Skip to content

Deal economics

AI dataset pricing evidence: what 16,147 data listings cannot tell you

By SourceX Editorial · Reviewed by Noah Loul ·

10 sections

Short answer

AI dataset pricing evidence comes in five grades, and only the last two are transactions: marketing ranges, posted asking prices, indicative offers, signed contracts and paid deals. The public record sits mostly in the lower grades: in 16,147 data-marketplace listings, 47.8 percent of paid listings show no price, and no field records exclusivity or permitted use. Value a dataset on signed or paid deals whose rights match.

Value drivers

AI dataset pricing evidence

This page covers AI dataset pricing evidence: what 16,147 data listings cannot tell you.Illustrative

Key takeaways

  • AI dataset pricing evidence comes in five grades: marketing ranges, posted asking prices, indicative offers, signed contracts and paid deals; only the last two are transactions.
  • In 16,147 public data-marketplace listings, 47.8 percent of the 9,441 paid listings show no public price, and visibility ranges from 0 to 100 percent by venue.
  • None of the listing dataset's 14 fields records exclusivity, permitted use or term, so posted prices cannot be rights-normalized from listing data.
  • Admit a price into a valuation range only as a signed or paid deal with stated rights that match the target; an illustrative naive average misses an exclusive-training range by 4 to 18 percent and an evaluation comparable by 146 percent.
  • Ask data companies, SourceX included, for price evidence by grade and for the share of signed deals that reach acceptance and payment.

Which AI dataset pricing evidence can an investor rely on?

Signed and paid transactions whose rights, scope and preparation are known. Everything else, from a marketing range to a posted list price to an indicative offer, can frame a negotiation but cannot tell you what a dataset clears at. The useful skill in AI dataset pricing evidence is grading each number before it is averaged with any other.

The problem is that the public record is lopsided. Asking prices are easy to collect because sellers publish them; transactions are hard to see because both sides usually keep them confidential. A valuation that blends the two gives the asking side most of the weight, simply because it supplies most of the numbers.

This page sets out a five-grade ladder for price evidence, measures how much price information one large public collection of data-marketplace listings actually contains, and gives a comparable record and an illustrative worksheet that admit a number only once its grade and rights are known. The thesis an investor can test is simple: in data licensing, the scarce asset is not a price list but a record of closed deals with their rights attached, and companies that generate that record are easier to value than companies that quote ranges.

Five grades of price evidence, from a range to cash received

Grade every price observation by how close it sits to money actually changing hands for a defined set of rights. The higher the grade, the more weight the number can carry; the lowest two grades describe what someone hopes for, not what anyone accepted.

Grades 4 and 5 are both transactions, but they are not equal. A signed contract can still fail acceptance, shrink at delivery or end early; a paid deal has survived those events. Keep them in separate columns, and report how much of any headline figure is still waiting on a condition.

Five grades of price evidence, from a range to cash received
GradeTypical sourceWhat it provesWhat it cannot proveUse in a valuation
1. Marketing range or estimateSeller websites, estimators, published payout rangesWhat a seller is willing to suggestThat any buyer would pay itContext only; never averaged
2. Posted asking priceMarketplace listings with a list priceA price at which a seller says it will transactWhether anyone bought, or at what discountOne side of a spread; a ceiling at most
3. Indicative offerBids, term sheets, non-binding lettersA buyer's interest at a stated levelThat terms survive diligence and draftingAn indication; not in the comparable range
4. Signed contractExecuted licenses with fixed termsAgreed price for stated rights and scopeThat delivery, acceptance and payment followA comparable once rights are normalized; flag conditions
5. Accepted and paid dealClosing records, invoices, recognized revenueMoney received for delivered, accepted dataValue of different rights or a different datasetThe anchor of the range, after rights normalization

What 16,147 data marketplace listings show about public prices

That posted prices cover a minority of paid products, never state the rights behind them, and include no transactions at all. DaDaDa, a public dataset built by researchers from 16,147 data-product listings on nine marketplaces, records each listing's pricing mode as labelled by its authors. Of those listings, 6,706 (41.5 percent) are free, 4,510 (27.9 percent) carry no public price because buyers must negotiate, and 4,931 (30.5 percent) show a posted subscription, one-off or usage-based price.

Among the 9,441 paid listings, 47.8 percent show no public price. Whether a price is visible depends heavily on the venue. Every paid listing on Shanghai Data Exchange and Databricks in the dataset is negotiation-only, as are 98.2 percent on Canton Data Exchange and 90.3 percent on Guiyang; Datarade sits at 54.8 percent; AWS Data Exchange, Snowflake and Beijing show a price on every paid listing, and Zhejiang on all but 5 of 838.

