Skip to content

Deal economics

Unique data investment value: six tests that separate rare from useful

By SourceX Editorial · Reviewed by Noah Loul ·

10 sections

Short answer

Unique data has investment value only when its uniqueness shows up as measured improvement on a task a buyer pays for, at a cost no substitute undercuts. Novelty matters most for private evaluation sets; in training, repeated, curated or synthetic data often replaces it. Score task lift and legal scope as gates, then novelty, replacement cost, preparation burden and repeat buying, each with a test that could prove the claim wrong.

Value drivers

Unique data investment value

This page covers unique data investment value: six tests that separate rare from useful.Illustrative

Key takeaways

  • Unique describes the data; useful describes what it does for a buyer's task, and investors should price the second.
  • Data-constrained scaling research found up to four epochs of repeated data came close to unique data, capping the premium on raw uniqueness in pretraining.
  • Novelty earns most in evaluation: GSM1k's fresh problems exposed accuracy drops of up to 8 percent for leading models.
  • Curated and partly synthetic data can rival rarer data on narrow tasks, as phi-1 showed for code, so replacement cost is a core test.
  • Repeat buying by distinct buyers is the best observable test of usefulness, and Reddit's fiscal 2025 Form 10-K shows how concentrated licensing revenue can be.

Is unique data the same as data with investment value?

No. Unique describes the data's relationship to the rest of the world: nobody else holds it and it is absent from public corpora. Useful describes its relationship to a buyer's task: adding it measurably improves a result the buyer pays for, compared with the best alternative. Unique data investment value exists only where the two overlap, and only to the extent the improvement costs less to get from this data than from anywhere else.

Pitch decks lead with uniqueness because it is easy to assert and hard to check. An investor's job is to turn the assertion into three measurements: how much of the data is genuinely unavailable elsewhere, how far it moves a defined task, and how much work it takes to put into a form a buyer can use. A dataset can earn top marks on the first and nothing on the second.

This page gives a scorecard for those measurements, each paired with a test that could disprove the claim, along with the public evidence on when novelty pays and when something cheaper replaces it. It applies to any company claiming a data advantage: a data supplier, a vertical software company sitting on years of workflow records, or a marketplace.

Why buyers ask for non-public data but pay for results

Non-public status gets a dataset into the conversation; it does not set the price. OpenAI's data partnerships page seeks large-scale datasets that the public cannot already reach easily online, and the company's statement on its approach to data says it does not pursue paid partnerships for purely publicly available information. Public availability is close to a disqualifier for a paid deal.

A stated preference is still the weakest kind of demand evidence. A request for non-public data says what a buyer will look at, not what it will renew. The price of a dataset follows from what it changes in the buyer's model or product, and from what the buyer would otherwise spend to get the same change. Those two quantities, lift and replacement cost, are where uniqueness either converts into value or quietly does not.

When does novelty pay, and when do repetition or synthesis replace it?

Novelty pays most in evaluation and least in bulk pretraining, with targeted training in between. Three lines of public research point the same way.

In pretraining, uniqueness has close substitutes. Muennighoff and colleagues, studying data-constrained language models in 2023, found that under a fixed compute budget, training for up to four epochs on repeated data changed loss very little compared with training on unique data, and that returns from further repetition eventually fell toward zero. A buyer short of tokens can reuse what it already has several times before new ones matter much.

In targeted training, quality and fit can beat rarity. The phi-1 team trained a 1.3 billion parameter code model on about 6 billion tokens of filtered, textbook-quality web data plus about 1 billion tokens of synthetic textbooks and exercises, and reported 50.6 percent pass@1 on the HumanEval benchmark. The data that mattered was chosen for usefulness, and part of it was generated rather than collected.

In evaluation, novelty is the product. On GSM1k, a set of new grade-school math problems researchers wrote to mirror the public GSM8k benchmark, leading models lost as much as 8 percent of their accuracy, according to the paper's current version, evidence of overfitting to the public set. The authors chose not to release GSM1k publicly so it would not be contaminated, which exposes the catch: unseen data keeps its value only while it stays unseen. The table summarizes where uniqueness earns its keep.

