Skip to content

AI data market

Permission, not compute? A six-constraint test for the AI data rights bottleneck

By SourceX Editorial · Reviewed by Noah Loul ·

9 sections

Short answer

The AI data rights bottleneck is real but narrow. Frontier training compute grew about 4-5x a year through May 2024 (Epoch AI), and Epoch now calls a public-text shortage surmountable, so permission rarely limits general pretraining. It binds where a capability needs private operating records that exist, fit a paying buyer and are affordable to prepare, yet cannot be exported or licensed. Test all six constraints before backing the thesis.

License terms

AI data rights bottleneck

This page covers permission, not compute? A six-constraint test for the AI data rights bottleneck.Illustrative

Key takeaways

  • Permission limits specific capabilities, not AI progress in general; published forecasts treat compute and public text as surmountable before 2030.
  • Permission has two halves that fail differently: export access, which platform terms increasingly set, and licensing authority, which contracts and ownership set.
  • The rights-versus-utility matrix separates permission-bound supply from records nobody needs; only the first supports the thesis.
  • A credible usable-data funnel reports stage denominators, rejection reasons and time-to-resolution with open cases censored, not dropped.
  • Preparation cost and concentrated demand are live competing explanations; Reddit attributes substantially all licensing contract value to two partners.

Is permission really the next AI bottleneck?

Not for AI in general, but yes for a specific class of capabilities. Permission becomes the binding constraint when a model needs records that only private companies hold, those records exist in usable form, a buyer will pay for the capability and preparation is affordable, yet nobody can export the records with their context or nobody has the authority to license them.

That conditional is the whole thesis, and the slogan version breaks easily. If permission limited AI progress in general, every frontier training run would wait on licenses, and the public record points elsewhere. Epoch AI, whose data forecasts scarcity pitches often borrow, argued in a May 2023 viewpoint that total data availability would not constrain general AI progress in the foreseeable future, while widespread automation could be bottlenecked by data on a subset of essential tasks that need scarce, high-quality sources.

That subset is where an AI infrastructure investment thesis about data should start. This article makes it testable: six constraints that can each be the real limit, a rights-versus-utility matrix that locates the cases permission actually explains, a funnel design that measures them, the competing explanations, and the evidence that would prove the thesis wrong.

The six-constraint test: which input runs out first?

For any target capability, permission is the bottleneck only when export access or licensing authority fails while compute, the public data stock, preparation and demand all pass. Run the six rows in order and stop at the first that fails; that row, not the loudest headline, sets the investment case.

The first two rows rarely fail for business capabilities. Epoch AI reported in May 2024 that training compute for frontier models had grown by about 4-5x per year. Its 2024 data study put the effective stock of quality, repetition-adjusted public human text at about 300 trillion tokens and projected full use between 2026 and 2032 on trend, earlier with heavy overtraining. Its later outlook, What will AI look like in 2030?, still finds a data bottleneck hard to rule out but judges it surmountable, partly through synthetic data.

Rows three to six are where enterprise AI data investment lives. A claims-handling agent, a dispatch planner or a procurement assistant needs records of real decisions and outcomes that were never published, and no amount of compute or public text substitutes for them. The open question is whether those records are blocked by permission or by something cheaper to fix.

The six-constraint test: which input runs out first?
ConstraintBinds whenPublic evidence, datedDiligence question
ComputeTraining or inference capacity, not inputs, limits the capabilityEpoch AI, May 2024: frontier training compute grew about 4-5x per yearIs the capability blocked by chips, or by inputs nobody can supply?
Public data stockThe capability could learn from public data, but the stock is used upEpoch AI, 2024: about 300 trillion tokens of public text, fully used between 2026 and 2032 on trendWould synthetic or multimodal data close the gap?
Export accessRecords exist but cannot leave their system with links and history intactSlack API terms: apps offered to other organizations may not train LLMs on API dataWhich route moves the records, and who controls it?
Licensing authorityNo one with authority can approve the outside useConsent in Crisis, 2024: over 28 percent of critical C4 sources became fully restricted in a yearWho signs, on which documents, and what do customer contracts allow?
Preparation costRights are clear, but records cannot be made usable affordablyScale AI, June 2025: a Meta investment valued the data company at over $29 billionHow many rights-cleared datasets are waiting on preparation?
Buyer demandUsable, licensable records have no paying taskReddit annual report for 2024: substantially all licensing contract value from two partnersHow many distinct buyers issued written requests?

