Skip to content

Deal economics

Dataset valuation framework: three anchors before the first offer

By SourceX Editorial · Reviewed by Noah Loul ·

10 sections

Short answer

A dataset valuation framework that works before any buyer names a price uses three anchors: the seller's floor, the buyer's ceiling, set by the lower of use-case value and the cost of the cheapest substitute that passes the buyer's acceptance test, and executed comparables normalized to the same rights and terms. Comparables narrow the space between floor and ceiling into a range. Asking prices are not evidence of value.

Value drivers

Dataset valuation framework

This page covers dataset valuation framework: three anchors before the first offer.Illustrative

Key takeaways

  • A defensible pre-offer range sits between the seller's floor and the buyer's ceiling, the lower of use-case value and the cost of a substitute that passes the buyer's acceptance test.
  • Replacement cost for a multi-year archive is mostly delay: a substitute that must collect new history costs the use-case value lost while it waits.
  • Comparables count only when they are executed and need few, small adjustments; U.S. rules for licensed intangibles list exclusivity, use limits, territory and duration among the factors to match.
  • In the illustrative worksheet, anchors of 22, 42 and 54 value units and two normalized comparables at 38 and 46 give a range of 38 to 42.
  • Listings sit outside the range in the example: a buyer with a 42-unit substitute would pay neither 70 nor 90, so asking prices set no anchor of their own.

What must a dataset valuation framework settle before any offer?

It has to produce a range that a seller or an investor can defend before any buyer names a number: the lowest price at which licensing still leaves the data owner better off, the highest price a buyer with alternatives would rationally pay, and evidence that places a deal between the two. One number from a formula cannot do that job, because the same records are worth different amounts to different buyers.

Three readers face the same decision. A company weighing whether to start a licensing process, an investor judging the inventory value a data company claims, and a strategic buyer preparing a first offer all need a range with its reasoning shown, so the first number on the table can be tested whoever names it.

The logic is not new. In its 2022 Transfer Pricing Guidelines, the OECD describes independent enterprises comparing a proposed transaction with the other options realistically available to them, and entering it only when none of those options is clearly more attractive (paraphrased). For a dataset, the buyer's options are building a substitute or doing without; the seller's options are keeping the records out of the market or licensing them elsewhere. The framework below turns those options into numbers.

The three-anchor worksheet, filled in for one archive

Three anchors set the range: the seller floor, the buyer ceiling, which is the lower of use-case value and replacement cost, and executed comparables, which narrow the space between floor and ceiling. Each anchor is computed from inputs a company can write down and defend, not from a market price that usually does not exist.

The worked example is Illustrative: six years of a regional field-service company's work orders, about 300,000 jobs with technician notes, parts used and revisit outcomes, offered exclusive for AI training. Every figure is in value units (U), an arbitrary unit used for all anchors so they can be compared; it is not a currency and implies no conversion.

Read the worksheet row by row. The floor is 22 U, use-case value 54 U and replacement cost 42 U, so the ceiling is 42 U; two executed comparables normalize to 38 and 46 U, and the range rule returns 38 to 42 U. The worksheet, with the inputs for each anchor:

The three-anchor worksheet, filled in for one archive
AnchorQuestion it answersHow it is computed (Illustrative inputs)Evidence that settles itWhat it cannot tell you
Seller floorBelow what price does licensing leave the owner worse off?Preparation support 5 + rights and legal review 4 + staff time 3 + exclusivity opportunity cost 10 = 22 UTime records, review quotes, the value of other uses given upWhat a buyer will pay
Use-case valueWhat is the buyer's task worth with these records?Benefit 40 U a year x 3 years x dataset share 0.6 x chance the capability ships 0.75 = 54 UThe buyer's evaluation with and without the recordsWhether a cheaper substitute exists
Replacement costWhat would the cheapest substitute that passes the buyer's test cost?Setup 15 U + 18 months of delay x 1.5 U of use-case value a month = 42 UA substitute scored on the same acceptance testValue if no substitute can pass
Buyer ceilingAbove what price does the buyer walk?Lower of use-case value and replacement cost = 42 UBoth anchors aboveWhere in the zone a deal lands
Executed comparablesWhere did similar rights actually trade?Two deals normalized to the subject's terms: 38 and 46 USigned terms, normalized factor by factorAnything, if they need large adjustments
Range ruleWhat range is defensible now?From the higher of floor and lowest comparable to the lower of ceiling and highest comparable = 38 to 42 UThe rows aboveThe price a negotiation will reach
A range chart on a scale of 0 to 100 value units shows the seller floor at 18 to 28, replacement cost at 33 to 51, use-case value at 36 to 72, executed comparables at 38 to 46 and excluded asking prices at 70 to 90, with a shaded band marking the reconciled range of 38 to 42.
Illustrative anchor ranges for one archive in value units, with the reconciled range shaded and asking prices drawn as excluded.