Just as telling is what the listings do not record. The dataset's 14 fields cover category, coverage, sample, description, dimension, platform, price, pricing mode, provider, size, title, update frequency, URL and volume. None records exclusivity, permitted use, license term or territory, the terms that decide what a price buys. In the listing text on the four English-language marketplaces, the words exclusive or exclusivity appear in 189 of 6,633 paid listings (2.8 percent), and the first eight of those, read in file order, use the word for geography, promotions, product descriptions or data content, not to grant exclusive license rights. Training phrases such as training data or model training appear in 140 (2.1 percent).

These listings are mostly market, location and financial data products, not AI training licenses, so the shares will not transfer one for one. The structure does: public prices exist for a minority of paid products, and none comes with the rights needed to compare it.

What 16,147 data marketplace listings show about public prices
Marketplace (as named in DaDaDa)ListingsPaid listingsPaid, no public priceShare of paid with no public price
Datarade4,3964,3962,41054.8%
AWS Data Exchange3,7021,60600.0%
Snowflake2,58523700.0%
Databricks2,004394394100.0%
Zhejiang Big Data Exchange83883850.6%
Canton Data Exchange79279277898.2%
Beijing International Data Exchange71420400.0%
Guiyang Global Big Data Exchange66952747690.3%
Shanghai Data Exchange447447447100.0%
All nine marketplaces16,1479,4414,51047.8%
A dot plot placing nine data marketplaces on a 0 to 100 percent scale by their share of paid listings with no public price, from 100 percent at Shanghai and Databricks to 0 percent at AWS, Snowflake and Beijing, with a dashed line at 47.8 percent for all nine.
Share of paid listings with no public price, by marketplace, in the DaDaDa dataset of 16,147 listings. Source: ZJU-DIVER/DaDaDa, counted 2026-10-28.

Why a posted price is not a clearing price, and the case that it is close enough

Because an asking price is one side of a negotiation, and the other side is usually invisible. Fair value accounting treats this explicitly: IFRS 13 paragraph 70 says, in paraphrase, that when an item has both a bid price and an ask price, the price within the bid-ask spread most representative of fair value in the circumstances is used. An ask on its own sets a boundary, not the answer.

The same standard lists, among signs that activity in a market has fallen significantly, that "there is a wide bid-ask spread or significant increase in the bid-ask spread". Data markets rarely publish bids at all, so the width of the spread cannot even be observed, which is a reason to treat a posted price as an upper boundary rather than an estimate.

The strongest objection is that asking prices are close enough: sellers anchor near what buyers accept, and a large sample averages out noise. Three things in the listing data cut against it. First, visibility tracks the venue more than the product: some marketplaces show a price on every paid listing and others on none. Second, the posted prices do not form one market; the median posted price is $10,000 for a 12-month subscription (2,854 listings), $194.20 for a one-off purchase (1,450) and $0.15 per use for usage-based access (625). Third, without rights fields, two posted prices for similar records may describe different grants entirely.

Asking prices still have a use. They show what sellers believe and set a ceiling on what a rational buyer would pay for a similar, posted product. They become evidence of value only when matched to a transaction for comparable rights.

The comparable record: what to fill in before a price enters the range

Before any observation enters a valuation range, fill in a comparable record that states its grade and every term that changes what the price buys. An observation with blank rights fields stays out of the range, however tempting its number.

At a glance

The comparable record: what to fill in before a price enters the range

Tax law faces the same problem when it prices intangibles between related companies, and its test is a useful model. Under the US transfer pricing rules, the comparable uncontrolled transaction method "will generally be the most direct and reliable measure of the arm's length result for the controlled transfer of an intangible", and its reliability "is affected by the completeness and accuracy of the data used and the reliability of the assumptions made to apply this method." The regulation also points, in paraphrase, to contractual terms and economic conditions as factors that can materially change the price. A dataset comparable needs the same completeness.

Record these fields for every observation:

  • Grade and date: which of the five grades, when the number was set, and the source document.
  • Rights granted: exclusive or non-exclusive, permitted uses (training, evaluation, both), term, territory and any sublicensing.
  • Scope delivered: records, years of history, fields, and the share of records that proved usable after review.
  • Preparation included: de-identification, labeling or linking, and who paid for them.
  • Buyer purpose and urgency: what capability the buyer wanted and whether timing pressure shaped the price.
  • Conditions and payment status: acceptance tests, milestones, holdbacks, and what has actually been paid.
  • Relationship and circumstances: related parties, distressed sellers, bundled services or refresh rights that make the price unlike an ordinary sale.
  • Normalization note: the adjustment, if any, that maps the observation onto the target dataset's rights and scope, with the reasoning written out.