When does novelty pay, and when do repetition or synthesis replace it?
Buyer jobWhat uniqueness addsWhat can replace itPublic evidence
Bulk pretrainingMore distinct tokens once existing stock has been reused several timesRepeating existing data; other public or licensed corporaMuennighoff et al., 2023: up to four epochs of repeats came close to unique data
Targeted fine-tuningExamples of decisions or domains models have not seenFiltered public data, synthetic examples, commissioned workphi-1, 2023: curated and synthetic data reached 50.6 percent pass@1 on HumanEval
Held-out evaluationTests a model cannot have memorizedLittle, while the set stays privateGSM1k, current version: accuracy drops of up to 8 percent against the public benchmark
Grounding a product in current factsRecords of what is happening now in a market or workflowThe buyer's own users and logs, where it has themNot measured in the sources used here; treat as a hypothesis

The unique-versus-useful scorecard: gates, size and durability

Score six tests separately and combine them with a rule rather than an average. Two tests are gates: if task lift or legal scope fails, there is no uniqueness premium, however novel the data. Two set the size of any premium: novelty and replacement cost. Two set its durability: preparation burden and repeat buying. Each test carries a disconfirmation test, the result that would prove the claim wrong, so the scorecard can be falsified rather than merely filled in.

Score each test 0 (fails or is contradicted), 1 (unproven) or 2 (passes with evidence). A 0 on either gate means no premium. Above the gates, a 2 on both novelty and replacement cost makes the data a moat candidate; durability scores below 2 mean the advantage needs retesting whenever the buyer's models or alternatives change. The table defines each test and the result that would disconfirm it.

The unique-versus-useful scorecard: gates, size and durability
TestRoleWhat it measuresEvidence to requestDisconfirmed if
Task liftGateImprovement on a buyer-defined task against the best available substitute dataA held-out evaluation with and without the data, reported with intervalsThe interval includes zero, or a substitute matches the lift
Legal scopeGateA documented right to license for training or evaluation, for the uses and term soldCustomer contracts, notices and license documents for each record familyAny family is limited to service delivery, or its license is missing
NoveltySizeShare of records absent from public corpora and unavailable from other holdersOverlap tests against public corpora and a list of other plausible holdersA material share appears publicly, or a rival can license equivalent records
Replacement costSizeWhat a buyer would spend to get the same lift another wayQuotes or pilots for commissioning, synthetic generation or direct sourcingA substitute reaches the same lift at lower cost or in less time
Preparation burdenDurabilityWork needed to turn raw records into a usable, rights-cleared deliveryHours per usable record and rejection rates from a pilot preparationPreparation cost approaches the replacement cost
Repeat buyingDurabilityRenewals and second purchases by distinct buyersCount of distinct buyers, renewals and the largest buyer's share of valueValue comes from one buyer or from one-off purchases

What evidence settles each test, and which claims are easy to fake?

Task lift is the test most often skipped and the hardest to fake. Ask for an ablation: the same model trained or evaluated with and without the dataset, on tasks the buyer defined before seeing results, reported with intervals rather than a single number. Data valuation methods can sharpen the picture. Data Shapley, proposed by Ghorbani and Zou in 2019, assigns each training point a value for its contribution to a model's performance; in their experiments, low-value points tended to be outliers and corrupted labels, while high-value points suggested what new data to acquire.

Legal scope is the easiest test to pass on paper. On popular dataset hosting sites, the Data Provenance Initiative's review of more than 1,800 text datasets found licenses missing more than 70 percent of the time and mislabeled at error rates above 50 percent (Nature Machine Intelligence, 2024). A license label is a claim, not evidence; ask for the underlying contracts and notices behind each record family.

Novelty and replacement cost have to be read together. Overlap tests show how much of the data is already public; replacement quotes show what it would cost to get the same effect without it. A dataset can be entirely novel and still cheap to replace if the buyer needs only its pattern rather than its particular records, which is the kind of gap the phi-1 work suggests curated and synthetic data can fill.

