The State of Business Data Licensing for AI, 2026

1. Introduction#

Developers of large language models and related systems have historically relied heavily on data that was cheap to obtain, such as web crawls and other openly available text.¹² A growing body of reporting and research describes a shift toward data that is neither public nor free: operational records, communications, documents, code, specialist imagery and other material held by organizations. This paper reviews what is publicly documented about that market as of October 2026, and it is deliberately limited to what can be verified from published sources.

For the purposes of this paper, "business data licensing for AI" means the licensing or sale of data that an organization controls and that is not freely available on the open web, to a party that intends to use it to train, fine-tune, evaluate or build environments for AI systems. The definition covers several kinds of holder: platforms with user-generated content, publishers, stock-media libraries, health systems, creators of video, and operating companies whose internal records have value to a model developer. It excludes the trade in labeled datasets produced by contracted annotators, except where that trade overlaps with the sources of the data.

Three questions organize the review. First, what is driving demand, and how robust is the evidence for that demand? Second, who is supplying data, through which intermediaries, and at what reported prices? Third, what legal and regulatory conditions govern a transaction, and how are they changing? A companion review covers individual reported deals in a deal-by-deal table; the present paper summarizes deals only at a high level. A deal-by-deal table appears in the companion review.

The market is difficult to study for three reasons. In the sources reviewed, contract terms are rarely disclosed. Prices, where they appear, are usually relayed by journalists from unnamed sources or marketed by interested parties. And the legal position of AI training is being decided in courts while transactions take place. This paper therefore emphasizes the provenance of each claim and marks where sources conflict.

2. Methods and sources#

This is a narrative review of public documents, not a systematic review and not a primary study. No interviews, surveys or proprietary transaction data were used. The review period runs from 2022 to October 2026, with most material falling between 2023 and 2026.

Sources were gathered by reading the original page, filing or paper for each cited item, and were classified into four tiers. The first tier contains primary legal and regulatory sources: statutes and regulations, regulator guidance, and court-related documents. The second contains primary company and institutional sources such as filings, press releases, product documentation and university program pages. The third contains peer-reviewed or preprint research; preprints are labeled as such in the reference list because they have not necessarily been peer reviewed. The fourth contains reputable news reporting and law-firm commentary, and vendor or marketing material. Vendor material was used only to document what a vendor claims, never as independent evidence of market facts.

Inclusion required that a source be publicly accessible and that the specific claim cited could be located in the text that was read. Where a figure appears only in a secondary article that cites an unnamed source, the text says so, and the figure is attributed to the secondary publication and to the original outlet it names. Where two sources disagree, both are reported. Quantities are given in the currency and units used by the source, and no figures have been converted, extrapolated or estimated by the authors.

Reported prices are not comparable across data types because the unit differs, and several informative accounts are recent journalism or practitioner commentary that has not been independently verified; Section 5 returns to these points.

3. Findings#

3.1 Demand drivers#

Data scarcity arguments. The most-cited quantitative argument for rising data demand is an analysis by Epoch AI, which estimated the stock of public human-generated text at on the order of 300 trillion tokens, with a 90 percent confidence interval of 100 trillion to 1,000 trillion, and projected that the stock would be fully used at some point between 2026 and 2032 on an 80 percent interval.³ The same source is internally inconsistent on one detail: a figure caption gives a 2025 to 2030 range that differs from the headline range, and the authors note that the estimate covers only public text and excludes private messages, images, video and synthetic data.³ The estimate is therefore best read as a statement about public text, not about the total quantity of data that could be used for training. Its relevance to licensing is indirect: it motivates the search for non-public sources but does not show that developers are paying for them.

