AI data market
Unavailable data and AI market constraints: six routes when records never sell
By SourceX Editorial · Reviewed by Noah Loul ·
10 sections
Short answer
Some of the most valuable AI data will never be sold as a copy: health records under sale bans, data large buyers refuse to take, and archives owners keep for their own models. That is a market constraint, not a dead end. Value moves to six permitted routes, from code-to-data environments to owner-built models and licensable proxies. Investors should ask which route recovers the most capability, and who captures it.
Value drivers
Unavailable data AI market
Key takeaways
- Unavailable data comes in five classes, barred by law, refused by buyers, withheld by owners, restricted by publishers and never captured, and each sends value somewhere different.
- The opportunity cost is the capability the best permitted route cannot recover, times the value of the task, adjusted for when the class might open.
- Six permitted routes recover part of the value without weakening rights standards, from code-to-data environments to licensable proxies.
- When copies cannot move, value shifts to environment operators, holder-built models and suppliers of lawful substitutes, not to sellers of copies.
- Synthetic data and stronger models are the main threat to the thesis; five dated forecasts, each with the signal to check it against, show a reader which way it breaks.
Which unavailable data creates real AI market constraints?
The constraint that matters is data that cannot change hands as a copy on any terms a lawful buyer will accept, not data that is merely hard to find. Some of the records most useful to AI sit in that group, and no price will move them. For an investor, the question is not whether this data exists but where its value goes instead.
Three public signals show the group is real. HHS guidance says the HIPAA Privacy Rule prohibits a covered entity from selling protected health information unless it obtains an authorization stating that it will receive remuneration for the disclosure. OpenAI's data partnerships page asks for large datasets the public cannot easily reach online, yet rules out anything containing sensitive or personal information, or information owned by a third party. And Bloomberg put its own archive to work in-house: it trained a 50-billion-parameter model, BloombergGPT, on a 363-billion-token dataset built from its own data sources, according to the 2023 paper describing it.
Those are three different kinds of unavailability: barred by law, refused by buyers and withheld by owners. Each sends value to a different place, which is why treating unavailable data AI market constraints as a single wall leads to the wrong investment conclusion. This article maps where the value goes, rates what each permitted route recovers, sets out four scenarios with signposts, and commits to five forecasts that can be graded later.
Five reasons the best records stay off the market
Records stay off the market for five reasons, and they differ most in how long they last. Durability matters more than size: a class that opens in three years is a timing bet, while a class that never opens is a structural limit the market has to route around.
Only the first and last classes are close to permanent. Buyer refusals and owner withholding are commercial choices that shift with prices and policies. Publisher restrictions are moving in the opposite direction from what a scarcity bull would hope: in the Consent in Crisis audit, which covered 14,000 web domains, more than 28 percent of the most actively maintained, critical sources in the C4 training corpus became fully restricted from use between 2023 and 2024. Public sources are closing, not opening.
| Unavailability class | Why it stays off the market | Dated public evidence | Durability (editorial) | What would change it |
|---|---|---|---|---|
| Barred by law | A statute or regulation forbids sale or limits reuse to approved purposes | HHS guidance: HIPAA bars sale of protected health information without an authorization that states the remuneration | High: change needs new law | A statute that creates a permitted research or innovation route |
| Refused by buyers | Buyers' own standards exclude the data even when a seller would sell | OpenAI data partnerships page: not seeking sensitive or personal information or third-party information | Medium: buyer policies can change quickly | A published change in a large buyer's sourcing policy |
| Withheld by owners | The holder expects more from using the records itself than from licensing them | BloombergGPT paper, 2023: a 363-billion-token dataset from Bloomberg's own sources trained an in-house model | Medium: depends on the owner's AI returns | Owner-built models that trail general models on the owner's own tasks |
| Restricted by publishers | Terms of service and crawler rules withdraw content that was once open | Consent in Crisis, 2024: more than 28 percent of critical C4 sources fully restricted within a year | Rising: restrictions spread faster than licenses | Paid licensing terms that allow AI use |
| Never captured | The decision or outcome was never written down, or was purged | No public measure exists; it shows up only in archive audits | Permanent for the past, fixable going forward | Logging that captures decisions and outcomes from now on |
What does the market lose when the best data never sells?
It loses the capability that the best permitted alternative cannot recover, and that loss is the opportunity cost an investor should model. Write it as three terms: the recovery gap between the best permitted route and full access, the value of the tasks that needed the data, and the expected time until the class opens. None of the three has a published market-wide measurement, so every estimate is an assumption to argue over, not a figure to quote.