Export access and licensing authority fail in different ways

Permission has two halves, and a deal can fail at either. Export access asks whether records can leave the system that holds them with their links and history intact. Licensing authority asks whether anyone can approve an outside use under the company's contracts, notices and ownership. Owning the records settles neither question.

Export access is increasingly set by platform terms. Under Slack's API terms, an app offered beyond its own organization is barred from training a large language model on API data and from bulk exporting message and file data unless an additional agreement allows it. The workspace owner can still use Slack's own export tools, so the rule does not lock the data away; it removes the easy third-party route and puts the data owner in the delivery path.

Licensing authority is the harder half, and public data shows how fast it can move. The Consent in Crisis audit of 14,000 web domains found that between 2023 and 2024, new restrictions left more than 5 percent of all tokens in the C4 training corpus, and more than 28 percent of its most actively maintained critical sources, fully restricted from use. Private records start further back: every outside use needs someone with authority to say yes.

Buyers price access and authority, not bytes. On its data partnerships page, OpenAI says it wants large-scale datasets the public cannot already get at easily online, and it rules out information owned by a third party. Reddit's 2024 S-1 shows what that kind of access can fetch: the data licensing deals it signed in January 2024 carried $203.0 million in combined contract value, over terms of two to three years. Either way, what commands the price is permissioned access to material others cannot reach.

The rights-versus-utility matrix: where the thesis lives

The permission thesis lives in one quadrant: records with high buyer utility whose rights are blocked or unproven. Place every candidate dataset on two axes, rights status and buyer utility, and the thesis explains only the cases that land there.

Rights status runs from blocked through unproven to clear, judged on documents rather than belief: customer contracts, privacy notices, platform terms and the authority of the person who would sign. Buyer utility runs from none to high, judged against a named task and a written acceptance test, never against volume. A large stock of records can sit in the low-utility half indefinitely, which is why a record count says little about a data business.

Three illustrative cases show why the quadrant matters more than the volume. Illustrative case A: a field services company holds a decade of dispatch records that a workflow-agent developer wants, but its customer contracts restrict disclosure and its software vendor's terms limit exports, so abundant records are permission-bound. Illustrative case B: a specialty insurer's adjuster notes cover a few thousand claims, but its contracts and notices permit de-identified licensing, so the records are licensable and the limit is volume. Illustrative case C: a distributor's order history is clean and fully licensable, but no buyer has a task that needs it, so clear rights meet no demand.

Only case A supports the permission thesis. Case B is a sourcing problem and case C is a demand problem, and a company that counts all three as supply overstates its moat. The figure plots the quadrants and the three illustrative cases; the positions are examples, not measurements.

The rights-versus-utility matrix: where the thesis lives
QuadrantRights statusBuyer utilityWhat actually limits supplyInvestor read
Licensable nowClear and documentedHigh: a named task and acceptance testPreparation capacity and delivery timeAn execution business; permission is not the moat here
Permission-boundBlocked or unprovenHigh: a named task and acceptance testExport access or licensing authorityThe thesis case: value comes from resolving rights others cannot
Rights without buyersClear and documentedLow: no named taskDemandStock without demand; volume claims mislead
Walk awayBlocked or unprovenLow: no named taskBoth rights and demandNo business here; screen it out early
A two-by-two matrix of rights status against buyer utility, with the permission-bound quadrant highlighted and three illustrative cases placed in three quadrants.
The rights-versus-utility matrix. Illustrative: case positions are examples, not measured data.

How to measure the usable-data funnel without fooling yourself

Measure stage conversion and time-to-resolution with fixed denominators and censoring, because a funnel that reports only finished cases always looks faster and cleaner than reality. Five stages map onto the matrix: do records exist, can they leave, may they be licensed, do they fit a buyer, and did a buyer accept them.

Time-to-resolution runs per stage, in calendar days, from entry to exit or rejection. Cases still open on the measurement date are censored: kept in the denominator, not counted as failures and not dropped. Summarize with a survival curve, such as a Kaplan-Meier estimate, or a median reported with its censored count; an average of closed cases overweights fast rejections and fast wins.

Report the sample size at each stage, the share censored, the entry date range and the rule that admitted a company to the funnel. Inbound applicants and outbound sourcing targets behave differently, and mixing them hides the constraint the funnel exists to find.

The verdict follows from where cases stall. Rejections and long medians at stages 2 and 3 mean permission binds; clusters at stage 4 mean demand or preparation binds. SourceX has not published results for this funnel. This is the measurement design, and the same definitions apply to every company in the market.

