# Pricing Models in AI Training Data Licensing

Source: https://sellyouraidata.com/research/ai-data-pricing-models/
Updated: 2026-10-20

## Abstract

**Background:** Licensing of data for AI training has produced a small number of large, widely reported deals and a larger number of small or private ones. There is no standard price list, and published figures differ in unit, term and reliability.
**Methods:** We reviewed deal announcements, company filings, press reports, buyer program pages, marketplace documentation and academic work on data valuation and data markets, using sources available as of October 2026. Each figure is labelled as disclosed, reported from named or anonymous sources, or indicative, and we include a source only if we opened and read it.
**Findings:** Reported structures include fixed annual fees, one-time payments for historical archives plus recurring fees, per-unit rates for images, video, text and books, tiered ranges on buyer program pages, per-task and per-environment prices for reinforcement learning environments, and annual list prices for licensed datasets. Reported amounts range from fractions of a cent per word or image to tens of millions of dollars per year for large content owners. Exclusivity is reported to multiply prices by roughly four to five in one segment. Academic valuation methods quantify the contribution of data to a model but are not used in any public deal we found.
**Implications:** Published figures are not comparable without their unit, term, exclusivity and scope, and almost none are rate cards for ordinary business data. Price appears to follow the buyer's estimate of replacement cost and expected training value, adjusted for rights, exclusivity and the cost of making the data usable.

## 1. Introduction

When a company licenses data to an AI developer, there is no established unit of account. Photographs have been priced per image, video per minute, books per title, news archives per year, online communities by multi-year contract, and corporate archives by auction. Each of these figures answers a different question, and placing them side by side without their context produces misleading comparisons.

This review sets out the pricing structures that appear in public sources, collects the figures that have been published or reported, and notes what each figure does and does not show. It also summarizes academic work on how data could be valued, because that literature shapes how researchers think about price even though it has not, as far as we found, set a price in a public deal.

The review does not advise on what any dataset is worth and does not suggest that any figure is a benchmark for a particular company. We also do not assert that any named company buys any category of business data beyond what its own cited page or a cited report states.

## 2. Methods and sources

We read sources from five groups between September and October 2026.

1. Primary company disclosures: deal announcements and annual reports filed with the U.S. Securities and Exchange Commission (SEC).
2. Press reports of deal values. Most of these cite anonymous people familiar with the matter or secondary reports of filings, and we say so for each figure.
3. Program pages published by AI data companies, which describe payout ranges or tiers for company data. We treat these as indicative statements of intent, not as agreed prices.
4. Marketplace and platform documentation, to see which pricing mechanisms are publicly described.
5. Academic work on data valuation, data pricing and data markets. Most of this literature is on arXiv, and we label preprints as such.

We labelled each figure with one of three reliability tags: **disclosed** (stated by a party to the deal or in a filing), **reported** (stated by a news outlet, with sources identified or anonymous) and **indicative** (a range on a program page that the page itself does not treat as an offer). We include only sources we opened. When sources disagree we report both. Currency amounts are as stated in the source and are not adjusted for inflation or exchange rates. All statements about company pages are as of October 2026.

## 3. Findings

### 3.1 Pricing structures

Table 1 lists the structures that appear in the sources, with an example of each. A given deal can combine several.