There is evidence that the recovery gap can be large. The BloombergGPT authors report that their model outperformed existing models on financial tasks by significant margins without sacrificing performance on general benchmarks, after training on roughly half domain data and half general data. NVIDIA's Nemotron training FAQ puts data and training infrastructure first among what matters, ahead of methodology and architecture, and says a strong dataset can overcome limitations in both.
The public substitute is also finite. Epoch AI estimated in 2024 that the effective stock of quality, repetition-adjusted public human text is about 300 trillion tokens, and that on current trends it will be fully used at some point between 2026 and 2032. As the public stock runs down, the recovery gap for private domains stops narrowing on its own, and the opportunity cost of unavailable records rises unless synthetic data or stronger models close it. That alternative is tested in the scenario matrix below.
Time is the term investors most often skip. A class barred by statute opens on a schedule set by legislation, which can be tracked; a class withheld by an owner opens when the owner's own returns fall, which cannot. That difference decides whether a fund should wait for supply or back the routes around it.
Six permitted routes around data that never sells, rated by what they recover
Six routes reach part of the value of unavailable records without weakening rights standards, and they recover very different amounts. The table rates each route on what an outside model developer can get from it, which unavailability classes it fits and who captures the value. The ratings are editorial assumptions as of October 26, 2026, built from the sources cited; replace them with your own.
The first route already works at national scale in health care. OpenSAFELY's documentation describes federated analysis in which researchers write code once, run it inside several secure backends that hold patient records and retrieve only summary outputs checked for safety, so raw clinical data never leaves its original location; its backends give researchers access to the records of nearly 60 million patients in England. The second route is now law in the EU: the European Health Data Space Regulation, published in the Official Journal on 5 March 2025, limits secondary use of health data to listed purposes under a permit from a health data access body, with processing in secure environments from which personal data cannot be downloaded.
The last route is the one most often overlooked. When the best records are barred, the nearest licensable substitute can carry a scarcity premium of its own. An insurer's claims-handling workflow records, stripped of clinical content and properly licensed, may teach a model much of the process knowledge a claims file holds. That is an editorial hypothesis, not a measured result, and it is testable: evaluate a model trained on the proxy against the barred records inside a secure environment, using the first route to measure the sixth.
| Route | How rights stay intact | Fits which classes | What an outside developer recovers (editorial) | Who captures the value |
|---|---|---|---|---|
| Code goes to the data | Analysis runs inside the holder's environment and only checked outputs leave | Barred by law; withheld by owners | High for analysis and evaluation; partial for training, because a trained model is itself an output | The environment operator and the data holder |
| Regulated access permit | A public body grants a permit for listed purposes inside a secure processing environment | Barred by law | Medium: only purposes the law lists, at the pace of permit review | Permit holders and secure environment operators |
| Owner-built model | The holder trains on its own records and offers the model, not the data | Withheld by owners | None as data; outsiders can only use the owner's model | The data holder |
| Individual authorization | Each person authorizes the use, including any payment, where the law allows it | Barred by law; refused by buyers | Low: small, self-selected cohorts that rarely represent the whole | Whoever can run consent at scale |
| Derived data where allowed | Aggregates, de-identified extracts or synthetic records made under a permitted purpose | Barred by law; restricted by publishers | Medium for common patterns; low for the rare cases a derivation smooths away | Holders that can document the permitted purpose |
| Licensable proxy | Adjacent records carry the same decisions without the barred element | All four classes above, not never captured | Medium to high for process knowledge; none for the barred content itself | Licensors and the intermediaries that find proxies |
Who captures the value when copies cannot move?
Value moves away from whoever sells copies and toward whoever controls the environment, owns the records or supplies the nearest lawful substitute. That is the investment consequence of an unlicensable best dataset, and it cuts against the simple view that scarcity always rewards data sellers.
Three positions gain. Operators of secure environments gain because every barred class that opens through a permit or a federated design runs through them. Holders able to build their own models gain because withholding becomes a strategy rather than a loss. Suppliers of licensable proxies gain because buyers that cannot get the best records still need the nearest lawful substitute, and OpenAI's own sourcing page describes that profile closely: non-public, large-scale and free of sensitive personal or third-party information.
One position loses: an intermediary whose business depends on holding and reselling copies of sensitive records. If the market's answer to a barred class is that code travels to the data, the intermediary never holds inventory and has to earn its place through rights review, preparation and access design instead. Investors should ask any data company which of these positions it occupies for each class it claims to serve, because the same company can be a winner for one class and a loser for another.
Four futures for data that never sells: the scenario matrix
Four scenarios cover the plausible outcomes through 2030, and each has a signpost to watch and a signal that would rule it out. They are not exclusive across data classes: health records could follow one path while confidential customer files follow another, so assign each class to the scenario its signposts currently favor.