Illustrative: one evidence set, two answers

The same nine observations give two very different answers depending on the rights being valued, and a naive average fits neither. Every value below is illustrative, in arbitrary units, and describes no real deal or company.

The rule admits an observation as a comparable only if it is a signed contract or a paid deal, its rights are stated, and those rights match the target's. Indicative offers are reported as indications only; ranges, listings and observations with unstated rights stay out.

For an exclusive training license, two observations qualify, giving a range of 90 to 105 units, of which only the 90-unit deal is paid; the 105-unit contract still awaits acceptance. For a non-exclusive evaluation license, one observation qualifies, at 35 units, which is too thin for a range and should be reported as such. The naive mean of all eight priced observations is 86.25 units: 4 percent below the low end of the exclusive range and 18 percent below its high end, and 146 percent above the only evaluation comparable. It is wrong in both directions at once, which is the point of grading first.

Illustrative: one evidence set, two answers
Observation (illustrative)GradeRights statedPrice, illustrative unitsAdmitted as a comparable for
Marketing range midpoint1Not stated160Neither; grade 1
Posted marketplace listing2Non-exclusive, use not stated45Neither; grade 2
Posted marketplace listing2Not stated70Neither; grade 2
Listing with no public price2Not statedNo public priceNeither; no number
Indicative offer3Exclusive training130Neither; indication only
Signed contract, acceptance pending4Exclusive training105Exclusive training, contingent
Signed contract4Non-exclusive evaluation35Non-exclusive evaluation
Paid deal5Exclusive training90Exclusive training
Paid deal5Not stated55Neither; rights not stated

Applying the grades to SourceX's own process

SourceX's published process maps onto the same ladder, which makes it a usable test case for the framework rather than evidence for it. The first step yields an indicative licensing estimate, grade 1. SourceX then proposes an indicative range from the company's inventory, which the company can agree or counter, and its process page says pricing stays indicative until a Dataset Deal Confirmation is signed: grade 3.

The Confirmation is the grade 4 event: it fixes the exact dataset, the payout, the rights, the processing route and the deadlines. Grade 5 comes later. SourceX reviews the delivered data, Closing follows written acceptance, and its terms state that data that does not reach Closing earns no payment. Licenses are exclusive for AI training, and the term and any alternatives sit in each Confirmation, so each SourceX deal is documented with the rights fields that listing data lacks.

What that means for diligence is a request, not a conclusion. SourceX does not publish its commercial terms, and its estimates are indicative. It shares company and commercial metrics with investors directly, so the useful questions are how many observations sit at each grade and how often grade 4 becomes grade 5. SourceX is raising from accredited investors under Rule 506(c); materials are available on request at sourcex.si/investors.

What could make this evidence framework wrong

The framework assumes that rights can be read and that closed deals will eventually be visible to a diligent investor. Each assumption can fail, and several of the measurements here have limits of their own:

  • The listings are not an AI training market. DaDaDa's products are mostly retail, location, marketing and financial data; AI training licenses may hide prices more often, or less, than these venues do.
  • Hidden prices could sit above or below posted ones. The listing data shows only that a price is withheld, not where it would fall, so the negotiation share says nothing about the level of value.
  • Confidentiality can starve the top grades. If buyers and sellers never disclose terms, grade 4 and 5 evidence stays private, and valuations default back to ranges whatever their quality.
  • Posted-price venues for AI data could mature. If standardized licenses with stated rights become common, asking prices would carry more information and this framework would weight them too lightly.
  • Few buyers can move the whole set. Where a handful of buyers account for most deals, one renegotiation can shift every comparable at once, and a range built on three deals is fragile however well graded.
  • Rights can change after signing. New law, a court ruling or a platform's terms can change what a license is worth mid-term, so a paid deal from one year may not be a comparable the next.

Diligence questions on pricing evidence, SourceX included

Ask any data company to show its pricing evidence by grade, with the rights attached, and to separate what is signed from what is paid. Put the same questions to SourceX:

  • How many price observations do you hold at each of the five grades, and from what date range?
  • For signed contracts, what share have reached acceptance and payment, and how long does that take on the median?
  • Which rights fields do you record for every deal: exclusivity, permitted uses, term, territory and sublicensing?
  • When you cite a market price, is it an asking price, an offer or a paid transaction, and who was the counterparty type?
  • How do you adjust a comparable for different rights or preparation, and can we see the adjustment written out?
  • How concentrated are your paid deals by buyer, and what happens to your comparables if the largest buyer renegotiates?