How to measure the usable-data funnel without fooling yourself
StageEntry testExit eventRejection reason codesDenominator
1. Raw availabilityRecords exist for the named systems and yearsInventory confirmed by system and date rangePurged, migrated without history, never capturedCompanies that completed a fit check
2. Export accessRecords can leave with links and history intactSample export reviewedPlatform terms, no export route, context lost in exportCompanies that passed stage 1
3. Licensing authoritySomeone with authority can approve an outside useRights review signed offCustomer contract, third-party content, no signing authority, prior exclusive licenseCompanies that passed stage 2
4. Buyer-usable scopePrepared records match a written buyer requestScope agreed against a task and acceptance testNo matching request, over-redaction, volume too smallDatasets that passed stage 3
5. Accepted deliveryDelivered scope meets the buyer's acceptance testAcceptance recordedFailed test, missing context, quality defectsDatasets delivered

Could preparation cost or buyer budgets be the real constraint?

Yes. Both explanations have public evidence behind them, so a permission thesis has to beat them on rejection data rather than assume them away.

At a glance

Could preparation cost or buyer budgets be the real constraint

Preparation cost first. De-identifying records, linking them across systems, labeling outcomes and documenting provenance take skilled work, and model developers pay for it at scale: Scale AI said in June 2025 that a new investment from Meta valued the company at over $29 billion. If preparation were the binding constraint, rights-cleared datasets would queue at stage 4 waiting for work, not stall at stage 3 waiting for signatures.

Buyer budgets second. In the risk factors of its 2024 annual report, Reddit traces substantially all the contract value of its licensing revenue to just two partners and cautions that the deals could lapse at renewal or come back on worse terms. A permission-rich supplier can still be demand-bound when only a few buyers sign.

Read the funnel for one of three patterns, and ask the company which one its own cases follow:

Checklist · 3 items

  1. Permission binds: rejection reasons at stages 2 and 3 dominate, and the same records would pass stage 4 if the rights resolved.
  2. Preparation binds: rights are cleared, but datasets wait at stage 4 for de-identification, linking or labeling.
  3. Demand binds: cleared, preparable records have no written buyer request that names a task.

0 of 3 ticked. Tick each item as you go.

What could make the permission thesis wrong

Four developments could shrink the value of resolving permission. Put them on the table before the upside; if a company cannot tell you which of the four would damage it most, it has not stress-tested its own case. Each comes with a dated source an investor can track.

  • Courts could make licensing optional for more content. In Bartz v. Anthropic (N.D. Cal., June 23, 2025), the court found that using the books at issue to train Claude was exceedingly transformative and a fair use, as quoted by the Southern District of New York in Lovo v. Lehrman (2025). Rulings like this lower the premium on licensed public content, though they give no route to records a company never published.
  • Synthetic data could substitute. Epoch AI's 2030 outlook counts on synthetic data to ease public-data limits; if synthetic workflows match real operating records on a buyer's acceptance test, the premium for permissioned records falls.
  • Demand could stay concentrated. If licensing budgets keep pooling with a couple of buyers, as the two-partner disclosure in Reddit's 2024 annual report suggests, a supplier carries renewal and pricing risk however clean its rights are.
  • Regulation adds cost as well as value. Since 2 August 2025, the EU AI Act has required general-purpose AI model providers to adopt a copyright policy and publish a summary of training content. That favors documented supply, but it also adds review work and delay to every license.

Questions to ask any company selling permissioned data, SourceX included

Ask for evidence at the stage where the company claims its advantage, using definitions you set rather than ones it chooses. SourceX should face the same list.

Treat SourceX as one test case for the thesis, not as proof of it. It converts business records that a company chooses into licensed AI datasets, and it owns the chain end to end: it sources and qualifies the records, reviews the rights, writes the contract, prepares the data, delivers it and settles payment. Nothing reaches SourceX until the supplying company has signed off on scope and terms and executed the agreement, so the licensing-authority stage comes before any records are handed over.

None of SourceX's funnel, deal or commercial figures are public; the company gives its company and commercial metrics to investors directly rather than printing them in articles. Its raise is limited to accredited investors under Rule 506(c). An investor can ask for the investor overview and put any question below to the company, starting with how stage 3 time-to-resolution is defined.

Checklist · 6 items

  1. Which matrix quadrant holds most of your active datasets, and who decides rights status and buyer utility?
  2. What were your stage denominators, rejection reason codes and censored counts over the last four quarters?
  3. What is the median time from entry to sign-off at the licensing-authority stage, with the censored count beside it?
  4. What share of rejections came from contracts or platform terms rather than from buyer fit or preparation?
  5. How many distinct buyers issued written requests that reached the buyer-usable scope stage?
  6. Which court ruling, synthetic-data result or regulation would remove the need to license a category you supply?

