Deal economics
Dataset price per record limitations: 7 fixes before comparing deals
By SourceX Editorial · Reviewed by Noah Loul ·
10 sections
Short answer
Dataset price per record limitations start with the ratio itself: the price buys a bundle of rights, access and services, while the count measures storage rows. Seven public AI licensing disclosures checked on 2026-10-26, including Reddit's $203.0 million of January 2024 arrangements, state values but no record counts. Before comparing deals, normalize the unit, usable scope, context, rights, term, bundled services and price evidence.
Value drivers
Dataset price per record
Key takeaways
- A price per record divides a negotiated bundle of rights, access and services by a storage count, so either side can move without any change in value.
- None of seven public AI licensing disclosures checked on 2026-10-26 states a record count; the values they do state are aggregates, project totals or segment revenue.
- Normalize seven terms in order: unit, usable scope, context, rights bundle, term and access, bundled services and price evidence.
- In the illustrative example, a 6.0x per-record gap becomes 1.33x per usable case-year and still needs a rights adjustment.
- Per-record comparison is legitimate only inside one standardized offer, where one agreement fixes the unit, rights and term.
What does a dataset price per record actually measure?
It measures a negotiated total divided by a storage count, a ratio of two things that rarely mean the same in two deals. The numerator buys a bundle: permitted uses, exclusivity, term, the access method and often preparation work. The denominator counts whatever the seller's system calls a record. Change the definition on either side and the ratio moves while the value stays where it was.
The decision this page supports is narrow and common: whether a quoted or reported per-record price can anchor a dataset valuation, a comparison between two data deals, or a diligence view of a data company's pricing. It can, but only after both sides of the ratio are normalized to the same unit, the same rights and the same kind of price evidence.
Useful evidence has four properties. It is a realized price rather than an asking price; it states what was licensed in units the buyer's task can use; it states the rights and the term; and it separates license value from services. Most public evidence fails at least two of these tests, as the disclosure check below shows, which is why this article builds a worksheet instead of a price table.
Why record counts make a weak denominator
Record counts are a weak denominator because a record is a storage choice, not a unit of value, and three routine adjustments can change the count by multiples.
Unit definition comes first. A support platform may store one customer issue as a ticket, a dozen messages, several status events and an attachment. One seller counts tickets, another counts messages, and their per-record prices differ by the messages-per-ticket ratio before anyone looks at quality. Convert both to the unit the buyer's task consumes, usually one complete case from request to outcome; the decision coverage method linked below shows how to count those cases without double counting.
Usable scope is second. Deduplication, filters for out-of-scope periods, redaction that empties a field and records missing an outcome all shrink what a buyer can use. A price spread over delivered rows flatters the seller whose rows contain the most filler.
Utility is third, and it is not a property of the count. In LESS (Xia and colleagues, ICML 2024), training on a 5 percent subset selected for a target capability often outperformed training on the full dataset across diverse downstream tasks. The same records were worth more or less depending on the task they were chosen for. A per-record price assumes every record carries the same value; selection research shows the buyer's task decides that.
Why the quoted price buys rights, access and services, not rows
The numerator is a bundle, and public documents show that the bundle varies more than the count does.
Reddit's S-1, filed in February 2024, describes data licensing arrangements entered into in January 2024 with an aggregate transaction price of $203.0 million, delivered through ongoing access to its data API plus quarterly data transfers, under contracts with an original expected duration of more than one year. That price buys access over time, which can include content created after signing, so no fixed record count could describe it.
Other terms move value without moving the count. Stack Overflow's May 6, 2024 announcement of its partnership with OpenAI paired API access with attribution to the Stack Overflow community inside ChatGPT. OpenAI's data partnerships page describes two tracks for contributed data: a public open-source archive anyone can use for model training, and private datasets for training proprietary models. The same records carry very different rights on each track.
Marketplaces price the bundle explicitly. On AWS Data Exchange, an offer's pricing defines the total price and the duration of the subscription, from 1 to 36 months, and a private offer can differ from the public offer on price, duration, payment schedule, data subscription agreement and refund policy. The unit of sale is a subscription under an agreement, not a row.
The rights side of a comparison therefore records six things for each deal: the uses granted (training, evaluation, retrieval), exclusivity, sublicensing, what happens to trained models after the term, attribution or display duties, and the access method. Two deals that differ on any of them are not comparable per record until the difference is priced or flagged.
What public AI licensing disclosures reveal: values without record counts
None of the seven public disclosures checked for this article states a record count, so none supports a price per record. Three state a deal or project value, one states only a segment total, and three state no value at all.
