AI data market
AI data market map of observed demand: 26 deals, no business records
By SourceX Editorial · Reviewed by Noah Loul ·
10 sections
Short answer
An AI data market map of observed demand should grade each data category by its strongest first-party evidence, not by AI spending. In SourceX's public tracker of 26 AI data deals, 21 cite a party's own announcement, 3 carry an amount a party disclosed, and none involves operational business records such as support, CRM or ERP histories. For that category the public evidence is stated buyer intent, so model it as an open question.
Value drivers
AI data market map observed
Key takeaways
- Observed demand means first-party evidence of a paid or agreed licence: a filing, a party's announcement or a tender; AI budgets and data volumes do not qualify.
- Of 26 tracked AI data deals, 21 cite a party's own announcement, but only 3 carry an amount a party disclosed itself.
- News text accounts for 14 of the 26 tracked deals yet none of the 3 party-disclosed amounts, so the category ranking depends on the evidence standard.
- No tracked deal involves operational business records; buyer calls and stated plans show intent, and confidentiality may hide completed deals.
- Seventeen of the 26 tracked deals name OpenAI, so the observed market is partly one buyer's pattern.
What should an AI data market map of observed demand show?
An AI data market map of observed demand should place each data category by the strongest public evidence that a buyer agreed to pay for licensed data in it, and keep every category without such evidence in a separate column of open questions. Observed demand here means evidence a reader can open: a seller's filing or results release, an agreement announced by a party to it, a public tender. A buyer's published call for data is weaker evidence of intent. AI budgets, data volumes and funding rounds are not evidence of licensing demand at all.
The thesis this page tests is narrow. Paid licensing is publicly visible today in published content, such as news, books, images and forum posts, while demand for operational business records, the support, CRM, ERP and workflow histories companies keep, is stated by buyers but not yet visible as completed, priced transactions. If that is right, a market model can count the first group as observed and must carry the second as a labeled scenario. If it is wrong, the evidence below says where to look.
The existing pages in this series answer neighbouring questions: how to size a market bottom-up from buyer budgets, how to read filings for data spending, how to separate potential from pipeline and how to lay out business models. None grades the categories themselves by evidence type, which is the job here. The page gives no market size.
Four evidence grades, and what each can carry in a model
Four grades cover the public record, and each supports a different amount of weight in a model. The first grade is an amount a party disclosed itself, in a filing or results release: it can anchor a floor for one seller in one year. The second is an agreement announced by a party to it, with no amount: it proves a transaction happened and says nothing about size. The third is a buyer's published call for data or stated plan: it shows intent and budget appetite, not spend. The fourth is a press report without a first-party source: useful for leads, unusable as an input.
A fifth state matters as much as the four: no public evidence found. It is not the same as no demand. Licensing contracts can be confidential, and SourceX's own how-it-works page says that neither side publicizes a SourceX deal, so a category built on confidential deals can stay invisible in every public tracker by design. The map below keeps that state explicit rather than filling it with an estimate.
Grading 26 tracked deals: where first-party evidence sits
Most tracked deals rest on first-party announcements, almost none carry an amount the parties disclosed, and none involves operational business records. The corpus is the 26 deal explainers in SourceX's public AI data deal tracker, read on 2026-10-28, each coded by data type, by whether it cites a source hosted by a party to the deal, and by where any money amount it states comes from.
Of the 26, 21 cite a party's own announcement and 5 cite only press coverage. Nine state a money amount: in 3 the amount comes from a party's own filing or disclosure, and in the other 6 only from press reports. Graded by its strongest evidence, each deal lands in one grade: 3 carry an amount a party disclosed, 20 rest on a party's announcement without one, and 3 on press reports alone. Seventeen of the 26 name OpenAI as a counterparty. The tracker describes itself as covering significant publicly reported deals; its selection rule is SourceX's, so these counts describe what has been reported and tracked, not every deal that exists. The table below gives the counts by data type.