Replacement cost for historical records is mostly waiting

For a multi-year operating archive, replacement cost is mostly the cost of waiting. A buyer cannot buy back six years of revisit outcomes; it can only generate imitations or start collecting new history, and the second route's main cost is the use-case value lost while the history accumulates.

Accounting gives the anchor a precise meaning. IFRS 13, the international fair value standard, describes the cost approach as a valuation technique that "reflects the amount that would be required currently to replace the service capacity of an asset". The words that matter are service capacity: a substitute counts only if it does the same job, which for a dataset means passing the buyer's own acceptance test. A synthetic set that covers 60 percent of that test is not a cheaper substitute; it is a different product.

In the example, generated work orders fail the revisit-outcome checks, so they drop out. The cheapest passing route is a collection partnership: 15 U of setup and 18 months before enough history exists, at 1.5 U of use-case value a month, for 42 U. With the delay at 12 or 24 months instead, replacement cost runs from 33 to 51 U, and that spread is what the chart draws.

Do not confuse replacement cost with what the records cost to produce. National accounts now treat data as an asset: the 2025 System of National Accounts, the international statistical standard, includes the information content of electronic data collected and used in production, and an IMF presentation hosted by the UN Statistics Division notes that there are not many sales or purchases of data, recommending a sum-of-costs method for data produced on own account. A sum of costs tells you what the records cost their owner to make, which bears on the floor, not on what a buyer would pay.

Use-case value: the ceiling seen from the buyer's side

Use-case value is what the buyer's task is worth with these records, discounted for the share of the gain the records cause and for the chance the capability ships. It is the anchor sellers know least about, so it is better treated as a range with its weakest input named than as a point.

The example takes a task worth 40 U a year for three years. The records' share of the gain, 0.6, is the input that matters most: it should come from the buyer's own evaluation run with and without the records, the kind of measurement a scoped utility pilot can supply. With that share between 0.4 and 0.8, use-case value runs from 36 to 72 U. The chance of shipping, 0.75, is an assumption about whether the buyer's capability reaches production at all, and it belongs in the range rather than buried in the benefit figure.

The ceiling is the lower of the two buyer-side anchors. Here replacement cost, 42 U, sits below use-case value, 54 U, so the substitute caps the deal: a buyer offered the records for more than 42 U would rather wait and collect. If a passing substitute did not exist at all, use-case value alone would set the ceiling.

Which comparable deals count, and how far can they be adjusted?

A comparable counts only if it was executed, covers rights close to the subject's and needs few, small adjustments. Every adjustment is an assumption, so a deal that must be stretched three ways to fit is a story about the analyst's inputs rather than evidence about price.

U.S. tax regulations on licensed intangibles state the same discipline. Under the comparable uncontrolled transaction method in 26 CFR 1.482-4, circumstances count as substantially the same only when the differences are minor, have a definite and reasonably ascertainable effect on the amount charged, and are adjusted for; comparability then turns on factors that include the terms of the transfer, whether rights are exclusive or nonexclusive, restrictions on use, geographic limits and the duration of the license (paraphrased). That is close to the counting rule below: few differences, each with a measurable effect. Those factors make a ready checklist for normalizing a data license: rights granted, exclusivity, permitted uses, term, history depth and any services bundled into the price.

This worksheet counts a comparable only if it needed at most one adjustment and that adjustment moved its stated value by no more than 25 percent. Comp 3 below shows why: three adjustments carry it from 24 U to 68 U, and the result depends entirely on the multipliers chosen. The candidates, with how each was treated (Illustrative):

Which comparable deals count, and how far can they be adjusted?
Candidate (Illustrative)Stated termsAdjustment to the subject's termsNormalizedCounts toward the band?
Comp 1Exclusive AI training license, similar history, preparation service bundled: 46 URemove the 8 U service38 UYes: one adjustment, 17%
Comp 2Exclusive AI training license, similar history, no outcome fields: 40 UAdd 15% for the subject's outcome fields46 UYes: one adjustment, 15%
Comp 3Non-exclusive, evaluation-only rights, 4 years of history: 24 Ux1.25 exclusivity, x1.5 training rights, x1.5 history68 UNo: three adjustments
Marketplace listingsAsking prices for two similar archives: 70 and 90 UNone possibleNot normalizedNo: not transactions

Are asking prices a usable shortcut? Testing the strongest alternative

No, and the framework shows why in its own terms. Asking prices and published ranges record what sellers hope for, not what buyers accepted. In the example, two listings at 70 and 90 U sit entirely above the 42 U ceiling: a buyer that can build a passing substitute for 42 U would pay neither price, so the listings say nothing about where this deal can close.

