What AI Labs Pay for Data: A Price Map of the Licensing Market

Until 2024, conversations about what AI companies pay for data relied almost entirely on rumour. Agreements were confidential, and the parties rarely discussed numbers. That changed quickly. Public companies began reporting AI licensing income in filings and earnings calls, and several high-profile agreements were reported in enough detail to give the market its first set of reference prices.

Those figures matter to anyone who owns data. A reference price turns an abstract opportunity into something that can be compared, modelled and negotiated. It also shows that the market is not a single price but a spectrum, with the value of each agreement shaped by what is being licensed, how often it is refreshed and what rights come with it.

This study brings together the most reliable public data points: amounts disclosed by the companies themselves in filings or investor material, and widely reported figures from established financial publications. Where a number is a press report rather than a company disclosure, it is described that way.

Taken together, the table shows a market with a wide range but a consistent logic. The largest figures belong to owners of vast, unique and constantly growing archives. Specialist publishers with deep but narrower libraries earn tens of millions a year across multiple buyers. And individual agreements between a single publisher and a single laboratory tend to sit in the low tens of millions per year.

Reddit’s registration statement (opens in a new tab) remains the most detailed public disclosure in the market. It reported that in January 2024 the company entered data licensing arrangements with an aggregate contract value of $203.0 million and terms of two to three years, with at least $66.4 million expected to be recognised in 2024. Google’s announcement (opens in a new tab) of an expanded partnership, including access to Reddit’s Data API, was published the same day.

For context, Reddit reported total revenue of $804.0 million for 2023. A data licensing line of that size, created within a single year, shows how quickly an archive of human conversation can become a material revenue stream once demand appears.

The structure is as instructive as the number. Reddit’s data is valuable not only because of its size, more than a billion posts and over 16 billion comments by the end of 2023, but because it keeps growing. Access through an API means buyers receive a continuing flow of fresh material, which supports multi-year contracts rather than a single sale.

Stack Overflow (opens in a new tab) followed a similar model in its May 2024 partnership with OpenAI, providing access to its developer knowledge through its OverflowAPI product, with attribution back to the community in OpenAI’s products. Developer questions and accepted answers are a near-perfect example of expert, structured and continually refreshed data.

In May 2024, News Corp and OpenAI announced (opens in a new tab) a multi-year global partnership covering content from mastheads including The Wall Street Journal, Barron’s, MarketWatch, The Times and The Australian. The Wall Street Journal reported the agreement to be worth more than $250 million over five years, including cash and credits for OpenAI technology.

IAC (opens in a new tab) disclosed in a shareholder letter that OpenAI pays its Dotdash Meredith division a minimum of $16 million a year, with additional variable payments. The Wall Street Journal later reported that Amazon agreed to pay The New York Times $20 million to $25 million a year for access to editorial content, including recipes and sports coverage, for use across Amazon products and AI models.

A pattern runs through these agreements. Training rights are only part of the package. Many also grant the right to display content inside AI products, often with attribution and links back to the source. That makes them partly licensing deals and partly distribution deals, and it explains why they tend to be structured as recurring annual payments rather than one-off purchases.

Shutterstock (opens in a new tab) reported $104 million in revenue from licensing its content library for AI training in 2023 and has set a target of $250 million by 2027. Its library of licensed images, video and associated metadata is exactly what developers of image and video models need: large, well labelled and rights-cleared. In January 2025, Getty Images and Shutterstock agreed to merge (opens in a new tab), combining two of the largest rights-managed visual libraries in the world.

Academic publishers show a similar profile. Wiley (opens in a new tab) reported $44 million in AI content licensing revenue in fiscal 2024, and Informa (opens in a new tab) said AI partnerships at Taylor & Francis would generate more than $75 million in 2024. Peer-reviewed research is expert content of the highest order, with clear structure, citations and editorial standards, and it can be licensed to several laboratories at once.

The lesson for owners of specialist material is that the market does not reward size alone. A focused library that is uniquely deep in one field, carefully organised and cleanly licensed can generate repeat income from multiple buyers.

Comparing these agreements side by side shows that price is driven by a small number of factors, each of which an owner can assess before entering a conversation.

• UniquenessMaterial that exists nowhere else, and cannot be recreated by scraping or synthesis, anchors the highest prices.
• Volume and coverageLarger and broader collections carry more training value, especially when they span years or many subject areas.
• FreshnessContinually updated sources support recurring, multi-year contracts instead of a single sale.
• ExpertiseContent written or reviewed by specialists teaches reasoning that general text cannot, and commands a premium.
• Structure and metadataClean formatting, labels and metadata cut a buyer’s preparation costs.
• Documented rightsClear ownership and consent remove legal risk, which buyers increasingly pay for directly.

Deal structure also affects value. Most public agreements combine a fixed annual fee with variable payments linked to usage or product performance. Exclusivity, where it is granted, typically raises the price; non-exclusive licences allow an owner to serve several buyers at once. Scope matters too: training-only rights are usually priced differently from rights that also allow display inside a product.