| Data type | Tracked deals | Cite a party's own source | Amount from a party | Amount from press only |
|---|---|---|---|---|
| News and magazine text | 14 | 14 | 0 | 4 |
| Books and scholarly | 3 | 1 | 2 | 1 |
| Images, video and music | 3 | 2 | 0 | 1 |
| Forum and Q&A posts | 5 | 3 | 1 | 0 |
| Marketplace or intermediary | 1 | 1 | 0 | 0 |
| Operational business records | 0 | 0 | 0 | 0 |
| Total | 26 | 21 | 3 | 6 |
The map: six data categories, their strongest evidence and the open question
Five categories show at least one paid or announced transaction in first-party sources; operational business records show only stated intent. Counted once at its strongest grade, news text has 14 announced deals; books and scholarly 2 with a party amount and 1 press only; images, video and music 2 announced and 1 press only; forum and Q&A posts 1, 3 and 1; and the marketplace deal is announced. Each row below names the strongest first-party evidence read for this page or cited by the tracker, what that evidence does not measure, and the question an investor should carry forward.
Two rows rest on documents read directly. OpenAI's own page on its Axel Springer partnership says the collaboration involves the use of Axel Springer content for advancing the training of OpenAI's large language models. Informa's 2024 full-year results report that AI partnerships brought Taylor & Francis more than $75 million of revenue in 2024, which Informa describes as non-recurring data access. For business records, OpenAI's data partnerships page invites private datasets that are not easily accessible online while saying it is not seeking datasets with sensitive or personal information or information belonging to a third party, and micro1's blog announces a commitment to spend $1 billion over 12 months acquiring and licensing enterprise operational data for reinforcement learning environments. That is a stated plan, not money spent, and it does not show that any particular company's data will be bought.
| Category | Strongest first-party evidence | What it does not measure | Open question |
|---|---|---|---|
| News and magazine text | 14 tracked deals, all citing a party's announcement; OpenAI's Axel Springer page names training use | Price: every amount the tracker gives comes from press reports | Do archive deals renew, and on what terms? |
| Books and scholarly | Informa: more than $75 million of AI partnership revenue at Taylor & Francis in 2024 | Repeat demand: Informa calls it non-recurring | Is this a one-time archive sale per buyer? |
| Images, video and music | 2 of 3 tracked deals cite a party's own release | Any amount disclosed by a party | How much is training licence versus product deal? |
| Forum and Q&A posts | 3 of 5 cite a party's own page; the tracker attributes 1 amount to a filing | Whether API access is a training licence | Does paid API access replace licensing? |
| Marketplace or intermediary | 1 acquisition announced by the buyer | Volume cleared through the venue | Do venues clear deals or only list supply? |
| Operational business records | Stated intent only: OpenAI's private-datasets call; micro1's stated 12-month plan | Any completed, priced transaction | Will confidential deals surface in public evidence? |
Do AI adoption and a shrinking public text stock prove licensing demand?
No. Both measure something real, and neither measures what a licensing market model needs. The strongest alternative explanation for rising data demand is that broad AI adoption, plus a finite stock of public text, must translate into purchases of licensed data. Each step of that chain needs its own evidence.
The U.S. Census Bureau reported in November 2023 that only about 4 percent of businesses used AI to produce goods and services, from its Business Trends and Outlook Survey. The figure measures how many businesses use AI, not purchases of training data by the firms that build models, and it dates from late 2023. Epoch AI estimates the effective stock of human-generated public text at roughly 300 trillion tokens, with a 90 percent confidence interval of 100 to 1,000 trillion, and projects that language models would fully use it between 2026 and 2032 if current trends continue. That is a supply estimate for public text; it says nothing about private records, their price or who will pay.
The bridge between the two is buyer behaviour, and one buyer describes it directly. OpenAI's page on its approach to data says it partners to access content that is not publicly available, such as archives and metadata, and does not pursue paid partnerships for purely publicly available information. That supports the direction of the thesis, toward non-public data, without supplying a single transaction for business records. Adoption rates, token stocks and funding rounds can frame a scenario; they cannot be counted as observed demand.
How the evidence standard reorders the map, and the inputs still missing
The ranking of categories depends on the evidence standard, which is why a model must state its standard before it counts anything. Counting any public report, news text is 14 of 26 tracked deals, or 53.8 percent. Counting only deals with a first-party source, it is 14 of 21, or 66.7 percent. Counting only amounts a party disclosed, news text falls to 0 of 3, and books and scholarly hold 2 of the 3. Operational business records stay at zero under all three standards.