The fourth scenario is the strongest alternative explanation and needs its own test. NVIDIA's Nemotron FAQ says its models use synthetic data in post-training, that the pretraining corpus also includes synthetic and rephrased sources, and that synthetic data is validated by training small models on it to check that they learn something. If synthetic routes keep closing gaps in data-poor domains, the opportunity cost of barred records falls, and so does the premium on proxies.
Do not average the four into one forecast. A fund backing secure environments wins in the second scenario and loses little in the first; a fund backing proxy suppliers needs the third and is hurt by the fourth. The useful output is a position per data class, revisited whenever a signpost moves.
| Scenario | What happens | Who gains | Signpost to watch | Kill signal |
|---|---|---|---|---|
| Walled gardens | Holders keep their best records and build their own models | Holders with AI teams; secure computing vendors | Data holders announcing domain models trained on their own archives | General models match holder-built models on the holders' own benchmarks |
| Code travels to data | Secure environments become the default way to train or evaluate on barred classes | Environment operators; permit holders | EHDS secondary-use rules phasing in; trusted environments approving model training projects | Permit reviews so slow or narrow that applicants give up |
| Proxy premium | Buyers pay more for lawful substitutes adjacent to barred data | Licensors of adjacent records; sourcing intermediaries | Buyer sourcing pages asking for non-public process records, not only content | Proxy-trained models fail when tested against the barred data |
| Constraint fades | Synthetic data and stronger models shrink the recovery gap | Model developers; synthetic data vendors | Technical reports relying more on synthetic data in pretraining and post-training | Domain benchmarks where models without the barred data keep trailing |
Public forecast scorecard: five dated calls to check by 2029
Forecasts are useful only if they can be wrong, so this page makes five dated calls. Each carries a date and the signal to check it against: what would make it true and the public record that would show it. The calls are editorial judgments as of October 26, 2026, not measured results, and none depends on SourceX data.
Once a call's date passes, a reader can check it against that signal. A call whose evidence has not become public by its date cannot be settled either way, which is itself a finding about how opaque this market remains.
| Forecast | Editorial call (2026-10-26) | Counts as true if | Check on or after |
|---|---|---|---|
| A national trusted research environment or EHDS secure environment approves a project that trains a machine learning model inside it | Likely | A public project register or official notice describes the approved training project | 2028-12-31 |
| Large model developers keep excluding sensitive personal data from what their public sourcing pages seek | Likely | No major developer's official sourcing page solicits identifiable sensitive records | 2027-12-31 |
| At least two large AI developers publicly seek non-public business process records, such as workflows and outcomes, not only content | Even odds | Two official developer pages or announcements describe such a program | 2027-12-31 |
| A general model trained without a holder's private data matches that holder's own domain model on the holder's published benchmark | Likely | A technical report or paper shows the match on the same benchmark and version | 2028-12-31 |
| A public tracker reports that frontier training has used most of the effective stock of public human text | Even odds | Epoch AI or a comparable tracker publishes that finding | 2029-06-30 |
What could make the unavailable-data thesis wrong
Five developments would weaken it, and the first two are already visible. Investors should watch them as closely as the thesis itself.
A sixth risk is self-inflicted: pressure to weaken rights standards to reach barred records. Any route that leans on thin consent, stretched de-identification or ambiguous terms turns a market constraint into a legal liability, and none of the six routes above requires it.
- Synthetic data closes the gap. Nemotron already uses synthetic and rephrased sources in pretraining and synthetic data in post-training; if that keeps working in data-poor domains, barred records matter less.
- Law opens a barred class. The EHDS creates a permitted route for health data, which lowers the premium on proxies even as it raises the value of secure environments.
- Buyer concentration. A few large buyers set sourcing standards, so one policy change can move a class from refused to wanted, or the reverse, with no change in law.
- Owner returns disappoint. If holder-built models trail general models, owners will license again and the walled-garden scenario fails.
- The gap was never measured. No public data measures the capability lost to unavailable records, so the opportunity-cost model rests on assumptions; ask for the evaluation design behind any claim of scarcity value.
Questions to ask a data company about records it cannot sell, SourceX included
Ask each company which unavailability class it serves, which route it uses and what evidence shows the route recovers value. The questions below work for any supplier, broker or platform in this market.
Applied to SourceX: the company helps businesses license selected records as AI datasets, and it carries out the sourcing and qualification, the rights review, the contract, data preparation, delivery and payment. Data moves only after the owning company has approved the scope and terms and signed; from then on, the de-identified dataset can be used, sublicensed and distributed by SourceX within the limits that agreement sets. Read against the routes above, that is a licensing model for records whose owners can approve their use; records that cannot lawfully be licensed fall outside it.