Shrinking access to the open web. A preprint audit by the Data Provenance Initiative examined consent signals across 14,000 web domains underlying common training corpora and reported that restrictions rose quickly between 2023 and 2024. The authors found that roughly 5 percent or more of all tokens in one widely used corpus (C4) became fully restricted, rising to more than 28 percent among its most actively maintained and critical sources, and that terms of service restricted crawling for about 45 percent of that corpus.² The authors also found that terms of service often conflicted with robots.txt files. If these trends persist, the open-web baseline becomes less complete, which strengthens the position of holders of data behind access controls. At the same time, public web data remains available on a large scale: Common Crawl describes a free and open repository that has operated since 2007 and holds over 300 billion pages, with 3 to 5 billion pages added each month.¹

Quality over quantity. Several research results suggest that smaller amounts of well-chosen data can matter as much as raw volume, which is the economic basis for paying for specific data. A preprint on phi-1 reported that a 1.3 billion parameter model trained on roughly 7 billion tokens of "textbook quality" data reached 50.6 percent pass@1 on HumanEval.⁴ A preprint on LIMA reported strong outputs from a 65 billion parameter model fine-tuned on only 1,000 curated examples.⁵ The DataComp-LM preprint reported that model-based filtering is a key step in assembling high-quality training sets.⁶ These results concern public or synthetic data, not private business data, so they support only a general inference: curation can be valuable, and it depends on access to suitable source material.

Concern about synthetic data. A widely cited preprint argued that training on model-generated content causes "irreversible defects" in resulting models, and that data about genuine human interactions would become increasingly valuable as model output spreads online.⁷ The finding concerns recursive training in specific model families and does not establish that synthetic data is unusable; some commercially distributed datasets are explicitly synthetic, as the dataset card for one popular customer-support example states.⁸ The evidence supports only a cautious claim: real human-generated data is one input among several, and its marginal value depends on the task.

Earlier signs of demand for private data. In November 2023, OpenAI announced a Data Partnerships program, stating that it sought large-scale datasets that reflect human society and are not easily accessible online, particularly data that expresses human intention such as long-form writing or conversations, in both open-source and private forms.⁹ It is a clear statement of demand from one developer, but it reports no prices or volumes.

Expert work products and workplace tasks. OpenAI's GDPval preprint introduced a benchmark covering 44 occupations drawn from the top nine sectors contributing to U.S. GDP, with tasks authored by professionals averaging 14 years of experience.¹⁰ Such benchmarks imply a need for realistic workplace material, although the paper does not describe purchases of business data. A TechCrunch report on remarks by Mercor's chief executive described AI labs hiring former employees of banks, consultancies and law firms as an alternative to negotiating with companies for their data; it also notes that some job postings sought people who could give access to company data such as a production codebase, and that Mercor said a few chief technology officers had accepted.¹¹ These claims come largely from the company's own representatives and were not independently verified in the article.

Reinforcement learning environments. The newest demand category concerns simulated workspaces in which agents practice multistep tasks. TechCrunch reported in September 2025 that major labs build such environments in house, that data-labeling firms and startups are expanding into the area, and, citing The Information, that Anthropic leaders had discussed spending more than $1 billion on environments over a year; the article characterizes this as discussion, not a confirmed commitment.¹² An Epoch AI analysis based on interviews with 18 practitioners reported that reinforcement learning is the primary use, that enterprise workflows such as expense reports and CRM updates are a major growth area, and that quality control of contractors is the core bottleneck; it also notes that several of its figures come from unverified third parties.¹³ These accounts indicate that demand for workplace-shaped material exists, but evidence on how much of it is met by licensed business data, as opposed to built environments and hired experts, is thin.

3.2 Supply side#

Supply of non-public data for AI comes from several kinds of holders, whose circumstances and constraints differ. The following overview is organized by holder type.

Platforms with user-generated content. Reddit is the best-documented case because its IPO prospectus disclosed data-licensing contracts. According to TechCrunch's reading of the prospectus, Reddit entered arrangements in January 2024 with an aggregate contract value of $203.0 million, running two to three years, with at least $66.4 million expected to be recognized in 2024.¹⁴ The article does not name the counterparties.¹⁴ Reddit's "other revenue" category, which includes data licensing, was reported at $35 million in the second quarter of 2025, up 24 percent from $28.1 million a year earlier.¹⁵ That category may include items other than licensing, so it should not be read as licensing revenue alone. Stack Overflow announced an API partnership with OpenAI in May 2024 covering access to its data through OverflowAPI, and its announcement gave no financial terms, exclusivity provisions or duration.¹⁶