The anchors also explain why listings run high. A listing is set without the buyer's replacement cost or use-case value in view; it is, at best, a seller's floor plus hope. Only the two buyer-side anchors and executed comparables bring the other side of the deal into the range.

Executed private terms are scarce for structural reasons. The IMF material cited above notes how few sales or purchases of data there are to observe, and SourceX's own standard terms say that neither side publicizes the deal and that an earlier estimate is indicative only, with the exact amount fixed in the signed Dataset Deal Confirmation. Confidential, deal-specific pricing is normal in this market, so a method that leans on public price points leans on the least representative ones.

The shortcut has one honest use: as an upper bound on what sellers believe. Grading every price observation by how close it sits to a closed deal is a separate exercise; in this worksheet the rule is simply that a listing never enters the comparable band, and a range built from listings is labeled as seller expectations, never as value.

Reconciling the anchors: five illustrative cases

The range rule returns one of four readings: anchored, unanchored, no deal or conflict, and each tells the reader what to do next. The script applies the rule to five cases built on the same archive (Illustrative). Each reading is a verdict on evidence, not a forecast of where a negotiation ends:

Reconciling the anchors: five illustrative cases
Case (Illustrative)FloorCeilingRangeReading and next step
Base case: Comps 1 and 2 count22 U42 U38 to 42 UAnchored: defend the range with the two deals
No executed comparable counts22 U42 U22 to 42 UUnanchored: report the whole zone and say so
A substitute passes the buyer's test for 14 U22 U14 UNoneNo deal at any price: rescope or walk
The only comparable closed at 60 U22 U42 UNoneConflict: recheck the use case or the substitute it contradicts
Asking-price shortcut: listings at 70 and 90 UNot computedNot computed70 to 90 UNot evidence of a clearing price

What could make a three-anchor range wrong

Six things can break the range: the buyer knows its use case better than the seller, substitutes improve, comparables age, rights turn out narrower than assumed, the exclusivity cost is misjudged, and the inputs themselves are guesses. Each has a check.

  • Information asymmetry. The buyer sees its own evaluation and roadmap; the seller sees neither. Treat use-case value as the widest anchor and seek a measured dataset share before leaning on it.
  • Substitutes improve. A generated or public alternative that fails the acceptance test today may pass next year, and the third case shows what happens to the ceiling when one does. Re-run the replacement anchor before each negotiation.
  • Comparables age. Data needs, model capabilities and the supply of licensed records can all shift between deals, and the IMF presentation cited above notes how few data transactions exist to begin with. Date every comparable and drop stale ones.
  • Rights may be narrower than assumed. Every anchor assumes the seller can grant exclusive AI training rights to the records offered; if a rights review narrows the scope, recompute all three anchors on the narrower set.
  • Exclusivity cost is a judgment. The floor's largest line is the value of other uses given up; a seller that sets it carelessly can price itself out of a deal or underprice exclusivity.
  • The inputs are illustrative. The worksheet's structure is general; its numbers are examples, and a real archive can reorder which anchor binds.

Valuation questions for any data company, SourceX included

Ask a data company how it builds a range before an offer, in these terms, and check that its answer separates evidence from hope. SourceX should face the same questions, and its public pages set out no pricing method. Its how-it-works page says pricing stays indicative until a Dataset Deal Confirmation is signed, that licenses are exclusive for AI training with the term and any alternatives set in each Confirmation, and, of its fit signals, that "final value depends on actual Acquisition Partner demand and due diligence".

Nothing on SourceX's public pages reports a price, a range or a transaction figure, and its investors page offers company and commercial metrics to investors on request. SourceX is raising from accredited investors under Rule 506(c), with investor materials available through that page. These questions work for any data company:

  • Which anchor set your last three ranges: the seller floor, the buyer's replacement cost, its use-case value or an executed comparable?
  • How many executed comparables did you use, and how many adjustments did each need?
  • Have any buyers tested a substitute against their acceptance test, and what did it cost them?
  • How do you measure the share of a buyer's gain that one dataset causes?
  • How often has a range come back unanchored or in conflict, and what did you do next?
  • Where do asking prices or public figures enter your process, and how are they labeled?