**Table 1.** Pricing structures observed in public sources, with examples and caveats.

| Structure | How it works | Example in sources | Caveat |
|---|---|---|---|
| Fixed annual fee | A set payment per year for access or training rights | OpenAI and Dotdash Meredith, reported at no less than about $16 million per year as a fixed component[^engadget-dotdash] | A fixed component may be supplemented by a variable one |
| Multi-year contract value | A total over a stated term | Reddit's aggregate contract value of $203.0 million for arrangements of two to three years[^tc-reddit] | A total is not an annual figure, and revenue is recognized over time |
| One-time archive payment plus recurring fee | Upfront payment for historical content and ongoing payments for current content | Axel Springer and OpenAI, as reported by the Financial Times[^decoder-axel]; Taylor & Francis and Microsoft, an initial fee of around $10 million plus recurring payments[^ii-informa] | Terms beyond the headline are usually undisclosed |
| Per-unit rate | A price per image, minute, word, book or hour | Video footage at $1 to $4 per minute[^decoder-video]; books at $5,000 per title in one reported program[^authorsguild] | Rates differ by quality, rarity and buyer |
| Revenue share | Split of proceeds between owner and intermediary or author | A per-title fee split 50 to 50 between author and publisher[^authorsguild] | This is a split between sellers and not a share of the buyer's revenue |
| Tiered or banded payout | Ranges tied to data quality, uniqueness or scope | micro1 tiers of $100k, $500k and $1M or more; Handshake's $100K to $4M+ range[^micro1-data][^handshake] | Program pages are indicative, not offers |
| Per-task or per-environment | Price per task or per simulated environment | Roughly $200 to $2,000 per task and about $20k for a basic website replica, as reported by interviewees and SemiAnalysis[^epoch-rl] | Reported by interviewees; contract totals also reported |
| Exclusivity premium | Higher price for sole rights | Roughly four to five times a non-exclusive price, per two founders[^epoch-rl] | One segment; two sources |
| Annual dataset license | A list price per dataset per year | $70,000 per dataset per year in a university program for FY25[^aimi] | Applies to a specific program |
| Auction | Competitive bidding in a sale of a company's assets | Google's $10 million winning bid for Spirit Airlines data[^time-spirit] | Price depends on the process and the privacy conditions |
| Option or renewal | Initial term with extension at the buyer's or seller's choice | Annual licenses renewable on request[^aimi] | We found no public deal with an explicit pricing option |
| Usage or subscription on a platform | Charges per access, per trial or per listing | Snowflake paid listings, trials and consumer-specific offers[^snowflake] | Applies to data products in general, not AI training in particular |

Two features of this table deserve emphasis. First, most reported structures are lump sums or fixed periodic fees. Usage-based pricing, in the sense that a data owner is paid according to how much a model uses its data, is not described in any AI training deal we found. Attribution of model output to training data is an unsolved technical problem, which is one reason academic methods in Section 3.7 matter. Second, the option and renewal structure is common in practice, since licenses have terms, but we found no public source that prices the option itself.

### 3.2 Deals by large content owners

Large owners of text, images and community content have produced the best-documented figures. Table 2 collects them. The table includes the tag for reliability, and the order is by owner type.

**Table 2.** Publicly reported or disclosed AI data licensing figures from large content owners, with structure and reliability.

| Owner and buyer | Figure | Structure as reported | Reliability |
|---|---|---|---|
| Reddit and Google | About $60 million per year[^reuters-reddit] | Content for training, term not stated in the report | Reported, from three anonymous sources |
| Reddit, all licensing arrangements entered in January 2024 | $203.0 million aggregate contract value; at least $66.4 million recognized in 2024[^tc-reddit] | Arrangements of two to three years | Disclosed in IPO filing, as relayed by TechCrunch |
| News Corp and OpenAI | Could be close to $250 million over five years, or about $50 million a year[^vb-newscorp] | Current and archived content from News Corp publications | Reported, via the Wall Street Journal |
| Axel Springer and OpenAI | Tens of millions of euros per year[^decoder-axel] | Upfront payment for historical content plus recurring licensing; non-exclusive | Reported, via the Financial Times |
| Dotdash Meredith and OpenAI | At least about $16 million per year[^engadget-dotdash] | Fixed component, with a variable component to be recognized later | Reported from public financial documents and an earnings call |
| Taylor & Francis (Informa) and Microsoft | Initial data access fee of around $10 million, plus recurring payments over three years[^ii-informa] | Data access for Microsoft; exclusivity not stated in the source | Disclosed by Informa, as relayed by a financial publication |
| Wiley | $44 million from AI partnerships: $23 million already earned and $21 million expected[^infodocket-wiley] | AI partnerships, not itemized in the source | Reported by The Bookseller, as relayed by infoDOCKET |
| Associated Press and OpenAI | Not disclosed[^pbs-ap] | Part of AP's text archive from 1985 | Financial terms not disclosed |
| Stack Overflow and OpenAI | Not disclosed[^so-openai] | API access to validated technical content with attribution | Financial terms not disclosed |
| Shutterstock and large technology companies | $25 to $50 million for initial deals with Apple, Meta, Google and Amazon, since expanded[^decoder-spree] | Hundreds of millions of images, videos and music files | Reported, via Reuters |

Three observations follow. First, the large figures concern owners with distinctive, continuously updated, human-written content, often of known provenance. Second, structures differ: Dotdash Meredith's deal has a variable element, and Axel Springer's combines an archive payment with ongoing fees.[^engadget-dotdash][^decoder-axel] Third, the quality of the evidence varies. The Reddit and Google figure comes from anonymous sources and is not the same as the $203.0 million total in the IPO filing, which covers all arrangements and a different period.[^reuters-reddit][^tc-reddit] The News Corp figure is a figure the Wall Street Journal said could be close to $250 million, and VentureBeat noted that the valuation was not officially disclosed.[^vb-newscorp]