Methods note: the listing count and the worksheet

Listing counts come from DaDaDa/data/final_data.csv in the ZJU-DIVER/DaDaDa repository at commit ead4f7f88f22933fcdc3a8bf2f239294a6cb6c3c (the commit read is dated 2026-01-30), read on 2026-10-28 with a script that only reads the CSV (dadada_price_evidence.py, kept for a recount); a byte-order mark, if present, is ignored. The repository does not state when the listings were collected. Every row counts once, and pricing modes, trimmed and lowercased, are taken as the dataset's authors labelled them: free, negotiation, subscription, one-off and usage-based. Paid listings are every row whose mode is not free; no public price means the negotiation mode; posted prices are the three priced modes.

Medians use priced listings with a price above zero (2 usage-based listings priced at zero are excluded); a median between two values is their mean, rounded half up to whole cents, and per the README subscription prices are 12-month costs. Rights language was measured only on the four English-language marketplaces (AWS Data Exchange, Snowflake, Databricks, Datarade), on title and description together, with case-insensitive whole-word patterns: exclusive or exclusivity, and training phrases (training data, model training, AI training, or train or training followed by AI, ML, machine learning, model, models, LLM or LLMs); the first eight exclusivity matches among paid listings, in file order, were printed and read. Percentages are computed from integer counts and rounded once, half up, to one decimal place.

The illustrative worksheet (xs0204_worksheet.py) uses fixed, invented inputs and no randomness. It admits a comparable only at grade 4 or 5 with stated rights equal to the target's, reports grade 3 as an indication, excludes grades 1 and 2, and takes the naive mean over every observation with a number. Gaps are computed from unrounded values and rounded once to whole percent.

Not measured: completed sales on any marketplace, the prices behind negotiation-only listings, discounts from posted prices, private offers made outside listings, and any SourceX deal, price or pipeline figure. The five grades, the comparable record and the worksheet are editorial tools drafted for this article on 2026-10-28.

Frequently asked questions

Are data marketplace list prices a good guide to what a dataset is worth?

Only as a ceiling for a similar, posted product. In the 16,147 listings in the DaDaDa dataset, 47.8 percent of paid listings show no public price, and none of the 14 listing fields records exclusivity, permitted use or term, so a posted price cannot be compared with a licensed dataset's rights without more information.

What makes a data transaction a usable comparable?

A signed or paid deal whose rights, scope, preparation and conditions are known, ideally for a similar dataset and buyer purpose. If the rights are not stated, the price cannot be mapped onto the dataset you are valuing, however close the records look.

Why do so many data listings show no price?

The listing data does not say why; it records only that buyers must negotiate. Visibility varies sharply by venue, from every paid listing on some exchanges to none on others, which suggests venue conventions matter. Treat a missing price as missing evidence, not as a high or low value.

How should an investor treat a company's estimated payout range?

As grade 1 evidence: a statement of what the company is prepared to suggest, useful for context but not for valuation. Ask how often estimates became signed terms and how often signed terms became paid deals, and value the company on those conversion records rather than on the ranges.

Does SourceX publish its prices or deal values?

No. SourceX does not publish its commercial terms; each deal's exact payout and terms are fixed in its Dataset Deal Confirmation, and estimates before that are indicative. SourceX shares company and commercial metrics with investors directly through its investor materials request.

Sources

  • The DaDaDa repository README describes 16,147 data-product listings from nine marketplaces, the five pricing modes, and that subscription prices are 12-month costs. Source
  • DaDaDa's final_data.csv holds the 14 listing fields and pricing modes counted in this article: 47.8 percent of paid listings have no public price. Source
  • IFRS 13 uses the price within the bid-ask spread most representative of fair value, and lists a wide or widening bid-ask spread among signs of decreased market activity. Source
  • US Treasury regulation 26 CFR 1.482-4 calls the comparable uncontrolled transaction method generally the most direct and reliable measure for intangibles, with reliability tied to data completeness. Source
  • SourceX's how-it-works page describes the indicative estimate and range, the Dataset Deal Confirmation that fixes payout and rights, and that data not reaching Closing earns no payment. Source
  • SourceX's FAQ says it does not publish its commercial terms, that estimates are indicative and that licenses are exclusive for AI training. Source
  • SourceX's investors page says it shares company and commercial metrics with investors directly. 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