined.ai price list put sensitive images at $5 to $7 each, several times its ordinary rate, because they need special handling. That is a useful reminder that "images" is not one market: the handling cost and the scarcity of a particular subject can multiply the price of an otherwise ordinary picture.

Demand: falling, at least for general libraries. The evidence comes from the companies that sell the most. Shutterstock reported that revenue from its Data, Distribution, and Services line " decreased by $7.1 million, or 19%" year on year in the fourth quarter of 2025, though the same line grew 16% for the full year - Shutterstock release, SEC. On Getty's earnings call, management said the two large deals "are not the pure data licensing deals", put traditional data licensing at "low single-digit percentages of total revenue", and said there was "nothing meaningfully baked in" for new deals in 2026 - Getty Q4 2025 earnings call. Getty's best-known 2026 deal, with OpenAI, is for showing its images inside ChatGPT, not for training - Getty Images newsroom. The buyer's page for Shutterstock tracks its licences.

What raises an image price: captions, keywords and metadata on every image; releases for people and property; and a specialist subject that stock libraries lack (medical, industrial, scientific, a particular region or period). What lowers it: everyday snapshots, which are abundant; recognisable people without releases; and images already on the open web. Bulk libraries of ordinary photographs are best listed directly on the image marketplaces that pay per image, which is what AIData.Broker tells people who bring one. Specialist libraries, where several buyers may compete, are a different matter, and the images page keeps the anchors.

9. Text and archives

Text was the raw material of the first generation of large language models, almost all of it scraped. What changed is the legal risk. Labs have been sued over books and articles they did not license, and one settled. The class settlement in Bartz v. Anthropic pays about $3,000 per work for pirated books used in training, from a $1.5 billion fund - Authors Guild. That figure is not a licence price, it is the cost of not having licensed, and it now sits in the back of every negotiation over edited, published writing.

Text prices therefore split in two. Bulk general text is cheap and priced by the word. Edited, published, expert writing from a clear rights holder is priced by the title or by an annual licence, and the large news and platform deals sit at the top of the whole market by value.

The gap between the first two rows is the whole story of text. HarperCollins' $5,000 for a book, at roughly a hundred thousand words (our assumption of a typical book length), works out to about fifty times the $1,000 per million words a marketplace quoted for general text (our arithmetic). Curation, editing, a clean rights chain and the fact that the work is not freely online all sit inside that multiple.

Demand: steady, concentrated at the top. Large publishers' licences are holding their value. Wiley said it " Delivered $49 million of AI revenue (+23%)" in fiscal 2026, with lifetime AI revenue past $110 million - Wiley results, SEC. Reddit's other revenue, which is mostly data licensing, grew 24% to about $43 million in the second quarter of 2026, still only about 5% of its revenue - Reddit results, SEC. And Meta's reported $50 million a year for News Corp in 2026 roughly matches the annual value of News Corp's 2024 deal with OpenAI, which suggests top-tier news prices held flat rather than collapsed. You can follow these licensors and buyers on Reddit's and OpenAI's pages.

Below the top tier, the picture is much thinner. Smaller and non-English publishers are mostly offered revenue shares or per-use payments rather than cash up front. The market is also moving from archive fees to pay-per-use: Microsoft's Publisher Content Marketplace, launched in February 2026, pays publishers per use with publisher-set pricing - WinBuzzer. Pay-per-use suits buyers who want fresh content for answers, and it makes a one-off archive payment rarer.

What raises a text price: edited, expert or professional writing; rights held cleanly by one party; a subject or language models are weak in; and a flow of new material rather than only an archive. What lowers it: text already freely available online; rights split between authors, agents and publishers; and user posts whose authors never agreed to licensing. A publisher often holds print rights without AI training rights, and authors may need to agree title by title, as they did in the HarperCollins deal. The text and archives page has the current anchors.

10. Datasets, tasks and RL environments

This is the category that most changed the market. Labs now train agents with reinforcement learning on verifiable tasks: the model attempts a task thousands of times inside a sandbox, and a grader scores each attempt. The thing being bought is the sandbox (an "environment", often a working copy of a piece of software filled with realistic records), the tasks inside it, and a grader for each task. Epoch AI's diagram below shows the loop for a simulated Slack workspace: a task goes in, the agent acts, the environment responds, and a grader turns the final state into a reward.

Every part of that diagram is something a vendor sells. The task is written by a domain expert. The environment is engineered, often by cloning real software. The grader has to be impossible to fool, because a model that finds a shortcut learns the shortcut instead of the skill. That is why tasks are priced in hundreds or thousands of dollars each, and why the best ones are written by people with real expertise in the field.