The Shutterstock 10-K gives the clearest primary view of how data licensing appears in financial statements. Shutterstock's Data, Distribution, and Services segment reported revenue of $38.5 million in 2022, $137.3 million in 2023 and $175.3 million in 2024, 19 percent of total revenue of $935.3 million in 2024.[^sstk24][^sstk23] The filing says customers have shown increased demand for access to its metadata for machine learning and generative AI model training, and it states that segment results fluctuate with the timing of delivery of large metadata sets and the achievement of agreed milestones.[^sstk24][^sstk23] The segment includes distribution and services as well as data, and the filings we read do not separate the AI training licensing revenue, so the segment totals should not be read as the price of AI licenses. The filing text does show that revenue from data deals is irregular, which is consistent with lump-sum or milestone-based structures.

### 3.3 Per-unit prices

Per-unit prices are the closest thing to a rate card, though they mostly concern media. Reuters, as relayed by The Decoder, reported in April 2024 that Photobucket's chief executive was negotiating to license 13 billion photos and videos at five cents to $1 per photo and more than $1 per video, that Freepik had licensed most of its 200 million image archive to two technology companies at two to four cents per image, and that Defined.ai, which licenses to large technology companies, charged $1 to $2 per image, $2 to $4 per short video, $100 to $300 per hour of longer film and $0.001 per word of text.[^decoder-spree] These are reported prices from one article and may not reflect current rates.

In January 2025, Bloomberg, as relayed by The Decoder, reported that companies were paying $1 to $4 per minute for unpublished video, with $1 to $2 for standard unused footage and up to $4 for high-quality 4K footage, drone video and 3D animation, often through intermediaries.[^decoder-video]

For books, HarperCollins announced a program in November 2024 under which participating authors opt in and receive a fee of $5,000 per title split equally between author and publisher, with outputs capped at 200 consecutive words or 5 percent of a book's text, according to the Authors Guild.[^authorsguild] In a different setting, Anthropic agreed to settle an authors' class action for $1.5 billion, which Boston Globe reported as about $3,000 per book across an estimated 500,000 books.[^bostonglobe] A settlement for past use of pirated material is not a license price, and we include it only to show a per-work figure that exists in the public record, alongside the licensing figure. The two should not be treated as comparable.

Per-unit figures for enterprise data such as emails, documents or tickets were not available in the sources we reviewed. The one intermediary that discusses company data pricing states that it has no rate card, only public price points, a floor and a set of factors it does not list on the page.[^troveo]

### 3.4 Buyer program tiers for company data

Several AI data companies publish payout ranges for company data. These are the closest public statements on pricing for ordinary business data, and they are all indicative. Table 3 summarizes them.

**Table 3.** Published program ranges and terms for company data licensing, as of October 2026. All amounts are indicative statements from the companies' own pages.

| Company | Stated amounts | Stated basis | Other stated terms |
|---|---|---|---|
| micro1 | $100k or more for qualified partnerships; $500k or more for large-scale or ongoing datasets; $1M or more for highly unique proprietary data; referral rewards up to $50k[^micro1-data] | Quality, uniqueness and value of the data | Companies keep ownership and may review samples; eligibility reported as more than 30 employees[^tnw-micro1] |
| Scale AI | "Illustrative" $10K to $1M or more per partnership; referral bonus up to $50,000[^scale-dp] | Scales with cadence; designed for recurring revenue | 40 or more employees preferred; years of records required[^scale-dp] |
| Handshake AI | $100K to $4M or more in payouts; an estimate field described as indicative and not an offer[^handshake] | Data sources, volume and quality, due diligence and acceptance of terms | 20 or more employees and three or more years in operation[^handshake] |
| Turing (Project Lazarus) | Headline payout of up to $1M, though the payment section on the page we read showed a placeholder; referral reward up to $10k[^lazarus] | How complete, rare and well preserved the record is | Single payment; no retainer or recurring commitment; one valuation and one contract after sampling[^lazarus] |
| Mercor | No amounts stated[^mercor-data] | Data volume, tools connected and data depth | Wire payment within 2 to 4 weeks of signed agreement; Mercor states no fees[^mercor-data] |