Repeat buying is the closest observable proxy for usefulness, because a buyer renews only what kept working. Ask for distinct buyers and renewals rather than total contract value, and for the share of value that comes from the largest buyer.

Illustrative: a rare archive and a useful one, scored side by side

Two invented datasets show how the gates change the answer. Dataset R is two decades of calibration logs from a specialty instrument maker, never published and held by no one else. Dataset U is five years of customer support conversations from a mid-size software company, similar in form to public support text but with resolution outcomes and follow-up records attached.

Dataset R scores 2 on novelty and 2 on replacement cost, but 0 on task lift, because no buyer task in the pilot improved, and 0 on repeat buying. It fails a gate, so its rarity carries no premium today; it is an option on a future buyer, not a moat. Dataset U scores only 1 on novelty, since much of its language resembles public data, and 1 on replacement cost, but 2 on task lift, because its outcomes improved an agent's resolution evaluation, 2 on legal scope and 2 on repeat buying. Both score 1 on preparation burden, and R scores 1 on legal scope.

U passes both gates with a modest size score: useful, and defensible only to the degree its outcome records are hard to recreate. The figure plots both datasets on the six tests. The scores are illustrative, chosen to show the method, and were not measured on any real company.

A dumbbell chart scoring two illustrative datasets from 0 to 2 on six tests, with the rare dataset failing the task lift gate and the useful dataset passing both gates.
Illustrative: scores for invented Dataset R (rare) and Dataset U (useful); task lift and legal scope are gates.

Could useful data still be a weak investment?

Yes, and this is the strongest alternative to the usefulness-first view. Useful data may still be easy to substitute or cheaper to source directly. If a buyer can approach the original record holders itself, commission equivalent work, or generate synthetic examples once it knows which pattern it needs, an intermediary's usefulness is real but its margin is temporary.

Usefulness can also be absorbed. A dataset that teaches a capability may be bought once; after the buyer's model has learned it, the next purchase has to teach something new. That is why repeat buying sits in the scorecard: it separates data that keeps paying from data that paid once. The investor question becomes whether a company can keep supplying the next useful set, which turns a data question into a sourcing and rights question.

What could make the unique-data thesis wrong?

Five risks belong in any investment memo on a data business, each traceable to public evidence where it exists. None of them argues that data has no value; each argues that uniqueness alone is the wrong thing to pay for.

  • Buyer concentration. Reddit's Form 10-K for fiscal 2025 states that "substantially all of the contract value associated with our licensing revenue is derived from two of our partners." Even data with proven demand can depend on very few buyers.
  • Curated and synthetic substitutes. The phi-1 results show that filtered and generated data can perform strongly on a narrow task, which lowers the replacement cost of any dataset whose value lies in a learnable pattern.
  • Repetition and diminishing returns. The data-constrained scaling results suggest buyers can reuse existing data several times before new unique tokens make much difference, which caps the premium on volume alone.
  • Decay of novelty. GSM1k's authors withheld their set to keep it uncontaminated; once unique data circulates, its evaluation value falls, and a license that allows broad redistribution can erode the novelty it sells.
  • Rights that do not hold. With license metadata often missing or wrong across public datasets, a uniqueness claim built on unclear rights may not survive diligence, let alone a dispute.

Questions to ask any data company about novelty and lift, SourceX included

Ask questions that force a measurement rather than an adjective. The same list applies to SourceX. It converts selected business records into licensed AI datasets, carrying each one from sourcing and qualification and a rights review through contracting, preparation, delivery and payment. It also keeps nothing in stock: AI buyers' datasets are sourced from U.S. businesses when requested, so the useful questions concern how each sourced dataset is tested, not the size of an inventory.

SourceX has not published results on these tests. It shares company and commercial metrics with investors directly, and it is raising capital from accredited investors under Rule 506(c), with investor materials shared on request through its investors page. Ask it the same questions.