A few rows need context. The Epoch figures come from interviews with 18 people who build and buy environments, so they are reported practice rather than a price list, but they are the most cited figures in the field. The Outlier figure is what the contributor is paid, not what the lab pays. The benchmark rows (Humanity's Last Exam, FrontierMath, SWE-Lancer) are evaluation sets rather than training sets, and they show how much a single expert-written, verified problem can be worth when the point is that frontier models cannot solve it.

Demand: rising, but it is a flow business. The spending signals are large. TechCrunch reported in September 2025 that " leaders at Anthropic have discussed spending more than $1 billion on RL environments over the next year" - TechCrunch. Mercor's chief executive said " certain frontier labs might buy 50,000 tasks a month from us" - Sequoia. Epoch reports contract sizes " often six to seven figures per quarter". At the same time, a task is worth little once models pass it, environments lose half their value within months, and narrow website clones are commoditising as dozens of vendors build them. Selling here means producing continuously, not selling an inventory once.

Notice what the website-replica price implies for anyone who owns an interface. Where a lab can have a $20,000 clone of a website built, it does not need to pay the website's publisher anything. The cost of a replica is the ceiling on what interface data is worth. What cannot be cloned cheaply is the realistic content that fills an environment (the company records of section 4) and the expert judgment that writes tasks and graders (section 12).

What raises a price here: graders that cannot be fooled; a professional domain where the builder has real expertise; difficulty calibrated so models pass sometimes but not always; and exclusivity, which buyers pay four to five times more for. What lowers it: tasks a model already passes, graders that reward shortcuts, and replicas of common apps. You must also own everything inside the environment: tasks derived from someone else's proprietary material carry that material's restrictions, as a 2026 appeals ruling over legal memos written from a competitor's database made clear. The datasets and environments page tracks current prices.

11. Physical-world data for robots

Robots cannot learn to fold a towel or turn a valve from the internet. They need to see hands doing real tasks from the worker's point of view (first-person, or "egocentric", video), and they need to be driven through tasks by a person (teleoperation) or to copy a person's movements (motion capture). In 2026 this became a large business, collected mostly by gig workers wearing head cameras in India, Africa and elsewhere, and its prices fell quickly as supply grew.

The prices form a ladder from raw footage to teleoperation on real robots, and the steps of the ladder show exactly what buyers pay for: specification, calibration, annotation, and above all hardware.

The collectors' side of the same market shows where the margin sits. Human Archive pays workers in India " a base rate of $1 per hour" for egocentric collection - TechCrunch. A worker in Nigeria filming household chores for micro1 is " paid $15 an hour" - MIT Technology Review. An analyst quoted by CNN put the hourly wages for this kind of collection at " anywhere between $5 and $20 depending on the region" - CNN, via KVIA. In the US, a robot operator doing VR teleoperation for Toyota Research Institute was advertised at $32.50 to $37.50 an hour - HireArt listing, and Tesla's motion-capture operators were offered $25.25 to $48 an hour in 2024 - Fortune.

Demand: split. Spending is real: micro1's chief executive estimated robotics companies are " spending more than $100 million each year to buy real-world data" - MIT Technology Review. But supply grew faster. " Data collection rates already fell 30% between January and June 2026 as suppliers multiplied, an industry executive said" - StartupFeed. Bulk footage of everyday chores is now close to worthless at the margin, while teleoperation on real hardware, skilled trades and unusual settings hold their price, because they cannot be produced by strapping a phone to more foreheads.

What raises a physical-data price: skilled trades and unusual settings (a metal shop, a lab, a ship's engine room rather than a kitchen); calibrated cameras, depth and hand tracking; annotation of every action; and teleoperation on real robots. What lowers it: undifferentiated bulk footage of chores; workers and bystanders who never consented; and footage a buyer will discard at review. Buyers pay for accepted hours, and DexSet's guide says buyers reject 15% to 25% at quality review - DexSet. Every worker filmed must consent, and so must anyone else in view; footage on a client's premises needs the client's agreement. This market works best to a buyer's specification rather than in bulk, and the physical-world page explains how AIData.Broker handles requests here.

12. Expert and human work

Most of the money AI labs spend on data does not buy data that already exists. It buys people's time: experts writing tasks and ideal answers, grading model outputs, building financial models, recording themselves doing real work on screen with narration. No archive contains an expert's judgment about a new problem, so labs commission it, mostly through a small number of large vendors. Those vendors are now some of the fastest-growing companies in the industry.