The tables reveal how differently programs are built. Scale describes recurring partnerships, Lazarus describes a one-time purchase of a historical record, and micro1 and Handshake describe banded payments tied to quality.[^scale-dp][^lazarus][^micro1-data][^handshake] The stated bases are qualitative: volume, depth, uniqueness, completeness and rarity. None of the pages publishes a formula, and several note that the figure depends on due diligence or sampling. The ranges are wide, spanning two to three orders of magnitude, and the pages do not say what share of applicants reaches any tier. Lazarus's headline and its placeholder are an example of why such pages should be read as sales material.[^lazarus]

The micro1 company announced in October 2026 that it would commit $1 billion over twelve months to its Company Data Partnerships program, with financing from Citi and Hercules Capital, and said some datasets may be rewritten synthetically to weaken links to original records.[^tnw-micro1] The commitment is a budget, not a count of deals.

### 3.5 Distressed and wind-down sales

Sales by failed companies provide a view of prices outside marketing material, though the sources are reports. Forbes, as relayed by Gizmodo in April 2026, reported payouts of $10,000 to $100,000 per company for Slack and email archives sold through SimpleClosure's Asset Hub, which the report said had processed nearly 100 such deals in the past year.[^gizmodo] Gizmodo's own lead mentioned "hundreds of thousands of dollars," which does not match the per-company range in the same article, and we therefore use the Forbes range.[^gizmodo] MIT Technology Review reported in October 2026 that a former crypto startup received about $25,000 for its code base, with about $15,000 more expected for its workspace data, that the Asset Hub had handled over 150 deals, and that buyers' valuations vary widely because they depend on research theses estimating how many useful tasks the data could yield.[^mit-dead] The two deal counts differ by report date and should not be treated as conflicting.

The largest reported sale is Spirit Airlines' corporate data. Time reported that Google won a bankruptcy auction with a $10 million bid, with Mercor bidding $7.5 million and micro1 later offering $12.5 million after the auction closed.[^time-spirit] The Next Web reported that Google opened at $5 million, that Mercor countered at $5.2 million and then $7 million if it could take the raw data and anonymize it itself, and that Google paid all de-identification costs on top of the price. The same report said that the agreement required that data stay linked to individual people (referential integrity) and that the customer side of the data, including about 97.5 million customer profiles, was excluded.[^tnw-spirit] CSO Online described the data as including about 600 million email and chat records from 17,000 employees, 17 million OneDrive files, 20.5 million SharePoint items and more than 30 million recorded customer service calls, and noted that former employees and unions had objected and the data's fate was undecided at the time of writing.[^cso] The accounts of the data differ slightly in how they count items, which is a reminder that these descriptions come from filings and press summaries and not from a single audited inventory.

Two lessons for pricing are visible. The first is that prices for scrubbed and unscrubbed data differ. The Next Web reported that a bidder valued the unscrubbed data above its own bid for the scrubbed version, and the cost of de-identification was borne by the buyer and was separate from the price.[^tnw-spirit] The second is that the same dataset attracted bids from several parties with different valuations, which is consistent with price discovery by auction when no benchmark exists.

### 3.6 Reinforcement learning environments

Prices for simulated work environments are reported by interviewees in an Epoch AI review. It reports around $20k for website replicas ("UI gyms"), according to SemiAnalysis, and around $300k for higher-fidelity replicas of complex products such as Slack, per one interviewee. Per-task prices were described as roughly $200 to $2,000 mostly, with $20k rare but possible. Contracts were described as often six to seven figures per quarter, and one researcher saw $300k to $500k depending on task count. Epoch also reports exclusive deals costing roughly four to five times as much as non-exclusive ones, per two founders.[^epoch-rl] TechCrunch reported in September 2025 that Anthropic leaders had reportedly discussed spending more than $1 billion on RL environments over the next year, which indicates the scale of the budget these prices come from.[^tc-env]

Epoch's interviewees also described what makes an environment worth paying for: soundness (a high reward means the task was actually solved), a range of difficulty with no task at a pass rate of zero or one hundred percent, and quality maintained at scale.[^epoch-rl] These criteria link price to measurable properties of the product, and they explain why such deals are priced per task. They apply to environments built by vendors, which may or may not be derived from licensed enterprise data. The review does not say how much of an environment's price reflects underlying enterprise records.