Publishers and rights holders. Publishers have signed multiple agreements, and their financial statements provide some of the clearest data. Wiley was reported to expect $44 million from AI partnerships, of which $23 million had already been earned, and not to offer authors an opt-out, although a spokesperson said authors would be paid according to their contractual terms.¹⁷ Created by Humans, a start-up, was reported to have raised $5 million to operate a marketplace where authors license work to AI developers directly, choosing which rights to license.¹⁸ The U.S. Copyright Office observed that direct licensing is most common and most promising with respect to corporate entities with catalogs of high-quality, easily identifiable content, such as large stock-photography companies, national news outlets, major record companies and film studios.¹⁹ This observation is useful for understanding who has been able to sell so far: concentrated rights holders with clear ownership.

Stock media. Shutterstock's 2024 annual report states that the company has seen increased demand for access to its metadata for machine learning and generative AI training, and that customers may use its library through a data offering to train models.²⁰ The filing reports revenue of $175.3 million in 2024 for its "Data, Distribution and Services" category, compared with $137.3 million in 2023, but it does not break out AI data licensing separately, so the increase cannot be attributed to AI alone.²⁰

Health systems and research institutions. Stanford's AIMI Shared Data program, which previously offered de-identified clinical data for non-commercial use only, now licenses datasets to commercial entities at $70,000 per dataset for one year of access (for agreements contracted in fiscal year 2025), with no negotiated discounts, and with applications reviewed by a committee against criteria such as scientific validity and the need for the data.²¹ Protege, an intermediary profiled in June 2026, says it connects healthcare and media data holders with AI developers and has raised a total of $65 million, according to the interview-based article; the same article reports company claims, which are not independently verified, that it helped one provider generate seven-figure new licensing revenue within a year.²²

Video and creator content. Bloomberg reporting relayed by The Decoder in January 2025 stated that OpenAI, Google and other AI companies were paying creators for unpublished video, with intermediaries Troveo AI and Calliope Networks handling licensing.²³ Troveo's own website describes a network of licensors and lists scale figures such as more than 8 million hours of video, which are self-reported marketing claims.²⁴

Defunct and failing companies. The most direct evidence of operating-company data entering the market comes from 2026 reporting. Gizmodo, summarizing Forbes, reported that shutting-down startups were selling internal emails, chat messages, source code and documents, and that a wind-down service called SimpleClosure had launched an "Asset Hub" to license source code and workplace data after removing personally identifiable information.²⁵ MIT Technology Review reported in October 2026 that buyers include foundation labs, vendors building simulated workspaces and mature technology companies, that Spirit Airlines sold data during bankruptcy (reported at $10 million for 100 million emails, 500 million Teams messages and 176,000 employee records) after Google outbid Mercor, and that SimpleClosure had completed more than 150 Asset Hub deals.²⁶ The same article reports that no company has published tests comparing real-world data with synthetic alternatives, that an attorney called such data "legally radioactive," and that proceeds typically go to creditors and investors, not employees.²⁶ These accounts are journalistic and rely on named but interested sources, so they indicate that a market exists without establishing its size.

Constraints on supply. Contracts, privacy promises and employee or customer data limit what can be sold (Section 3.5). Ownership may be diffuse, a point the Copyright Office raised for content created outside professional creative industries.¹⁹ And preparation is burdensome: Stanford's program requires an application, committee review and a signed agreement before access begins.²¹

3.3 Intermediaries and marketplaces#

A distinct feature of the market is the number of intermediaries that sit between holders and developers. Table 1 summarizes categories documented in the sources reviewed. The classification is the authors' own organizing device, and individual firms often span several categories.

Table 1. Categories of intermediaries and enabling infrastructure, with examples named in the sources reviewed.

Expert-labor marketplaces. The TechCrunch account of Mercor reports that its customers include OpenAI, Anthropic and Meta, that it said annualized recurring revenue was roughly $500 million, that most revenue comes from a few labs, and that it benefited when many labs stopped working with Scale AI following Meta's investment.¹¹ These are company statements and reporting at a conference, and they should not be treated as audited figures. The pattern of concentrated buyers is nonetheless notable for anyone assessing counterparty risk.