Prices here are posted openly, on job boards and vendor sites, because vendors compete for experts. What is public is the rate paid to the person, not the rate the lab pays the vendor. Section 14 explains the gap.

Demand: rising for specialists, falling for generalists. The vendor revenue figures say the money is growing fast: Mercor hit " over $2 billion in gross annualized revenue in June 2026" - Dealroom, and its homepage showed an average contracted rate of $112 an hour when we read it on 17 October 2026, and " $4M+" in daily payouts on 11 October - Mercor. But the averages hide a split. DataAnnotation's own pay bands put generalist work at $25 to $50 an hour against $40 to $125+ for credentialed fields - DataAnnotation, and one Outlier tasker reported effective pay cut to about $20 an hour from $50 after Meta's investment in Scale AI. Generalist labelling is being automated; credentialed judgment is not, yet.

What raises a rate: a credential and real experience in the field; a scarce language; skill in specialist software few people know; and a record of reliability over many accepted units. What lowers it: general tasks anyone can do, and work a model can already check by itself. The rights rule here is simple but often broken: work made to a buyer's brief belongs to the buyer under the contract, and past work written for an employer usually belongs to the employer, and work written for a client often belongs to the client under its contract, so check the contract before you license it. The Handshake documents offer is explicit that only documents the person owns or may share qualify.

For a professional, the practical takeaway is that the posted rate is what you are paid per accepted unit, not per hour spent, and that the scarce combinations (a regulated profession plus a less common language, or a niche software tool plus deep domain knowledge) are where rates hold up. The expert work page and the vendor page for Mercor track posted rates as they change.

13. Quoted prices versus realised prices

Every figure in the tables above is either a price someone paid, a price someone offered, or a range someone published. Those are three different things, and the difference between them is where most holders lose money. A published range is a ceiling and an advertisement. Buyers and marketplaces publish ranges to attract supply; the top of a range is what an exceptional holder got, and most holders land near the bottom.

The cleanest public example is creator footage. In January 2025 the reported rate for unused creator video was $1 to $4 a minute. The one creator who later published his actual earnings, Peter Hollens, made $33,000 from licensing about a thousand hours, which Semafor described as " more than $33 per hour of content" - Semafor. Per minute, that is about 55 cents, a little over half of the bottom of the quoted range and less than a seventh of its top. Nobody lied: big catalogues of general footage simply fetch less per minute than the curated, high-spec clips that set the top of the quote.

The same pattern shows up wherever both numbers are public. The table below puts quoted and realised figures side by side.

There are four mechanisms behind the gap, and each one can be managed. The first is acceptance: buyers pay for the hours, tasks or pages they accept, not for what was delivered, and rejection rates of 15% to 25% are normal in robot data. For recorded work, the ratio of effort to accepted output can be steep: one tutor in Delhi spends far longer making each video than it runs, because he keeps having to invent new chores - MIT Technology Review. The second is the sample review: the published range is set before anyone has looked at your data, and the offer is set after. The third is costs that come off the top, such as de-identification, which the buyer paid for in the Spirit sale but which other buyers net out of their offer. The fourth is the single-buyer offer, which never shows you what a competing buyer would have paid.

The practical defence is simple to state. Treat any published range as the upper bound, expect your first offer near the bottom, and get a second and third offer before accepting the first. That last point is the core of what an agent does: AIData.Broker's approach when acting for a holder is to put the data before every buyer that fits at the same time, and the same datasheet before each one that is interested, because a buyer competing with others prices differently from one that knows it is alone. The market tracker shows which buyers are currently active by kind of data.

14. Follow the dollar: unit economics

To understand prices, it helps to follow a single dollar from the lab that pays it to the person or company whose data or work it buys. The intermediaries in between are not a detail. They take a large share of every dollar, and their margins explain why some kinds of data get resold many times while others are bought once.

Start with the largest flow, expert hours. Mercor " pays out 60% to 70% of its topline revenue to contractors" - Dealroom. At the average contracted rate of $112 an hour shown on its homepage, that implies the lab pays somewhere around $159 to $185 for the hour and Mercor keeps roughly $48 to $74 (our arithmetic: the rate divided by the share paid out). Mercor's reported gross margin was " only 27% last year" and " 33% in the second quarter of this year" - BigGo Finance, citing The Information. Vendors that run delivery from lower-cost regions do better: one practitioner writes that they " keep 70-80%+ of every dollar" - Thoroughly Intrigued.