### 3.7 Marketplaces and list prices

Public list prices for AI training datasets are rare. The Stanford AIMI program is an example: it states that de-identified Stanford clinical datasets, originally limited to non-commercial use, can be licensed commercially for an annual fee of $70,000 per dataset for agreements in the fiscal year ending August 2025, with one-year renewable terms, committee review and no negotiated fees or discounts.[^aimi]

General data platforms document how pricing mechanisms work without publishing prices. Snowflake's documentation describes free listings, limited trial listings of 1 to 90 days, paid listings, and pricing plans and offers that set consumer-specific billing terms, payment schedules and contract dates.[^snowflake] The AWS Data Exchange page says over 1,000 data sets are free and directs readers to individual listings for paid terms.[^awsdx] An Extruct directory of AI training data marketplaces lists fifteen companies by model, namely collect and license, rights brokerage and creator marketplace, and states no prices or revenue splits.[^extruct] Defined.ai's catalog could not be read in our fetch because it loaded dynamically, so we rely on the 2024 Reuters-based figures above for that company.

Law firm commentary describes contract terms that bear on price. A Malaysian law firm note says AI companies typically commit to a fixed annual license fee, that the fee should consider the value of existing content and content added later, that exclusivity should be settled at the outset, and that removing data from a trained model may not be feasible, so licensors may have to leave datasets in the model.[^hhq] A sample agreement memo recommends a perpetual license for training and use of the trained models, paired with a finite license for ongoing dataset access, and extension to other models on additional fee.[^termslaw] These are practitioner views and not market data, but they identify the points on which fees are negotiated: scope of models, term, access versus training rights, and exclusivity.

### 3.8 Academic work on data valuation and pricing

A separate literature asks how much data is worth in principle. It has not, as far as we found, been used to set a price in the public deals above.

Data Shapley, introduced by Ghorbani and Zou in a preprint, applies the Shapley value from cooperative game theory to measure each training example's contribution to a predictor's performance, shows that it satisfies natural properties of equitable valuation, and reports that it assigns low values to outliers and corrupted data and indicates what kinds of new data to acquire.[^dshapley] Jia and coauthors note that Shapley values often require exponential time to compute and propose faster approximation algorithms.[^jia] Data valuation with reinforcement learning (DVRL) learns a value for each example using a small validation set as a reward signal and reports that it identifies corrupted samples and improves domain adaptation by 14.6 percent and robust learning by 10.8 percent.[^dvrl] Influence functions, from Koh and Liang, trace a model's predictions to the training examples that most influence them, and have been used for debugging and detecting dataset errors.[^koh] These methods attribute value inside a given model and dataset. They require access to the data and the model, which a seller and buyer may not share before a deal.

A second line of work concerns markets. Agarwal, Dahleh and Sarkar propose a marketplace for training data and identify four properties that make data hard to price: it can be copied at no cost, its value depends on combination with other data, the value of accuracy varies by task, and its usefulness is hard to verify before use.[^agarwal] Fernandez, Subramaniam and Franklin propose data market platforms in which owners are rewarded for sharing and buyers post needs, and outline a research agenda for market design.[^fernandez] Pei's survey of data pricing covers economics, marketing, data management and machine learning and addresses how pricing models have evolved.[^pei]

Practitioner valuation draws on accounting-style approaches. Deloitte describes an income approach (including relief from royalty and with-and-without methods), a cost approach (replacement cost adjusted for obsolescence) and a market approach, and notes that the market approach is limited because transaction data is scarce and many deals do not disclose key value drivers.[^deloitte] The replacement-cost view matches how buyers reason in sources such as Epoch's per-task prices, and the scarcity of disclosed transactions explains why this review depends on reports.

### 3.9 Factors reported to affect price

Drawing the sources together, the following factors are associated with price. We list them as stated in the sources; none is a measured elasticity.

- **Uniqueness and rarity.** micro1 ties its tiers to uniqueness, Lazarus to rarity and completeness, and Scale to domain expertise that is hard to replicate.[^micro1-data][^lazarus][^scale-dp]
- **Volume and depth.** Mercor ties payout to data volume, tools connected and data depth, and Scale and Handshake mention volume and years of records.[^mercor-data][^scale-dp][^handshake]
- **Exclusivity.** Epoch's interviewees report a multiple of roughly four to five for exclusive environment deals; the Axel Springer deal was reported as non-exclusive.[^epoch-rl][^decoder-axel]
- **Quality and soundness.** Environment buyers reportedly prioritize soundness and difficulty calibration.[^epoch-rl]
- **De-identification state and cost.** In the Spirit sale the buyer paid de-identification costs separately and reportedly valued unscrubbed data more highly.[^tnw-spirit]
- **Rights and use limits.** The HarperCollins program caps output at 200 consecutive words or 5 percent of a book, and unwinding a license after training may not be feasible.[^authorsguild][^hhq]
- **Term and structure.** Fixed fees, variable components and recurring payments change the headline figure's meaning.[^engadget-dotdash][^decoder-axel][^ii-informa]
- **Process.** Auctions produced different bids from different parties for the same asset.[^time-spirit]
- **Buyer thesis.** Valuations of wind-down archives rest on estimates of how many useful tasks the data could yield.[^mit-dead]