The check covered primary documents on each publisher's own domain that describe AI content or data licensing by a named company: SEC filings, a company press release and an AI developer's partnership page, retrieved on 2026-10-26. The table records what each one states.
Where values appear, they are measured in different accounting terms. Reddit's $203.0 million is an aggregate transaction price for multi-year arrangements, of which it expected at least $66.4 million to be recognized in 2024. Wiley's fiscal 2024 results attribute part of fourth-quarter revenue to a $23 million content rights project for training generative AI large language models, and its 10-Q for the quarter ended October 31, 2024 describes a second, $21 million project recognized as $17 million and $4 million in consecutive quarters. Shutterstock's 2024 annual report places its data offering, access to its content and metadata for machine learning and generative AI training, inside a data, distribution and services segment that earned $175.3 million, up 28 percent, with the data offering growing 15 percent. Its data revenue has no line of its own.
Excluded cases matter as much as the included ones. Announcements from two news publishers could not be retrieved on their own domains in this environment, so they are left out rather than cited secondhand. Deal values reported only by the press are excluded because no primary document confirms them, and marketplace list prices are excluded because they are asking prices. A larger sample could change the counts but not the finding: a document that states a value without a unit count cannot be turned into a price per record.
| Disclosure | Value stated | Unit count stated | Term or delivery stated | Per-record price computable? |
|---|---|---|---|---|
| Reddit S-1, filed February 2024 | $203.0 million aggregate for January 2024 arrangements | No | API access plus quarterly transfers; duration over one year | No |
| Reddit 2024 annual report | No deal value; substantially all contract value from two partners | No | Renewal risk noted; terms not given | No |
| Wiley fiscal 2024 results release | $23 million content rights project | No | No | No |
| Wiley 10-Q, quarter ended October 31, 2024 | $21 million project, recognized over two quarters | No | Recognition timing only | No |
| Shutterstock 2024 10-K | Segment total only: $175.3 million, data plus distribution and services | No | No | No |
| Stack Overflow and OpenAI release, May 6, 2024 | No | No | API access with attribution in ChatGPT | No |
| OpenAI data partnerships page | No | No | Public archive or private training datasets | No |
The rights-normalized valuation worksheet
Normalize seven terms in order, recording each one for both deals, and compare only the price that survives all seven.
Fill the worksheet from documents rather than descriptions: the signed license, a schema sample, the deduplication report and the delivery schedule. Where a value is an assumption, label it as one in the cell. The order matters because later adjustments depend on earlier ones; term is expressed per usable unit, so usable scope has to be settled before term can be.
Rights rarely reduce to a number. When two deals differ on uses, exclusivity or post-term model rights, either price the extra right separately with a stated assumption or mark the pair not comparable. Marking it is the honest default, and it is also a finding: it tells a buyer which right it would have to purchase twice.
| Term to normalize | Record for each deal | Normalization rule | Evidence that settles it |
|---|---|---|---|
| 1. Unit | What one record is: message, row, ticket, case or document | Convert both deals to the buyer's task unit, such as one complete case | Schema sample and a unit crosswalk |
| 2. Usable scope | Share left after deduplication, filters and redaction | Divide price by usable units, not delivered rows | Deduplication report and filter log on the delivered set |
| 3. Context | Share of units carrying every field and link the task needs | Count only complete units for the target task | Field-completeness audit on a random sample |
| 4. Rights bundle | Uses, exclusivity, sublicensing, post-term model rights, attribution | Compare like bundles; price extra rights separately or mark not comparable | Signed license text, not a term sheet summary |
| 5. Term and access | One-time copy, refreshes or API access, and duration | Express price per usable unit per year of access | Delivery schedule in the contract |
| 6. Bundled services | Preparation, de-identification, delivery, evaluation support | Strip services out before comparing license value | Line items, or a services estimate labeled as an assumption |
| 7. Price evidence | Asking, contracted, recognized or press-reported | Compare realized with realized; drop press-only figures | Invoice, revenue recognition note or filing |
Worked example (Illustrative): a 6x gap that is not a gap
In this illustrative case, a quote that looks six times more expensive per record ends up a third more expensive per usable case-year of license value, and still not comparable on rights.
Illustrative: two invented deals, described with counts and ratios only. Deal A licenses 50,000 records, each one a complete case, at half of Deal B's total price. Deal B licenses 600,000 rows in which every case spans six rows. On the raw count, A's price per record is 6.0 times B's.