Licensing platforms. Troveo's website describes licensing of video, audio and gaming data and, separately, a program under which companies can license their own business data such as email, chat, CRM records, tickets and files for recurring revenue; it lists more than 7,000 data providers and 350 companies, which are self-reported.²⁴ A vendor-authored article by Extruct, dated June 2026, names 15 AI-training-data marketplaces, classifies nine as "collect and license," and reports disclosed funding for 12 of the 15 totaling about $178 million.³² That article is promotional, and its categories are the vendor's own, but it indicates the number of venture-backed entrants. Neither source reports transaction volume or price.

Cloud marketplaces. Snowflake's documentation describes free, limited-trial and paid listings, with paid listings priced by provider-set plans and offers.²⁷ AWS states that its Data Exchange offers more than 3,500 third-party datasets from more than 300 providers.²⁸ Neither page says how many listings are intended for AI training, so these venues are available infrastructure, not evidence of an AI data market.

Access-control and payment layers. Cloudflare announced "pay per crawl" in private beta in July 2025: publishers set a flat per-request price and choose to allow, charge or block each crawler, with Cloudflare acting as merchant of record.²⁹ The RSL 1.0 specification, reported in December 2025, adds machine-readable licensing terms to the robots.txt model; The Register noted that, like robots.txt, it is not a technical access control, and that one payment layer was in beta with no bots yet being billed.³⁰ These mechanisms address web content more than private business records, but they show per-use licensing infrastructure being built.

Provenance standards. The Data & Trust Alliance published Data Provenance Standards version 1.0.0 in July 2024, defining 22 metadata fields across source, provenance and use groups, including consent documentation location, license to use and intended use; adoption levels are not reported in the source.³¹

3.4 Reported prices and pricing signals#

What prices are reported. Public price evidence is fragmentary and heterogeneous. Table 2 collects figures from the sources reviewed, with their units and attribution. Entries should be read as reported statements, not as market rates, and several are from 2024 or 2025 and may have changed.

Table 2. Reported price points and deal values in AI data licensing, by type, as of October 2026 (reported figures, not verified rates).

Unit heterogeneity. The units in Table 2 do not reconcile: per minute, per image, per task, per company, per dataset-year, per contract. Even within a category the range is wide. The 2024 Reuters-derived figures include ranges from a fraction of a cent to several dollars per image, depending on the seller and the negotiating context.³³ The Epoch analysis of reinforcement learning environments reports that exclusive deals were estimated by two founders at roughly four to five times the cost of non-exclusive ones, which is one of the few direct statements about how contract terms affect price.¹³ Several prices in the table come from named sellers or buyers with an interest in the number; for example, the per-image ranges attribute figures to Defined.ai, a data vendor.³³

Valuation methods. Deloitte describes three approaches to valuing data assets (income, cost and market), and states that the market approach, which relies on prices paid for comparable assets, is conceptually valid but rarely used alone, partly because reliable transaction data is scarce and many deals do not disclose pricing drivers.³⁶ The lack of comparables is itself a finding: sellers cannot easily benchmark.

Market-size forecasts. Commercial research products report very different sizes for "AI training datasets," with scopes that are not visible in the summaries reviewed. Grand View Research, as distributed by Global Information in March 2026, estimates the market at about $3.2 billion in 2025 and $16.3 billion by 2033, a compound annual growth rate of 22.6 percent over 2026 to 2033.³⁷ Verified Market Research reported $1.56 billion in 2023 rising to $7.56 billion in 2031 at 21.86 percent.³⁸ Technavio forecasts an increase of $9.12 billion over 2026 to 2030 at a 28.9 percent growth rate (its page gives both 2025 to 2030 and 2026 to 2030 as the window), covering text, image or video, and audio service types.³⁹ These cannot be directly compared because base years, end years and definitions differ, and they should not be read as measuring licensed business data specifically; the segment is likely a fraction of whatever each report counts, but none of the summaries reviewed provides that share.