Now compare a dataset sold off the shelf. micro1's off-the-shelf datasets, sold to more than one customer, carry margins " as high as 80% to 90%", according to a person familiar with its finances - TechCrunch. That margin gap, roughly 30% on labour against 80% to 90% on resold data, is the single most important fact about where this market is going. It is why the labour vendors are now buying company archives and building environments: data that can be packaged once and sold several times is far more profitable than hours that must be bought fresh for every order.

The company-archive row deserves a closer look, because it is where this market is newest. A principal buyer such as micro1 pays the company, then turns the records into training environments and datasets, which it can license to several labs. The company is paid once; the buyer earns on every resale. That is not unfair in itself (the buyer carries the cost of cleaning, de-identifying and packaging), but it is why a company should know whether it is selling outright or licensing non-exclusively, and whether its records will be resold. The finder's fee of $50,000 is also informative: a buyer willing to pay that much just for an introduction is telling you something about the margin it expects.

The growth of the largest labour vendor shows how fast the money arrived. The chart below traces Mercor's reported revenue pace from early 2025 to mid-2026. These are gross figures (before the 60% to 70% paid out to contractors) and run rates, which one practitioner warns are " usually the best month multiplied by 12" - Thoroughly Intrigued. Read them as the pace of lab spending through one vendor, not as profit.

Sources: February 2025, March 2025, September 2025, early 2026, June 2026.

The individual points come from separate reports: $75 million ARR in February 2025 - TechCrunch, " a $500 million pace in September 2025" and " over $1 billion in annualized gross revenue" in early 2026 - The Information, and over $2 billion in June 2026. The H1 2026 figure, $614 million of gross revenue, with more than 90% from AI model companies, is a useful reality check on the run rate. Other vendors grew in the same direction: micro1 went from $100 million to a $500 million gross run rate in eight months, and Snorkel AI said its run rate " now stands at $375 million" - TechCrunch.

What this means for pricing is direct. For data made by people, the price a lab pays is roughly the person's rate grossed up by a vendor margin of 30% to 40%. For data that exists and can be packaged, the price a buyer will pay the holder is bounded by what the buyer expects to earn reselling it, minus its cost of cleaning and de-identifying. A holder who understands that second calculation knows why a buyer who plans to resell non-exclusively can afford more than one who will use the data once, and why asking about resale rights is never a rude question.

15. The biggest disclosed deals

The largest training-data deals by disclosed value are dominated by text and platforms, with stock imagery and scholarly publishing behind them. Almost every per-deal figure is reported by press from unnamed sources rather than stated by the parties; the figures stated by the parties tend to be aggregate revenue lines in filings, covering several buyers at once. The table below lists the largest money figures on record, with their basis.

Treat the table with two cautions. First, it ranks what was disclosed, not what was paid: the largest deals in the market may well be among the 92 in the ledger without a figure. Second, several entries mix training with display (showing content inside AI answers), and a few are revenue totals rather than single deals. The "Basis" column says which.

Three patterns come out of that list. Recurring sources get annual licences: news, Reddit and anything that refreshes is licensed by the year, for two to five years, while static archives (the Wiley projects, the Spirit archive, a book backlist) sell once or for a fixed term. The top tier is narrow: these are household names with archives nobody else has, and a mid-sized publisher or platform should expect an order of magnitude less. And disclosed revenue lines are small next to the headlines: Reddit's licensing is about 5% of its revenue, and Wiley's $49 million covers every AI licensee it has. The one figure that dwarfs all of them is not a deal at all: micro1's commitment of $1 billion over twelve months to buy company data, which is a budget, not a closed price.

If you want to work with the underlying rows (every deal, its terms, its basis and its source), the full ledger is browsable on the deals page and downloadable in CSV and JSON from datasets, and the buyers behind them have their own entries on the buyer board.

16. How to estimate your own data's range, and its limits

The anchors in this guide make it possible to put a rough range on a dataset before talking to anyone, as long as you are honest about what the range means. AIData.Broker's valuation tool does this arithmetic in the open, and its method is worth explaining because the same method works on paper. It takes a unit price range from the public price board, multiplies it by the size you enter, and adjusts for the few things the evidence says move a price (linkage, scarcity, rights); for company records, codebases and task sets it runs a fuller model that starts from what buyers publish. It shows its working and its sources under every result.