For a bottom-up model of business-record demand, the consequence is that every input is a scenario today. The list below gives the inputs such a model needs and the strongest public evidence available for each; label every value drawn from them illustrative and show it as a range.
- Number of buyers seeking non-public data: one published call read for this page (OpenAI's data partnerships page), grade three; no count of such buyers exists in first-party sources.
- Buyers seeking enterprise operational records specifically: micro1's stated 12-month plan, grade three; no completed deal found.
- Signed deals per buyer per year: no public evidence found for this category.
- Records or datasets per deal: no public evidence found for this category.
- Renewal or refresh rate: an analog only, from another category, where Informa labels its 2024 AI revenue non-recurring.
- Price basis: no first-party amount for this category; in the tracker, six of nine stated amounts in other categories come only from press reports.
Signposts that would move a cell
Four kinds of new evidence would move a cell from open question to observed, and an investor can watch for each.
The first is regulatory. Under Article 53(1)(d) of the EU AI Act, providers of general-purpose AI models must publish a summary of the content used to train them, on a template the European Commission presented on 24 July 2025. The obligation has applied since 2 August 2025, models placed on the market before that date have until 2 August 2027, and the AI Office can enforce it from 2 August 2026, according to the Commission. The summaries are high-level and cover models offered in the EU, but they create a first-party record of the kinds of sources a model used, licensed data included.
The second is a seller disclosing a line item, as Informa did for 2024. The third is a buyer announcing a completed purchase of operational records rather than a plan. The fourth is a public tender for licensed business data; this page did not search procurement databases, which could not be reached from the research environment, so that column stays unexamined rather than empty.
Five ways this evidence map could mislead
Each of the five weaknesses below comes with a test an investor can run.
Selection: the tracker is compiled by SourceX from publicly reported deals, so its zero for business records may reflect what reporters cover. Test it against other first-party sources as they appear. Confidentiality: SourceX's how-it-works page says neither side publicizes a deal, and its investors page says labs can source data without revealing their roadmap or identity, so the deals the thesis concerns are designed to stay out of public view. That cuts both ways: absence of evidence is expected, and it also means demand claims for the category can be checked only privately.
Concentration: 17 of the 26 tracked deals name OpenAI as a counterparty, so the observed market is partly one buyer's purchasing pattern, and a change in that buyer's strategy would reshape every row. Non-recurrence: Informa's figure, the one party-disclosed amount this page states (read in a search excerpt of the results), is revenue Informa itself calls non-recurring, which warns against annualizing any single deal. Substitution and rights: synthetic data, paid API access and unresolved rights questions could cap licensing demand in any category; the map records transactions, not whether they will continue.
Diligence questions for a company that cites observed demand
Ask any company that cites observed demand to grade each claim on the scale above, and expect it to show the evidence privately where contracts are confidential. SourceX is a fair test case. Its public pages give no demand or pipeline figures, and its investors page says it shares company and commercial metrics with investors directly. It is also seeking capital under Rule 506(c), which limits purchasers to verified accredited investors, so that private channel is where its own evidence would be graded.
These questions turn the map into a diligence list:
- Which evidence grade supports each demand claim, and which claims rest only on adoption rates, data-stock estimates or funding?
- Which data categories do signed agreements cover, and which are stated buyer interest only?
- What share of signed value or volume comes from the largest single buyer?
- Which agreements are one-time archive licences and which include repeat deliveries, as the agreement states them?
- For confidential deals, what redacted documents can an investor inspect to confirm grade one or two evidence?
- Which signposts on this page, such as training-content summaries or seller disclosures, would the company expect to confirm its category first?
Methods note: corpus, coding rules and what was not measured
Corpus: the DEAL_PAGES array in src/lib/kb/deals.ts of the SourceX site repository (last changed 2026-10-19, SHA-256 prefix 9f66b037), which feeds the live tracker at sourcex.si/deals; the live index fetched on 2026-10-28 listed the same 26 deal explainers and 8 roundups. Roundups were excluded because they summarize the deals. Script: bin/xs0374_tracker.py in the writer's scratch folder, with output and a snapshot of slugs, source hosts and codes beside it.