Checklist · 6 items

  1. Which buyer tasks has the data been evaluated on, against which substitute, and with what interval?
  2. What share of records was tested for overlap with public corpora, and by what method?
  3. For each record family, which documents establish the right to license, for which uses and term?
  4. What would it cost a buyer to get the same lift by commissioning, generating or sourcing directly?
  5. How many distinct buyers have renewed or bought a second time, and what share of value comes from the largest?
  6. How is preparation effort measured, and what share of raw records is rejected?

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

Methods note: sources and limits of the unique-versus-useful scorecard

The scorecard is an editorial framework built for this article on 2026-10-26. Its tests draw on the evidence cited above: data-constrained scaling and phi-1 for substitution, GSM1k for the value of unseen evaluation data, Data Shapley for measuring contribution, the Data Provenance Initiative audit for legal scope, OpenAI's published data partnership criteria for stated demand, and Reddit's fiscal 2025 Form 10-K for buyer concentration. Every source was checked on its publisher's own site on that date.

The scorecard does not measure prices, deal values or any company's actual data. The gate and size rules are judgments rather than estimates fitted to outcomes, and the 0-to-2 scale is ordinal, so scores should not be added or averaged across datasets. Dataset R and Dataset U are illustrative, no SourceX data was used, and SourceX has not published results for this scorecard.

Frequently asked questions

What is the difference between unique data and useful data?

Unique data is unavailable elsewhere; useful data measurably improves a task a buyer pays for, compared with the best alternative. The two overlap less than pitch decks suggest, and investment value sits only in the overlap, sized by what it would cost a buyer to get the same improvement another way.

Is proprietary data automatically a moat?

No. Proprietary data becomes a moat candidate only after it passes two gates, measured task lift and documented legal scope, and then shows high novelty and high replacement cost. Repeat purchases by distinct buyers are the best evidence that the advantage lasts.

How can an investor check whether a dataset improves a model?

Ask for an ablation on tasks defined before the results: the same model with and without the data, against the best substitute, reported with intervals. Data valuation methods such as Data Shapley can show which records contribute and which add noise.

Can synthetic data replace unique business records?

Sometimes. Where the value lies in a learnable pattern, curated and generated data can substitute, as the phi-1 code model showed. Synthetic data is a weaker substitute for private evaluation sets, whose value depends on models never having seen them.

Does SourceX publish evidence on the usefulness of its datasets?

No. Task-lift, deal and commercial metrics are not public; SourceX provides its company and commercial figures to investors directly. Materials are sent on request through sourcex.si/investors, and any offering is open solely to verified accredited investors.

Sources

  • OpenAI's data partnerships page says it is interested in large-scale datasets that reflect human society and are not already easily accessible online to the public, and excludes sensitive personal information and third-party data. Source
  • OpenAI's statement on its approach to data and AI says it uses partnerships to access non-public content and does not pursue paid partnerships for purely publicly available information. Source
  • Muennighoff et al. (2023) found that under a fixed compute budget, training for up to four epochs on repeated data yields negligible changes in loss compared with unique data, with returns from further repetition diminishing toward zero. Source
  • phi-1, a 1.3 billion parameter code model trained on about 6 billion tokens of textbook-quality web data and about 1 billion tokens of GPT-3.5-generated textbooks and exercises, reported 50.6 percent pass@1 on HumanEval. Source
  • The current version of the GSM1k paper reports accuracy drops of up to 8 percent for leading LLMs on fresh problems mirroring GSM8k, and the authors do not intend to release GSM1k publicly to avoid contamination. Source
  • Data Shapley (Ghorbani and Zou, 2019) values each training datum by its contribution to predictor performance; in experiments, low-value points captured outliers and corrupted labels and high-value points suggested data to acquire. Source
  • The Data Provenance Initiative audit of more than 1,800 text datasets (Nature Machine Intelligence, 2024) found license omission rates above 70 percent and error rates above 50 percent on popular dataset hosting sites. Source
  • Reddit's Form 10-K for fiscal 2025 states that substantially all of the contract value associated with its licensing revenue is derived from two of its partners. 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