Signals that matter more than headline prices. Four signals recur. Exclusivity appears to command a premium.¹³ Data with clear rights and curation appears to be favored over ambiguous material, consistent with the Copyright Office's observation that centralized, high-quality catalogs are easier to license.¹⁹ Intermediaries such as SimpleClosure and Protege bundle preparation and de-identification into the offer.²⁵²² And buyers can substitute hired experts, built environments or synthetic data, which limits pricing power for sellers of generic material.¹¹¹²

3.5 Legal and regulatory environment#

Copyright and training. For content that is protected by copyright, the legal status of training is unsettled and is being clarified through litigation. In May 2025, the U.S. Copyright Office published a pre-publication report that said licensing markets are still developing, that voluntary licensing is already happening in some sectors, and that where licensing options exist or are likely to be feasible, this consideration will disfavor fair use under the fourth factor; it also said that if barriers to licensing prove insurmountable for some works, there may be no functioning market to harm.¹⁹ The report is not binding on courts.

Court decisions have so far pointed in different directions on different facts. In June 2025, in Bartz v. Anthropic, Judge Alsup was reported to have found that training chatbots on copyrighted books was not illegal but that acquiring millions of books through pirate sites was wrongful; the case then settled for $1.5 billion, or about $3,000 for each of an estimated 500,000 books.⁴⁰ A commentary on the Meta ruling places the Alsup decision two days earlier.⁴¹ In Kadrey v. Meta, Judge Chhabria granted summary judgment to Meta on June 25, 2025, but an advocacy-group summary of the opinion stresses that the result rested on the plaintiffs' failure to show market harm, and that the judge described market dilution as a potentially significant theory.⁴¹ In Thomson Reuters v. Ross, the District of Delaware held in February 2025 that use of Westlaw headnotes to train a non-generative legal search tool was not fair use, and the judge limited the holding to non-generative AI.⁴² The Third Circuit affirmed on September 29, 2026, in what was reported as the first federal appellate decision on fair use in AI training; as of the report reviewed, the opinion was sealed and its reasoning not public, so its scope is unclear.⁴³ The practical implication for this paper's subject is limited but real: courts and regulators are treating the existence of a licensing market as relevant to the fair use analysis, which is a reason developers may prefer documented licenses for content with clear owners.

Most of these cases concern published creative or editorial works. Internal business records such as emails, tickets and code involve different issues, including contract, confidentiality, trade secret and privacy law, and the sources reviewed here do not report a court decision on training with licensed or unlicensed business records.

Contract and terms of service. Reddit sued Anthropic in June 2025 in San Francisco Superior Court on theories that include breach of contract and unfair competition, alleging that Anthropic accessed content in violation of Reddit's user agreement and that bots continued to access Reddit after it was said to be blocked; Reddit seeks damages for lost licensing revenue and deletion of models containing Reddit content.⁴⁴ The allegations had not been adjudicated in the sources reviewed.

Consumer protection and promises. The U.S. Federal Trade Commission has said in staff guidance that quietly changing terms of service or privacy policies to allow AI training may be unfair or deceptive, and that a company that collected data under one set of privacy promises cannot drop those promises on its own.⁴⁵ Separate FTC guidance to companies offering model-as-a-service products stated that there is "no AI exemption from the laws on the books," that companies must honor commitments however they were made, and that in past cases the agency has required deletion of models built on unlawfully obtained data.⁴⁶ For a business that considers licensing its data, the relevant implication is that prior privacy promises to customers and users constrain secondary uses, and that a remedy can reach the trained model, not just the data.

Bankruptcy. Where a debtor's privacy policy prohibits transferring personally identifiable information to unaffiliated parties, 11 U.S.C. section 363(b)(1) bars the trustee from selling it unless the sale is consistent with the policy or a court approves after the appointment of a consumer privacy ombudsman and a hearing.⁴⁷ In March 2025 the FTC chairman wrote to the U.S. Trustee about the 23andMe bankruptcy, stating that consumers should be able to trust that companies will keep their promises about sensitive information.⁴⁸ These provisions are relevant to the wind-down market described in Section 3.2: the MIT Technology Review account reports union privacy objections to the Spirit sale and an academic's statement that there is no federal law banning the sale of workers' data, which is the academic's characterization.²⁶