The important part is what is assumption and what is published. The unit ranges come from sourced public prices. The multipliers (how much more linked records are worth, how much a scarce language adds, how much a closed company is discounted) are our reading of those prices, not published rates, and the tool labels them as assumptions. Here are the unit ranges the tool uses and where each comes from.

A worked example shows how the arithmetic behaves. Take the tool's starting company: 40 people at its largest, eight years of records, operating in the United States, some linked systems, a common field. The tool runs the buyers' shared formula on it, halves the result for what deals actually realise, and, because it has only your answers, shows four times either side of that middle: $74,000 to $590,000. Make the work traceable end to end and the range rises by 0%; make it a specialised trade and it rises by 0%; make the company closed and loose and it falls by 81%. Compare that with Handshake AI's own estimator, which, on our reading of its outputs, matches the same formula without the halving: $383,083 to $896,969 for a 100-person US company with ten years. The tool's range is wide on purpose, and neither has seen your data, which is the point.

Three more examples using the same tool and its starting answers: 400 hours of general footage with releases comes out at about $13,000 to $48,000; 200 hours of consented English two-way conversation at $5,000 to $30,000; and 400 graded software tasks at $17,000 to $1,100,000. The width of that last range is honest. Task prices genuinely span an order of magnitude depending on domain, difficulty and grader quality, and no formula can tell you where yours sit without a buyer testing them.

Now the limits, which matter more than the numbers. First, no buyer has priced your data; only a buyer's written offer prices it, and offers routinely land below published ranges (section 13). Second, the tool shows non-exclusive figures; an exclusive sale can fetch four to five times more, and a non-exclusive one can be sold more than once, so the right number depends on a decision you have not made yet. Third, rights can zero the value: speech without every speaker's consent is not valued at all, and company records with uncleared personal data or restrictive software terms may not be sellable. Fourth, expert work has no valuation, because it is not an archive; it is priced by the hour or the accepted unit when a buyer commissions it.

How to use a range, then. Use it to decide whether a sale is worth the effort (a range of $100 to $1,300 for a small image library says list it on a marketplace and move on; a range of $74,000 to $590,000 for a specialised company's records says build a rights file and get competing offers). Use it to recognise an offer that is far below public anchors. Do not use it as an asking price. If the range justifies a sale, the selling guide explains how AIData.Broker, as the holder's agent, puts a holder's data before every buyer that fits, and if you are on the buying side looking for a specific kind of data, the acquire page is where to describe what you need.

17. Where prices are heading: a two-tier market

Put all nine kinds side by side and the direction of 2026 is clear. The market is splitting into two tiers. Scarce, verifiable data is getting more expensive: company records that can fill a realistic workplace, expert-built tasks and environments, credentialed human judgment, uncommon languages, teleoperation on real robots. Commodity data is getting cheaper: general stock images, generic footage, bulk first-person video, plain read speech, ordinary code, generalist labelling. The dividing line is not the format. It is whether a buyer can get the same thing some other way.

The first-principles reason is the shift in how models are trained. When the bottleneck was reading, any large body of text or images was valuable, and the first deals priced volume. Now the bottleneck is practising real work and being graded on it, so what is valuable is whatever makes practice realistic and grading reliable: real records, real expertise, real physical demonstrations. At the same time, models themselves have become a substitute for anything they can generate and verify. Together AI's synthetic coding trajectories at about 50 cents each are a small example of a large force: whatever a model can produce for itself caps what anyone will pay for it.

Three further shifts are worth watching, because they change how prices are expressed as well as their level. The first is the move from flat fees to pay-per-use, led by news: Microsoft's marketplace pays per use, and Perplexity said " 80% of Comet Plus revenue will go to participating publishers" - Shelly Palmer. The second is big holders building instead of selling: holders of unique data (studios, hospital systems, large platforms) increasingly keep it and co-build models, which leaves intermediaries the fragmented supply held by many small holders. The third is decay: environments have a half-life of months, exclusivity erodes as similar data reaches rivals, and the window for any given kind of data can close quickly, as the creator-footage market showed.

What it means in practice depends on which side you are on. A holder of scarce data has leverage now and should use it deliberately: decide exclusivity on purpose, insist on a rights file, and get competing offers while demand is high, because the premium for being first will not last. A holder of commodity data should sell through the cheapest efficient channel (a marketplace that pays per unit) and not spend months negotiating over cents. A buyer should expect to pay more each quarter for linked, expert, consented data, and less for anything that synthetic generation or a hired crowd can replace. The market tracker records which way each kind is moving as new evidence comes in.