Converting to complete cases gives B 100,000 cases and brings the ratio to 1.0. Usable scope is 90 percent for A and 60 percent for B once duplicates and cases without outcomes are removed, which makes A the cheaper deal at 0.67. A runs two years and B three, so per usable case-year the ratio returns to 1.0. A quarter of B's price covers preparation that A's buyer does separately; stripping it out leaves A at 1.33 times B.
The last row cannot be computed. A permits training only, while B also permits evaluation, so a buyer that needs both must price a second license for A before the comparison is complete. The ratio moved from 6.0 to 0.67 and back above 1.0 on adjustments any diligence team can check. The raw figure carried no information about relative value.
| Step | Deal A | Deal B | A's price per unit relative to B's |
|---|---|---|---|
| Quoted per record | 50,000 records at half B's total price | 600,000 rows | 6.0x |
| Common unit: complete cases | 50,000 cases, one row each | 100,000 cases, six rows each | 1.0x |
| Usable scope | 45,000 usable cases (90 percent) | 60,000 usable cases (60 percent) | 0.67x |
| Term: per usable case-year | 2 years of access | 3 years of access | 1.0x |
| Bundled services removed | License only | A quarter of the price is preparation | 1.33x |
| Rights bundle | Training only | Training and evaluation | Not comparable until the extra use is priced |
Are asking prices a fair shortcut? Testing the strongest alternative
No. The strongest case for per-record prices is that markets settle on them anyway, but asking prices, public ranges and contract totals are not realized, market-clearing prices.
The argument deserves its best form. If buyers and sellers routinely quote per record, a per-record benchmark works as a coordination device. Inside one standardized offer, where the unit, rights bundle and term are fixed by the same agreement, comparing price per record is legitimate; two tiers of one marketplace product under one subscription agreement are an example.
Outside that case the shortcut fails on the evidence. AWS's own documentation lets a private offer differ from the public offer on price and terms, so a listed price may not be the price paid. Contract totals are not annual prices either: Reddit expected at least $66.4 million of its $203.0 million aggregate to be recognized in the first year, and Wiley recognized its $21 million project across two quarters. The OECD's 2022 paper on measuring the value of data describes market-based, cost-based and income-based approaches and finds that market statistics capture only a small portion of data's value, which is another way of saying that comparable transactions are scarce.
Ranges that data programs publish for what they may pay have the same limit. A range describes an offer, not a transaction, and says nothing about the scope, rights or units behind any amount actually paid.
Risks that could make a rights-normalized comparison wrong
Five things could make a normalized comparison wrong, and an investor should know which one a valuation leans on before trusting it.
At a glance
Risks that could make a rights-normalized comparison wrong
- Thin comparables: with few public transactions, two or three deals set the benchmark, and the OECD finds that market statistics capture only a small portion of data's value.
- Buyer concentration: by Reddit's own account in its 2024 annual report, two partners supply substantially all of its licensing contract value, and the company warns that those deals could lapse at renewal or come back on worse terms. A price set by two buyers may not hold.
- Task drift: utility follows the buyer's task, as selection research such as LESS shows, so a normalization built for one task can misprice the same records for another.
- Bundled reporting: when a filer reports data revenue inside a wider segment, as Shutterstock does, even the numerator cannot be isolated from public documents.
- Assumption creep: services estimates and usable-scope rates are often assumptions. Label each one and test how far it can move before the comparison flips.
Questions for any data company quoting per-record value, SourceX included
Ask for the evidence behind every per-record figure, using your own definitions rather than the seller's. Put the same questions to SourceX.
SourceX is one application of this question. It turns a company's selected business records into datasets licensed for AI use, running the sourcing and qualification, rights review, contracting, data preparation, delivery and payment in between. Its rights bundle is written down before anything moves: the supplying company approves the scope and terms, and executes the agreement, before delivering any data, and from then on SourceX's right to use, sublicense and distribute the de-identified dataset reaches only as far as that agreement allows. Each deal's signed agreement also fixes its exclusivity and term. Those are the rights-bundle rows the worksheet asks every seller to document.
SourceX publishes no price list or per-record prices; its company and commercial metrics are provided to investors directly. Because its raise runs under Rule 506(c), every purchaser must be an accredited investor. Investors can request materials and bring a worksheet question, such as how usable scope is measured before a price is agreed.
- What is one record in this figure, and how many records make one complete case for the buyer's task?
- What share of delivered units survived deduplication, filters and redaction?
- Is the figure an asking, contracted or recognized price, and over what term?
- Which uses, exclusivity, sublicensing and post-term model rights does the price include?
- How much of the price pays for preparation, de-identification or delivery work?
- How many independent buyers paid prices in this range, and over what period?