Although the commercial terms of most licensing agreements are confidential, the public disclosures reveal a consistent structure. Understanding it helps owners frame their own conversations and recognise which levers matter most.

The first element is scope: which datasets are included, over what time period, and in what formats. Reddit’s agreements, for example, are built around continuing access to its content through an API, while publisher agreements typically cover defined archives plus new material as it is published. The narrower and better defined the scope, the easier it is for both sides to value.

The second element is the permitted use. Training rights allow a buyer to use the data to build or improve models. Display rights allow content to appear inside AI products, often with attribution and links back to the source, as in OpenAI’s publisher agreements. Some agreements cover both, and the combination usually commands a higher price than training alone.

The third is payment structure. IAC’s disclosure that OpenAI pays Dotdash Meredith a minimum of $16 million a year plus variable payments is a typical shape: a guaranteed annual amount that gives the owner certainty, combined with an upside linked to usage or performance. Reddit’s disclosure of an aggregate contract value spread over two to three years shows the same preference for multi-year terms.

The fourth is exclusivity. Most public agreements appear to be non-exclusive, allowing owners such as Wiley and Taylor & Francis to license the same material to several buyers. Exclusivity, where granted, is usually priced at a premium because it removes the owner’s ability to sell elsewhere.

Finally, agreements set out obligations on both sides: how data is delivered and secured, how personal information is handled, how attribution works, and what happens to the data when the agreement ends. These terms rarely make headlines, but they are where much of the long-term value of an agreement is protected.

The figures in the price map are significant not only in absolute terms, but relative to the businesses that earn them. Reddit’s $203.0 million in contracts sits against total 2023 revenue of $804.0 million. Shutterstock’s $104 million in AI licensing revenue in 2023 was achieved alongside its traditional stock content business, and its target of $250 million by 2027 shows how central it expects the line to become.

For academic publishers, AI licensing has become a new and high-margin income stream built on content that was already created, edited and paid for. Because the underlying archive exists regardless, much of the licensing revenue flows directly to profit. That combination of recurring income and high margins explains why publishers have moved quickly to put agreements in place.

The broader point is that AI licensing is no longer a curiosity in investor presentations. For the companies that have entered the market, it is a reported line, tracked by analysts and built into forward targets.

Much of the early licensing market centred on text for language models. As laboratories build models that generate and understand images, video, speech and music, demand has widened. Shutterstock’s results show how a rights-cleared visual library can generate substantial AI revenue, and the planned combination of Getty Images and Shutterstock brings together two of the largest such libraries.

Code is another category with clear value. Stack Overflow’s partnership with OpenAI reflects the importance of high-quality technical questions and answers for models that write and explain software. Developer knowledge is structured, expert and constantly refreshed, which makes it close to an ideal training source.

Each of these categories has its own buyers and its own pricing dynamics, but the same principles apply: unique, well-labelled, rights-cleared collections command the strongest terms.

The public reference points give owners a sense of the market, but every dataset is priced on its own merits. The owners who achieve the strongest terms tend to arrive at a conversation with a clear picture of what they hold and what they want.

That begins with a description a buyer can evaluate: the volume of material, its format, the date range it covers, the subjects it spans and how it is refreshed. It continues with a clear rights position, setting out what the owner holds outright and what consent or contractual basis applies. And it includes a view on the terms that matter most to the owner, such as permitted uses, exclusivity, duration and attribution.

A representative sample is often the most persuasive part of the package. A small, cleaned extract lets a buyer test quality and fit without exposing the full dataset, and allows a conversation about value to rest on evidence rather than description.

Owners should also think about the long term. Because much of the value in this market comes from recurring arrangements, the first agreement is often the start of a relationship. Terms that work well over several years, and that allow for fresh data to be supplied as it is created, tend to produce more value than a single, larger upfront payment.

Finally, the market rewards owners who understand where their data fits. A dataset that is ordinary to a frontier laboratory can be exceptional to a company building a model for a single industry. Identifying the buyers for whom a dataset is most relevant is often the most important step in achieving the best price.

Timing plays a part as well. Demand for high-quality data is strongest while laboratories are racing to build and improve their next generation of models, and Epoch AI’s research places the period of greatest pressure on public data between 2026 and 2032. Owners who prepare during that window can meet buyers when their need is most acute, which tends to support firmer terms and longer relationships than waiting until the market has matured and pricing references have settled.

The number of public reference points will keep growing as more listed companies report AI licensing income. That transparency benefits owners. A market with known prices is one where agreements can be reached faster, terms can be compared and value is less likely to be left on the table.

Demand is also broadening. The first wave of agreements focused on text for language models. The next is reaching into images, video, audio, code and highly specialised professional data, each with its own buyers and price dynamics. As laboratories build models for specific industries, the value of narrow but deep datasets is likely to rise further.

For a company that owns unique data, the practical takeaway is straightforward. The market now has enough evidence to support a serious valuation conversation, and the owners who prepare well, with a clear inventory, documented rights and a view on the terms they want, are the ones best placed to capture it.