SourceX has not published how often requested record classes turn out to be unlicensable or which routes buyers choose instead; its company and commercial metrics go to investors directly. SourceX raises under Rule 506(c) from accredited investors only, with investor materials available through sourcex.si/investors on request. A useful diligence question to bring: which unavailability classes appear most often in declined requests, and how are they counted?
Checklist · 5 items
- Which requested record classes have you declined, and on what grounds: law, buyer policy, owner choice or missing history?
- For each route you offer, who holds the copy, who checks the outputs and what leaves the environment?
- Have you tested a proxy dataset against the data it substitutes for, and how was that evaluation designed?
- Which buyer sourcing policies would, if changed, remove demand for your largest data class?
- How do you document the permitted purpose behind any derived or synthetic dataset you sell?
0 of 5 ticked. Tick each item as you go.
Methods note: sources, dates and what was not measured
The taxonomy, route ratings, scenarios and forecast calls are editorial frameworks built on October 26, 2026 from the sources listed below: HHS guidance on the sale of protected health information, the European Commission's EHDS page, OpenSAFELY's federation documentation, OpenAI's data partnerships page, the BloombergGPT paper, NVIDIA's Nemotron training FAQ, Epoch AI's 2024 study of the public text stock and the Consent in Crisis audit. Each source was checked on its publisher's own domain.
Not measured: the size of any unavailable class, the capability gap any route leaves, the prices buyers pay for proxies and SourceX's own experience with declined classes. The recovery ratings and forecast calls are judgments a reader can overwrite. Descriptions of HIPAA and the EHDS are general information about how their administrators describe them, not legal advice; counsel should review any specific question.
Frequently asked questions
What does unavailable data mean for AI markets?
Unavailable data is data that cannot change hands as a copy on terms a lawful buyer accepts, because law bars it, buyers refuse it, owners withhold it, publishers restrict it or it was never recorded. It limits which capabilities can be bought, but value then moves to other routes, such as secure environments, owner-built models and licensable proxies.
Can AI be trained on data that cannot be shared?
Sometimes. In federated or trusted-research designs the code runs inside the data holder's environment and only checked outputs leave, as OpenSAFELY's documentation describes for patient records in England. A trained model is itself an output, so training in such environments needs its own review of what the model could reveal.
Does the European Health Data Space make health data licensable?
Not as a copy. The EHDS Regulation, published in March 2025, allows secondary use of health data for listed purposes under a permit from a health data access body, with processing inside secure environments from which personal data cannot be downloaded. It is a regulated access route, not a commercial license.
Will synthetic data replace data that can never be licensed?
It may shrink the gap in some domains; NVIDIA's Nemotron models already use synthetic data in pretraining and post-training. Whether it replaces barred records in a given domain is an empirical question, best answered by comparing models on real held-out cases evaluated inside a secure environment.
Does SourceX license records that are barred from sale?
SourceX licenses selected business records through rights review and an agreement the company approves and signs before any data is delivered, so records that cannot lawfully be licensed fall outside that process. SourceX has not published metrics on declined record classes and shares company metrics with investors directly.
Sources
- HHS guidance says the HIPAA Privacy Rule prohibits selling protected health information unless an authorization states that the covered entity will receive remuneration. Source
- OpenAI seeks large-scale datasets not easily accessible online and says it is not seeking sensitive or personal information or information that belongs to a third party. Source
- BloombergGPT is a 50-billion-parameter model trained on a 363-billion-token dataset from Bloomberg's data sources plus 345 billion general tokens; the authors report it outperforms existing models on financial tasks without losing general benchmark performance. Source
- The Consent in Crisis audit of 14,000 web domains found that between 2023 and 2024 more than 28 percent of the most critical sources in C4 became fully restricted from use. Source
- NVIDIA's Nemotron training FAQ ranks data first among what matters in training, uses synthetic data in pretraining and post-training, and validates synthetic data with small-scale training ablations. Source
- Epoch AI estimated the effective stock of public human text at about 300 trillion tokens, fully used between 2026 and 2032 on current trends. Source
- OpenSAFELY runs analysis code inside secure backends and releases only checked summary outputs, so raw clinical data never leaves its location; researchers reach records of nearly 60 million patients in England. Source
- The EHDS Regulation, published on 5 March 2025, allows secondary use of health data for listed purposes under a permit, in secure processing environments from which personal data cannot be downloaded. Source
Request SourceX investor materials
SourceX shares company and commercial metrics with investors directly. Any offering is made only to verified accredited investors.