Methods note: the disclosure check, its sampling frame and what was not measured
The seven-term worksheet, the price-evidence categories and the illustrative example are editorial frameworks built for this article on 2026-10-26. The disclosure check used primary documents on each publisher's own domain: Reddit's S-1 (February 2024) and 2024 annual report, Wiley's fiscal 2024 results release and its 10-Q for the quarter ended October 31, 2024, Shutterstock's 2024 10-K, Stack Overflow's May 6, 2024 release and OpenAI's data partnerships page. The sampling frame was named-company AI licensing disclosures retrievable on the publisher's domain on that date; press-only deal values, marketplace list prices and two publisher announcements that could not be retrieved were excluded.
Method sources are AWS Data Exchange documentation, the OECD's 2022 paper on measuring the value of data and LESS (ICML 2024). All were verified through search excerpts on their own domains, and a reviewer should open each before release.
Not measured: any realized price per record, at SourceX or elsewhere; the record counts behind any public deal; the value of specific rights such as exclusivity; and how often private offers differ from public ones. The worked example's counts, shares and ratios are invented to show the arithmetic and describe no real deal.
Frequently asked questions
How much do AI companies pay per record for data?
There is no public per-record benchmark. Filings such as Reddit's S-1 and Wiley's results state deal or project values but not record counts, and marketplace list prices are asking prices that private offers can change. Any per-record figure needs its unit, usable share, rights, term and price type stated before it means anything.
Is price per record ever a valid way to compare datasets?
Yes, inside one standardized offer where the unit, rights bundle, term and services are fixed by the same agreement, such as two tiers of one marketplace product. Across different deals it is valid only after those terms are normalized, and rights differences may still need to be priced separately.
Why does exclusivity change a per-record comparison?
Exclusivity changes what the buyer gets without changing the count, because it stops the seller licensing the same records to others for the restricted scope and period. Two prices for the same records, one exclusive and one not, are prices for different products. Compare like bundles or price the exclusivity separately.
What is the difference between contract value and revenue in data licensing?
Contract value, or aggregate transaction price, is what a contract is expected to bring over its whole term; revenue is the part recognized in a period as the seller performs. Reddit expected at least $66.4 million of a $203.0 million aggregate in its first year. Dividing a multi-year total by records and comparing it with an annual price mixes two periods.
How should an investor test a data company's per-record pricing claims?
Ask for the unit definition, the usable share after deduplication and filters, the price type, the term, the rights bundle and the services included, then rebuild the figure yourself. Ask how many independent buyers paid prices in the claimed range. A figure that cannot survive those questions is a marketing number, not a valuation input.
Sources
- Reddit's S-1 reports data licensing arrangements entered into in January 2024 with an aggregate transaction price of $203.0 million, delivered through data API access and quarterly transfers under contracts expected to last more than one year, with at least $66.4 million expected to be recognized in 2024. Source
- Reddit's 2024 annual report says substantially all of the contract value associated with its licensing revenue is derived from two partners, and that the arrangements may not be renewed or may be renewed on less favorable terms. Source
- Wiley's fiscal 2024 results release attributes part of fourth-quarter revenue to a $23 million content rights project for training generative AI large language models. Source
- Wiley's 10-Q for the quarter ended October 31, 2024 describes a $21 million content rights project for training generative AI models, with $17 million recognized in the quarter ended July 31, 2024 and $4 million in the quarter ended October 31, 2024. Source
- Shutterstock's 2024 10-K reports data, distribution and services revenue of $175.3 million, up 28 percent, with data offering growth of 15 percent; the data offering provides access to content and metadata for machine learning and generative AI model training. Source
- Stack Overflow announced on May 6, 2024 that OpenAI would use its OverflowAPI product and provide attribution to the Stack Overflow community within ChatGPT. Source
- OpenAI's data partnerships page describes an open-source archive that would be public for anyone to use in model training and private datasets for training proprietary models, and says it is not seeking sensitive or personal information or information belonging to a third party. Source
- AWS Data Exchange private offers can differ from the public offer on price, duration, payment schedule, data subscription agreement or refund policy. Source
- AWS Data Exchange offer pricing defines the total price and duration of a subscription, with durations from 1 to 36 months. Source
- The OECD's 2022 paper Measuring the value of data and data flows describes market-based, cost-based and income-based approaches to valuing data and finds that market statistics capture only a small portion of data's value. Source
- LESS (Xia and colleagues, ICML 2024) reports that training on a 5 percent subset selected for a target capability often outperforms training on the full dataset across diverse downstream tasks. Source
Request SourceX investor materials
SourceX shares company and commercial metrics with investors directly. Any offering is made only to verified accredited investors.