## 4. Discussion

**Headline numbers conflate units.** Annual fees, contract totals, per-unit rates and settlements are all called "price" in popular coverage. The Reddit figures show the problem: a reported $60 million per year with one buyer and a $203.0 million aggregate across arrangements of two to three years are not in conflict, but neither is a price for a defined quantity of data.[^reuters-reddit][^tc-reddit] A comparison is meaningful only if unit, term, exclusivity and scope are aligned.

**Large-owner deals do not transfer to other data.** The sums in Table 2 relate to well-known corpora with large audiences and strong rights. The only public ranges for ordinary company data are on buyer program pages and are indicative, and the only reported transactions for such data are small sales by failed companies in the low five figures per company and a single $10 million auction for a very large archive.[^micro1-data][^gizmodo][^time-spirit]

**Replacement cost plus expected value seems to be the working logic.** The per-task prices for environments, the research-thesis valuations of startup archives and the accounting-style approaches all reason from what it would cost to produce comparable material and how much training value it could yield.[^epoch-rl][^mit-dead][^deloitte] Buyers who can commission experts to write examples have a ceiling set by that cost, and data that cannot be reproduced cheaply has a higher one. This is an interpretation of the sources and not a stated rule.

**Academic valuation and market practice are far apart.** Shapley-based methods can say which examples help a model but require access to the model and data, and they are expensive. Public deals are negotiated on volume, uniqueness, rights and the buyer's thesis.[^dshapley][^jia] The data market literature identifies verification before purchase as a core difficulty, which matches the use of samples, due diligence and valuation after sampling on program pages.[^agarwal][^lazarus][^handshake]

**Cost shifting matters as much as price.** Programs state that they cover extraction and anonymization, and the Spirit sale put the cost of de-identification on the buyer.[^mercor-data][^tnw-spirit] A seller comparing offers would need to compare who bears these costs, who holds liability for residual privacy risk and what use limits apply, which are legal questions beyond this review.

## 5. Limitations

This is a narrative review of public sources. Most large-deal figures come from press reports citing anonymous sources, and some are secondary reports of filings. We could not verify them against the contracts, and parties may have reasons to leak favorable numbers. Program pages are marketing and do not show how many companies are paid. Our reading of two SEC filings relied on automated text extraction of very long documents, and we did not read every note to the financial statements, so we make no claim about how Shutterstock recognizes AI licensing revenue. Several academic papers are preprints. We did not find public sources for deals in several categories, including customer support, CRM and marketing performance data. Currency amounts are not inflation adjusted, and prices are time-sensitive in a market that is changing quickly.

## 6. Conclusion

Public information about the price of AI training data is a patchwork. Fixed annual fees, archive payments with recurring components, per-unit rates, tiered ranges, per-task prices, annual dataset licenses and auctions all appear, and exclusivity is reported to multiply prices by around four to five in one segment. The best-documented figures belong to large content owners and a few media types. For ordinary business data, public evidence is limited to indicative program ranges and small reported sales by failed companies, plus one large auction. Academic valuation methods describe how a dataset's contribution to a model could be measured, but we found no public deal that used them. A reader should treat each figure as a statement about a particular unit, term and set of rights, and should not treat any of them as a benchmark for a different kind of data.