0 of 6 ticked. Tick each item as you go.

Methods note: evidence behind the six-constraint test and funnel design

The six-constraint test, the rights-versus-utility matrix and the funnel design were drawn up for this article on 2026-10-26 as editorial tools. They rest on Epoch AI publications (2023 and 2024), the Consent in Crisis audit (2024), OpenAI's data partnerships page, Reddit's 2024 S-1 and annual report, Slack's API terms, Scale AI's June 2025 announcement, the European Commission's AI Act guidance and the court opinion cited above. The three cases are illustrative and describe no real company or deal.

Not measured: the share of real datasets in each quadrant, stage conversion or time-to-resolution at SourceX or any other company, buyer budgets, and the effect of rights status on price. None of these can be measured from outside without company records that nobody has published.

Frequently asked questions

Is data or compute the bigger constraint on AI in 2026?

For general pretraining, neither looks like a hard wall on published evidence: Epoch AI measured frontier training compute growing about 4-5x a year through May 2024 and judges a public-text shortage surmountable. For capabilities that depend on private business records, compute is rarely the limit. Access to those records and authority to license them usually decide whether the capability can be built.

What is the difference between data access and data rights?

Access is the technical and contractual ability to get records out of a system, which platform terms such as Slack's API terms can restrict. Rights are the authority to let someone else use the records for a stated purpose, which depends on ownership, customer contracts, notices and third-party content. A license needs both, and a company can easily have one without the other.

Do fair use rulings make data licensing unnecessary?

Not for private records. A ruling such as Bartz v. Anthropic (2025) concerns content a developer already holds; it gives no route to records a company never published, such as support histories or claims files. Rulings can lower the premium on licensed public content, which is why they belong in the risk section of any thesis.

How can an investor verify a data supplier's supply claims?

Ask for stage-level denominators, rejection reason codes and time-to-resolution with censored cases reported, plus the number of distinct buyers behind written requests. Then place the supply on the rights-versus-utility matrix: datasets in the permission-bound quadrant support a rights advantage, while datasets with clear rights and no buyer task do not.

Does SourceX publish its funnel or deal metrics?

No. Its funnel, deal and commercial figures are not public; SourceX provides company and commercial metrics to investors directly. Materials can be requested at sourcex.si/investors, where an investor can also ask how each funnel stage is defined. Any offering it makes is open only to accredited investors whose status has been verified.

Sources

  • Epoch AI reported in May 2024 that training compute for notable and frontier AI models grew about 4-5x per year. Source
  • Epoch AI estimated the effective stock of quality, repetition-adjusted public human text at about 300 trillion tokens and projected full use between 2026 and 2032, earlier with heavy overtraining. Source
  • Epoch AI's 2030 outlook finds a data bottleneck difficult to fully rule out but surmountable, with synthetic data producible in large quantities. Source
  • An Epoch AI viewpoint (May 31, 2023) argued total data availability will not constrain general AI progress, while automation may be bottlenecked by data on essential tasks needing scarce, high-quality sources. Source
  • Slack's API terms bar apps offered outside the developer's organization from using API data to train a large language model and from bulk exporting message and file data without an additional agreement. Source
  • The Consent in Crisis audit of 14,000 web domains found that over 2023-2024 restrictions left 5 percent or more of C4 tokens, and 28 percent or more of its most actively maintained critical sources, fully restricted. Source
  • OpenAI's data partnerships page seeks large-scale datasets not already easily accessible online and says it is not seeking sensitive or personal information or information that belongs to a third party. Source
  • Reddit's S-1 disclosed data licensing arrangements entered in January 2024 with an aggregate contract value of $203.0 million and terms of two to three years. Source
  • Reddit's annual report for 2024 says substantially all contract value associated with its licensing revenue comes from two partners and that the arrangements may not be renewed or may be renewed on less favorable terms. Source
  • Scale AI announced in June 2025 a significant new investment from Meta that values Scale at over $29 billion. Source
  • Lovo v. Lehrman (S.D.N.Y., 2025) quotes Bartz v. Anthropic (N.D. Cal., June 23, 2025): using the books at issue to train Claude was exceedingly transformative and a fair use under Section 107. Source
  • The European Commission's AI Act guidance says general-purpose AI model obligations, including a copyright policy and a public summary of training content, apply from 2 August 2025. Source

Request SourceX investor materials

SourceX shares company and commercial metrics with investors directly. Any offering is made only to verified accredited investors.

See if your company qualifies