18. FAQ

How much is AI training data worth per gigabyte or terabyte? There is no meaningful per-gigabyte price, and nobody in the market prices that way. Data is priced per company, repository, minute, hour, image, title or task, because those units track what a buyer can use. Size matters only after deduplication and only after linkage, scarcity and rights; about 37% of raw code disappears in cleaning alone.

What kind of data sells for the most right now? By single transaction, whole archives of operating companies and the large news and platform licences sit at the top, from hundreds of thousands to tens of millions of dollars. By unit, expert-written tasks and environments command the highest prices per piece of work ( $200 to $2,000 per task, up to $300,000 per complex environment). By growth, company records and RL environments are where demand rose fastest in 2026.

Can I sell my company's Slack, email and tickets? Often yes, but owning the files is not the right to license them. Customer contracts, employee notices, software terms and third-party material inside the archive all need checking, and personal data has to be removed. Buyers publish ranges from $10,000 for a small closed startup to $1 million and more for a large operating company; see company records.

Is it better to sell exclusively or non-exclusively? Exclusive sales have been reported at roughly four to five times the non-exclusive price for environments. Non-exclusive licences can be sold to several buyers. If three or more buyers would license the same data, non-exclusive can be worth more in total; if one buyer badly wants to keep rivals out, exclusive pays. Decide before you negotiate, not during.

Why was my offer so far below the published range? Published ranges are ceilings meant to attract supply. Offers come after a sample review, net of de-identification and other costs, and paid on accepted units. A single buyer also has no reason to bid against itself. Getting several offers at once is the most reliable way to close the gap.

Do AI companies still buy images and stock video? Yes, but less than before. The largest stock agencies reported falling data revenue in late 2025, and new deals favour display rights over training. Specialist libraries with metadata and releases still sell; everyday photography is close to a commodity at a few cents an image.

How do I know a buyer's price is fair? Compare it with the public anchors for your kind of data (the price board has them, dated and sourced), check what the offer does and does not include (exclusivity, resale rights, who pays for de-identification), and get at least one competing offer. A fair price is the one that survives competition, not the one in a single email.

Does AIData.Broker buy data? No. AIData.Broker acts as the holder's agent: you keep ownership of your data and license it to the buyer yourself. We do not buy a holder's data, take ownership of it, or license or sell it in our own name. You answer the rights file, we prepare the datasheet from your answers, and once you approve it, every buyer that fits hears of the data without your name, and each one that is interested gets the datasheet after its written undertaking, on one deadline. You, or a provider you choose, deliver the data to the buyer, and the buyer pays you. For single repositories or small bulk image libraries, we usually suggest a self-serve platform instead, because an agent adds little there.

19. Conclusion: a decision framework

The prices in this guide span nearly nine orders of magnitude, from two cents per image to ten million dollars per archive, and almost every one of them comes with a caveat. That is not a failure of research; it is what this market is. Training data is valued by what it teaches a particular model at a particular moment, there are few buyers, and most deals are never disclosed. What a holder or buyer can do is know the anchors, know the multipliers, and know where their own data deviates from both.

For a holder of data, the decision runs in five steps. First, identify the kind of data and its unit, and find its public anchors in the tables above. Second, check rights before anything else, because rights can take the value to zero and no buyer will sign without them. Third, ask whether your data is scarce (a buyer could not get it another way) or commodity (a buyer could); scarce data deserves an agent and competing offers, commodity data deserves the cheapest efficient marketplace. Fourth, decide exclusive or non-exclusive on purpose, and decide who controls de-identification. Fifth, estimate a range with honest assumptions (the valuation tool shows its working), treat it as a ceiling, and never accept a single offer without testing it.

For a buyer, the same facts point the other way. Use comparables to set a starting point, insist on a per-asset rights file and documented consent, measure the lift of each purchase on held-out tests before buying again, and expect to pay more each quarter for linked, expert, consented data while paying less for anything a model can generate. If you are looking for a specific kind of data that is hard to find, the acquire page is where to describe it.

AIData.Broker updates the price board, the deal ledger and the market tracker as new figures are published, each with its source. If you hold data and want to know what the public prices imply for it, start with the valuation tool; if you want competing offers from buyers, the selling guide explains how that works.

This guide reflects public prices, deals and vendor figures as of 11 October 2026. Prices in this market move within a quarter, several figures are reported by press rather than stated by the parties, and published ranges are ceilings rather than offers. Verify current figures at their sources before relying on them.