[^engadget-dotdash]: [OpenAI will pay DotDash Meredith at least $16 million per year to license its content](https://www.engadget.com/ai/openai-will-pay-dotdash-meredith-at-least-16-million-per-year-to-license-its-content-191756662.html) (Engadget, citing Adweek, November 2024)
[^tc-reddit]: [Reddit says it's made $203M so far licensing its data](https://techcrunch.com/2024/02/22/reddit-says-its-made-203m-so-far-licensing-its-data) (TechCrunch, citing Reddit's IPO filing, February 2024)
[^reuters-reddit]: [Exclusive: Reddit in AI content licensing deal with Google](https://tech.yahoo.com/articles/exclusive-reddit-ai-content-licensing-041012503.html) (Reuters, via Yahoo, February 2024)
[^decoder-axel]: [Axel Springer and OpenAI license agreement is worth tens of millions of euros per year](https://the-decoder.com/axel-springer-and-openai-license-agreement-is-worth-tens-of-millions-of-euros-per-year/) (The Decoder, citing the Financial Times, December 2023)
[^ii-informa]: [ii view: confident Informa does AI data deal with Microsoft](https://ii.co.uk/analysis-commentary/ii-view-confident-informa-does-ai-data-deal-microsoft-ii531624) (interactive investor, May 2024)
[^decoder-video]: [OpenAI and Google are buying YouTubers' unpublished videos for up to $4 per minute](https://the-decoder.com/openai-and-google-are-buying-youtubers-unpublished-videos-for-up-to-4-per-minute/) (The Decoder, citing Bloomberg, January 2025)
[^decoder-spree]: [Tech giants are on a billion-dollar shopping spree for AI training data](https://the-decoder.com/tech-giants-are-on-a-billion-dollar-shopping-spree-for-ai-training-data/) (The Decoder, citing Reuters, April 2024)
[^authorsguild]: [HarperCollins AI licensing deal](https://authorsguild.org/news/harpercollins-ai-licensing-deal/) (Authors Guild, November 2024)
[^bostonglobe]: [Anthropic authors lawsuit settlement](https://www.bostonglobe.com/2025/09/05/business/anthropic-authors-lawsuit-settlement-chatbot) (Boston Globe, September 2025)
[^vb-newscorp]: [OpenAI partners with Wall Street Journal publisher News Corp](https://venturebeat.com/ai/openai-partners-with-wall-street-journal-publisher-news-corp) (VentureBeat, citing the Wall Street Journal, May 2024)
[^infodocket-wiley]: [Report: Wiley set to earn $44M from AI rights deals, confirms no opt-out for authors](https://www.infodocket.com/2024/08/30/report-wiley-set-to-earn-44m-from-ai-rights-deals-confirms-no-opt-out-for-authors/) (infoDOCKET, citing The Bookseller, August 2024)
[^pbs-ap]: [ChatGPT creator OpenAI signs deal with AP to license news stories](https://www.pbs.org/newshour/economy/chatgpt-creator-openai-signs-deal-with-ap-to-license-news-stories) (PBS NewsHour, Associated Press, July 2023)
[^so-openai]: [Stack Overflow and OpenAI partnership announcement](https://stackoverflow.co/company/press/archive/openai-partnership) (Stack Overflow, May 2024)
[^sstk24]: [Shutterstock annual report (Form 10-K) for fiscal 2024](https://www.sec.gov/Archives/edgar/data/1549346/000154934625000011/sstk-20241231.htm) (Shutterstock, SEC filing, 2025)
[^sstk23]: [Shutterstock annual report (Form 10-K) for fiscal 2023](https://www.sec.gov/Archives/edgar/data/1549346/000154934624000007/sstk-20231231.htm) (Shutterstock, SEC filing, 2024)
[^troveo]: [Troveo](https://www.troveo.ai/) (Troveo, accessed October 2026)
[^micro1-data]: [Turn existing workflows into revenue](https://www.micro1.ai/data) (micro1, accessed October 2026)
[^tnw-micro1]: [micro1 commits $1bn to buy company data for training AI agents](https://thenextweb.com/news/micro1-1bn-enterprise-data-ai-training-citi-hercules) (The Next Web, October 2026)
[^scale-dp]: [Scale AI data partnership](https://scale.com/data-partnership) (Scale AI, accessed October 2026)
[^handshake]: [Handshake AI data licensing: get started](https://joinhandshake.com/ai/data-partnerships/license/) (Handshake, accessed October 2026)
[^lazarus]: [Project Lazarus](https://lazarus.turing.com/) (Turing, accessed October 2026)
[^mercor-data]: [Mercor Data: Monetize your data](https://www.mercor.com/data/) (Mercor, accessed October 2026)
[^gizmodo]: [Failed companies are selling old Slack chats and email archives to train AI](https://gizmodo.com/failed-companies-are-selling-old-slack-chats-and-email-archives-to-train-ai-2000747916) (Gizmodo, citing Forbes, April 2026)
[^mit-dead]: [AI labs are buying data from dead startups to train models](https://www.technologyreview.com/2026/10/13/1145755/ai-labs-buying-data-from-dead-startups-to-train-models) (MIT Technology Review, October 2026)