Privacy law. In the European Union, the European Data Protection Board's Opinion 28/2024 stated that AI models trained on personal data cannot in all cases be considered anonymous, and that anonymity requires both extraction of personal data and its retrieval through queries to be insignificant in likelihood, assessed case by case; it also set out a three-step legitimate interest test for model development and addressed the consequences of unlawful processing in the development phase.⁴⁹ In the United States, there is no single comprehensive federal privacy law, and the IAPP counted 19 states with comprehensive privacy laws as of its last update on October 27, 2025.⁵⁰ California's statute defines "sell" to include disclosing personal information to a third party for monetary or other valuable consideration, and separately defines "deidentified" information as information that cannot reasonably be used to infer information about, or be linked to, a particular consumer, subject to the business taking reasonable measures to prevent association.⁵¹ For health information, the HIPAA rule at 45 CFR 164.514 recognizes two methods of de-identification: expert determination, and removal of 18 categories of identifiers together with an absence of actual knowledge that the remaining information could identify an individual.⁵² Whether a particular de-identification method suffices for AI training is fact specific and outside the scope of this paper.

Transparency obligations for developers. The European Commission's AI Office published a template on July 24, 2025 for the public summary of training content that providers of general-purpose AI models must publish under Article 53(1)(d) of the AI Act. According to a law-firm summary, the summary must state whether licensed, scraped, user-generated and synthetic data were used; new models had to comply from August 2, 2025, existing models have until August 2, 2027, and from August 2, 2026 the AI Office may verify compliance.⁵³

3.6 Trends#

The evidence reviewed supports the following trends, listed with the strength of support for each.

From text corpora toward workplace-shaped data. Demand statements and benchmarks increasingly concern professional work, agent environments and enterprise workflows. The OpenAI Data Partnerships announcement (2023), the GDPval benchmark (2025), the reinforcement learning environment reporting (2025 to 2026) and the wind-down data accounts (2026) form a consistent sequence.^910121326 Support is moderate, since most items are reporting or practitioner commentary, and quantitative evidence on spending is limited to a reported discussion of more than $1 billion at one lab.¹²

Substitution among supply channels. Hired experts, built environments, synthetic data and licensed archives are partial substitutes. Mercor's description of hiring former employees instead of negotiating with companies, Epoch's report of "substantially more in-housing" by labs, and the MIT observation that no one has published head-to-head tests of real versus synthetic data all point to an unsettled mix.¹¹¹³²⁶ Support is moderate and mostly qualitative.

More structured, per-use access on the open web. Pay-per-crawl, machine-readable licensing and default blocking show a trend toward priced access to web content.²⁹³⁰ The Register's note that no bots were yet being billed on one RSL payment layer indicates the early state.³⁰ Support for adoption is weak; support that infrastructure is being built is strong.

Increasing disclosure and provenance expectations. The EU template, the Data & Trust Alliance standards and the FTC's statements on promises all raise the cost of undocumented provenance, though the sources do not measure the effect on pricing.⁵³³¹⁴⁶

Rights holders enforcing and licensing in parallel. Licensing revenue (Reddit, Wiley) coexists with litigation (Reddit v. Anthropic, Bartz, Kadrey, Ross), which suggests both strategies are used at once.^141744404142

Concentration among buyers. Mercor's reported dependence on a few labs and the reported auction between Google and Mercor for Spirit's data point to a small number of large buyers.¹¹²⁶ This is an inference from anecdotes; no source quantifies buyer concentration for business data.

4. Discussion#

What the evidence supports. There is credible, multi-source evidence that a market exists for licensing non-public content and data to AI developers, and that it includes platform content, publisher archives, stock media, clinical datasets, creator video and, since 2025 or 2026, internal records of defunct companies. The best-documented transactions involve parties with centralized rights and a public company's disclosure obligations. For operating businesses that are not shutting down, the documented evidence is much thinner: the sources reviewed describe platforms that invite companies to license business records,²⁴ and labs that say they prefer hired experts to negotiating with companies,¹¹ but they do not document completed licensing transactions of an ordinary operating company's internal data, apart from the wind-down cases.

Why prices are hard to compare. Section 3.4 shows that reported prices vary by unit, data type and context, and that the few public comparables are tied to well-known publishers or unusual events such as bankruptcies, consistent with Deloitte's observation that comparable transactions are scarce.³⁶ A seller inferring a price from Table 2 would need to adjust for exclusivity, rights clarity, preparation and the buyer's alternatives, and the sources do not supply data for those adjustments.