Coding rules: data type hand-assigned by one coder from each entry's title and answer. First-party source: any cited URL whose host equals or ends with a domain of a company named as a party, compared case-insensitively after removing www; a press-release wire page carrying a party's own announcement counts (prnewswire.com, once), as does Google Cloud's press site (googlecloudpresscorner.com, once, in an entry that also cites the other party's own page). Money amount: party when the entry attributes an amount to a party's own filing or disclosure, press when every stated amount is attributed to press reports. OpenAI concentration: entries naming OpenAI as a party, including one where the deal was reported but not confirmed. Strongest grade per deal: grade one when the amount is coded party, otherwise grade two when a first-party source is cited, otherwise press only. Percentages are rounded once, to one decimal, from exact counts.
Sources read for this page: OpenAI's data partnerships, approach to data and Axel Springer pages; Informa's 2024 full-year results; micro1's announcement; the Census Bureau's November 2023 release; Epoch AI's data-stock publication; and the Commission's template page, each through a search limited to its own domain. Not measured: deal values, which the tracker mostly does not have from first-party sources; procurement notices; and the first-party pages the tracker cites but this page did not open, whose existence is taken from the tracker's citations. No market size, growth rate or share is estimated.
Frequently asked questions
What counts as observed demand for AI training data?
Evidence a reader can open that a buyer agreed to pay for licensed data: a seller's filing or results release, an agreement announced by a party to it, or a public tender. A buyer's published call or plan shows intent, and AI budgets or data volumes do not show licensing demand at all.
Can first-party evidence give a total size for AI data licensing?
Not today. The few amounts parties disclose describe one seller in one year, such as Informa's report of more than $75 million of AI partnership revenue at Taylor & Francis in 2024, and this page deliberately gives no total.
Which data types have the most public licensing deals?
News and magazine text: 14 of the 26 deals in SourceX's public tracker, all citing a party's own announcement. None of the 14 carries an amount disclosed by a party, so the category with the most deals has no first-party price.
Is there public evidence that AI companies buy operational business records?
There is stated intent, such as OpenAI's call for private datasets and micro1's announced 12-month plan, but no completed, priced transaction among the tracked deals. Confidential contracts may explain part of the gap, so investors should verify such claims privately.
What would make demand for business records observable?
Training-content summaries that EU rules require from general-purpose AI model providers, a seller disclosing licensing revenue from operational data, a buyer announcing a completed purchase, or a public tender for licensed business data.
Sources
- SourceX's public AI data deal tracker lists 26 deal explainers and 8 roundups and describes itself as covering significant publicly reported AI data licensing deals. Source
- SourceX's how-it-works page says neither side publicizes a deal. Source
- SourceX's investors page says labs can source data without revealing their roadmap or identity and that SourceX shares company and commercial metrics with investors directly. Source
- OpenAI's data partnerships page seeks large-scale datasets not already easily accessible online, including private datasets, and says it is not seeking sensitive, personal or third-party information. Source
- OpenAI says it partners to access non-publicly available content such as archives and metadata and does not pursue paid partnerships for purely publicly available information. Source
- OpenAI's Axel Springer partnership page says the collaboration involves using Axel Springer content to advance the training of OpenAI's large language models. Source
- Informa's 2024 full-year results report more than $75 million of AI partnership revenue at Taylor & Francis in 2024, described as non-recurring data access. Source
- micro1 announced a commitment to spend $1 billion over 12 months acquiring and licensing enterprise operational data for reinforcement learning environments. Source
- The U.S. Census Bureau reported in November 2023 that about 4 percent of businesses used AI to produce goods and services. Source
- Epoch AI estimates the effective stock of human-generated public text at about 300 trillion tokens and projects full use between 2026 and 2032 if trends continue. Source
- The European Commission's template for training-content summaries under Article 53(1)(d) applies from 2 August 2025, with legacy models due by 2 August 2027 and enforcement from 2 August 2026. Source
Request SourceX investor materials
SourceX shares company and commercial metrics with investors directly. Any offering is made only to verified accredited investors.