[^time-spirit]: [Google's $10 Million Bid for Spirit Airlines' Data Reveals AI's Next Frontier](https://time.com/article/2026/08/25/google-spirit-airlines-ai-data-RL/) (Time, August 2026)
[^tnw-spirit]: [Google picked and paid for the firm anonymising the Spirit Airlines data](https://thenextweb.com/news/google-spirit-airlines-data-10m-bankruptcy-auction-mercor) (The Next Web, August 2026)
[^cso]: [Bidding war for defunct Spirit Airlines' employee data will not die](https://www.csoonline.com/article/4218769/bidding-war-for-defunct-spirit-airlines-employee-data-will-not-die.html) (CSO Online, September 2026)
[^epoch-rl]: [Gradient Updates: the state of RL environments](https://epoch.ai/gradient-updates/state-of-rl-envs) (Epoch AI, January 2026)
[^tc-env]: [Silicon Valley bets big on environments to train AI agents](https://techcrunch.com/2025/09/16/silicon-valley-bets-big-on-environments-to-train-ai-agents/) (TechCrunch, September 2025)
[^aimi]: [Stanford AIMI datasets for commercial use](https://aimi.stanford.edu/datasets-commercial-use) (Stanford AIMI, accessed October 2026)
[^snowflake]: [Snowflake Marketplace documentation](https://docs.snowflake.com/en/user-guide/data-marketplace) (Snowflake, accessed October 2026)
[^awsdx]: [AWS Data Exchange](https://aws.amazon.com/data-exchange/) (Amazon Web Services, accessed October 2026)
[^extruct]: [AI training data marketplaces](https://www.extruct.ai/data-room/ai-training-data-marketplaces/) (Extruct, June 2026)
[^hhq]: [Licensing of data for AI model training: things to take note of](https://hhq.com.my/posts/licensing-of-data-for-ai-model-training-things-to-take-note-of/) (Halim Hong & Quek, February 2024)
[^termslaw]: [AI training data licensing: a usable agreement](https://terms.law/insights/ai-training-data-licensing-usable-agreement.html) (terms.law, undated)
[^dshapley]: [Data Shapley: Equitable Valuation of Data for Machine Learning](https://arxiv.org/abs/1904.02868) (Ghorbani and Zou, arXiv preprint, April 2019)
[^jia]: [Towards Efficient Data Valuation Based on the Shapley Value](https://arxiv.org/abs/1902.10275) (Jia et al., arXiv preprint, February 2019)
[^dvrl]: [Data Valuation using Reinforcement Learning](https://arxiv.org/abs/1909.11671) (Yoon, Arik and Pfister, arXiv preprint, September 2019)
[^koh]: [Understanding Black-box Predictions via Influence Functions](https://arxiv.org/abs/1703.04730) (Koh and Liang, arXiv preprint, March 2017)
[^agarwal]: [A Marketplace for Data: An Algorithmic Solution](https://arxiv.org/abs/1805.08125) (Agarwal, Dahleh and Sarkar, arXiv preprint, May 2018)
[^fernandez]: [Data Market Platforms: Trading Data Assets to Solve Data Problems](https://arxiv.org/abs/2002.01047) (Castro Fernandez, Subramaniam and Franklin, arXiv preprint, February 2020)
[^pei]: [A Survey on Data Pricing: from Economics to Data Science](https://arxiv.org/abs/2009.04462) (Pei, arXiv preprint, September 2020)
[^deloitte]: [Valuing data assets](https://www.deloitte.com/us/en/insights/topics/digital-transformation/valuing-data-assets.html) (Deloitte, December 2025)

## FAQ

**Is there a standard price for AI training data?**

No. Public figures come in different units, such as per image, per minute of video, per book, per task, per year or per deal, and many come from anonymous sources. One data intermediary states that there is no rate card for company data.

**What are the most common pricing structures?**

In the sources reviewed, the most common are a fixed annual fee, a one-time payment for an archive combined with recurring fees for current content, per-unit rates, and tiered payout ranges that depend on the quality and uniqueness of the data.

**Do exclusive licenses cost more?**

One review of reinforcement learning environments reports that exclusive deals cost roughly four to five times as much as non-exclusive ones, according to two founders it interviewed. We found no comparable published multiple in other segments.

**Are the large publisher and Reddit figures relevant to a normal company?**

They show that large owners of distinctive content have been paid sums in the tens of millions of dollars per year, according to filings and reports. They are not benchmarks for other kinds of data, which differ in scale, uniqueness and rights.

**Do data valuation methods from research set prices?**

We found no public deal in which a published valuation method such as the Shapley value set the price. Those methods are used in research to attribute a model's performance to training examples and to design hypothetical data markets.