Where evidence conflicts. At least four conflicts are visible. The Epoch data-exhaustion window is 2026 to 2032 in the headline and 2025 to 2030 in a figure caption.³ The Forbes-derived payout range for failed startups differs from the lead of the article that reports it.²⁵ Market-size estimates from different publishers differ widely and use different base years.³⁷³⁸ And commentary on the Meta ruling varies with the commentator's perspective.⁴¹ Any single number should be treated as provisional.

The role of legal risk. Legal conditions seem to push toward documented licenses for content with clear owners, as the Copyright Office's fourth-factor reasoning suggests,¹⁹ while privacy and consumer protection rules constrain what an operating company can license from its own customers, employees and users.⁴⁶⁴⁹⁵¹ A business may be a clear owner of non-personal operational data yet hold data about people that it is not free to sell. The wind-down accounts illustrate the tension.²⁶⁴⁷

Implications for holders of business data. Without giving advice, the evidence suggests that the following factors are salient to any assessment: rights and permissions attached to the data (including prior privacy promises), the presence of personal data and the de-identification method, the availability of substitutes to the buyer, whether the arrangement is exclusive, and who bears preparation costs. Each of these appears in at least one cited source as a driver of feasibility or price.^4552132125

Implications for researchers and policymakers. The absence of transaction-level disclosure limits analysis. Public company filings are the most reliable source but are infrequent, as the lack of a separate AI licensing line in the Shutterstock filing shows.²⁰ The EU training summary requirement may eventually provide standardized data on which sources were licensed, though it does not require pricing.⁵³

5. Limitations#

This review has several limitations that affect how its findings should be used.

First, it relies on public documents, and the market is largely confidential; absence of a reported transaction is not evidence that none occurred. Second, many sources are secondary: figures from Forbes, Bloomberg, Reuters, the Financial Times, the Wall Street Journal and The Information appear here through other outlets, and the original reports were not all read. Third, vendor websites, a vendor-authored marketplace article and market-research sales pages describe their own products or forecasts, and their self-reported scale numbers are unverified. Fourth, practitioner interview material such as the Epoch AI analysis rests on a small number of respondents.¹³

Fifth, the legal section summarizes secondary descriptions of some decisions and does not substitute for reading the opinions; the Third Circuit opinion in Thomson Reuters v. Ross was sealed at the time of the report reviewed, so its reasoning may differ from the characterizations here.⁴³ Some descriptions come from advocacy or law-firm commentary with their own perspectives.⁴¹⁴² State privacy laws other than California's, laws on employee data, and non-U.S., non-EU regimes are not examined.

Sixth, facts change rapidly. Figures and legal positions are given as of October 2026, and several have already shifted (for example, Epoch's estimate revised an earlier prediction of exhaustion by 2024).³ Seventh, the classification in Table 1 and the trend assessments in Section 3.6 are interpretive. Finally, nothing in this paper is legal, tax, financial or investment advice.

6. Conclusion#

The public record as of October 2026 shows a real but uneven market for non-public data licensed to AI developers. Demand drivers include constrained open-web access, the perceived value of quality and human-generated data, and a growing appetite for workplace-shaped material for agents and evaluations. Supply is best documented for platforms, publishers, stock media and clinical datasets, with newer reporting on data from closing companies. A layer of intermediaries has formed, and infrastructure for per-use licensing and provenance is under construction.

Prices are reported in incompatible units, are often relayed secondhand and span orders of magnitude, so no reliable benchmark for ordinary operating-company data can be drawn from the sources reviewed. The legal environment is moving on several fronts at once: copyright litigation, contract claims, privacy and consumer-protection obligations, bankruptcy rules and EU transparency requirements. Each shapes what can be licensed and on what terms.

Priorities for further evidence are: transaction-level disclosures (for example in public company filings), independent tests of the value of real versus synthetic business data, appellate and trial rulings that address internal business records, and consistent provenance metadata. Until such evidence exists, claims about the size of this market or the price of a given dataset should be treated as provisional.