Methods note: worksheet rules, inputs and limits

The three-anchor worksheet, the range rule and the five cases were built for this article on 2026-10-28 as an editorial framework, computed by a short Python script (xs0203_three_anchor_range.py, kept with its output in the writer's working folder) with exact fractions, rounding half up once at output and no randomness. All inputs are examples in value units (U), and no SourceX figure appears.

Rules as the script applies them. Floor is the sum of four costs. Use-case value is annual benefit times years times dataset share times shipping probability. Replacement cost is setup plus months of delay times the monthly value lost, which is annual benefit times dataset share times shipping probability, divided by 12 (1.5 U). Ceiling is the lower of use-case value and replacement cost. A comparable is normalized by applying its multipliers, then its additive amounts, and counts only if it is an executed deal with at most one adjustment that moves its stated value by no more than 25 percent; asking prices never count. The rule then returns no deal if the ceiling is below the floor, the floor-to-ceiling zone if no comparable counts, a conflict if every counted comparable lies below the floor or every one lies above the ceiling, and otherwise the overlap of the zone with the counted comparables' band. The chart's spans vary one input each: exclusivity cost 6 to 16 U puts the floor at 18 to 28 U, dataset share 0.4 to 0.8 puts use-case value at 36 to 72 U, and delay of 12 to 24 months puts replacement cost at 33 to 51 U.

Public evidence: the OECD Transfer Pricing Guidelines (2022), 26 CFR 1.482-4 on eCFR, IFRS 13 on the IFRS Foundation's site, the 2025 System of National Accounts and an IMF presentation on recording data, both on the UN Statistics Division's site, all checked through search excerpts on those domains and paraphrased unless quoted; and SourceX's how-it-works and investors pages, fetched directly on 2026-10-28. Not measured: any real dataset's value, any executed price, how often listings differ from closing terms, buyers' substitution costs, and SourceX's prices or results.

Frequently asked questions

How do you value a dataset before a buyer makes an offer?

Set three anchors: the seller's floor, the buyer's ceiling as the lower of use-case value and the cost of the cheapest substitute that passes the buyer's acceptance test, and executed comparables normalized to the same rights and terms. The defensible range is where the comparables overlap the space between floor and ceiling.

Is replacement cost the same as what the data cost to create?

No. Replacement cost is what the buyer would spend to get a substitute that does the same job, which for historical records is mostly the value lost while new history accumulates. The cost of producing the records bears on the seller's floor, not the buyer's ceiling.

Can asking prices on data marketplaces value a dataset?

Not on their own. A listing records what a seller hopes for, set without the buyer's substitute or use case in view, and few data transactions exist to test it against. Treat listings as an upper bound on seller expectations and label them that way.

How many adjustments can a comparable deal take before it stops being useful?

There is no universal limit, but each adjustment is an assumption. This article's worksheet counts a comparable only if it needed one adjustment that moved its value by no more than 25 percent, and sets aside deals that must be stretched further.

Does SourceX publish its dataset prices or valuation method?

No. SourceX's public pages set out no pricing method or price list. Its how-it-works page says pricing stays indicative until a Dataset Deal Confirmation is signed and that final value depends on actual Acquisition Partner demand and due diligence.

Sources

  • The OECD's 2022 Transfer Pricing Guidelines describe independent enterprises comparing a transaction with the other options realistically available to them and entering it only when no alternative is clearly more attractive. Source
  • 26 CFR 1.482-4(c) treats circumstances as substantially the same only when differences are minor, have a definite and reasonably ascertainable effect on price and are adjusted for, and lists comparability factors including exclusivity, use restrictions, geographic limits and license duration. Source
  • IFRS 13 describes the cost approach as a valuation technique reflecting the amount that would be required currently to replace the service capacity of an asset. Source
  • The 2025 System of National Accounts includes the information content of electronic data collected and used in production in the asset boundary. Source
  • An IMF presentation on recording data, hosted by the UN Statistics Division, notes there are not many sales or purchases of data and recommends a sum-of-costs method for own-account data. Source
  • SourceX's how-it-works page says pricing stays indicative until a Dataset Deal Confirmation is signed, licenses are exclusive for AI training, neither side publicizes the deal, and final value depends on actual Acquisition Partner demand and due diligence. Source
  • SourceX's investors page says it shares company and commercial metrics with investors directly and offers investor materials on request. 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