Skip to content

Deal economics

More suppliers, fewer licenses? Testing data supply growth against demand fit

By SourceX Editorial · Reviewed by Noah Loul ·

10 sections

Short answer

Data supply growth pays only where it meets demand fit. In this article's illustrative cohort model, quadrupling supplier applications while preparing every qualified supplier in arrival order nearly doubles staff hours and cuts accepted datasets from 13.6 to 10.2, because unmatched suppliers take preparation capacity. Gating preparation on a buyer signal lifts the same intake to 22.8. Investors should ask for funnels split by category and cohort.

Value drivers

More suppliers

This page covers more suppliers, fewer licenses? Testing data supply growth against demand fit.Illustrative

Key takeaways

  • Preparation comes before buyer acceptance, so intake that lands outside requested categories mostly becomes stranded work.
  • In the illustrative model, 400 applications prepared in arrival order produced fewer accepted datasets than 100, while staff hours rose 89.7 percent.
  • Intake growth is not the problem in itself: with the base mix held, doubling applications lifted accepted datasets from 13.6 to 17.5.
  • Doubling preparation capacity without a buyer-fit gate still trailed the gated rule at the original capacity, 20.4 against 22.8 accepted datasets.
  • Ask for six category-level measures, including preparation started without a buyer signal and the wait matched suppliers face.

Does data supply growth create value before demand fit?

Only where the new supply lands in categories buyers have asked for. Data supply growth and demand fit separate quickly in a licensing business because the expensive work, preparing records into a dataset a buyer can test, comes before any buyer accepts it. Growth in applications or qualified suppliers outside those categories mostly adds work that no license will pay back.

The thesis to test is narrow: scaling supplier intake before buyer fit converts preparation capacity into stranded work, and when that capacity is shared, it can lower the number of accepted datasets, not just raise their cost. Demand fit here means a category with a buyer request that carries a written acceptance test, not a category a pitch deck calls hot.

Two-sided markets have shown versions of this before. Jun Li and Serguei Netessine studied a holiday rental platform that absorbed a one-time migration of listings and estimated, in Management Science (2020), that doubling market size cut traveler confirmation rates by 15.4 percent and host occupancy by 15.9 percent, which they attributed to search friction. The closest theory is Nick Arnosti, Ramesh Johari and Yash Kanoria's paper in Manufacturing and Service Operations Management (2021), which finds that when applying is cheap, applicants flood a market with applications that are never read, and warns platforms against focusing only on attracting users and making contact easy.

How a supply funnel turns intake into stranded preparation

Every supplier who enters the funnel consumes effort at up to three stages before any buyer decides: screening the application, qualifying the company and its rights, and preparing its records into a dataset. In this article's model preparation carries the most hours per supplier, and it is the stage that strands when no buyer accepts the result.

Scaling intake changes the mix as well as the volume. Outreach that reaches more companies tends to reach categories where records are plentiful, and records are often plentiful because nobody has been paying for them. Those suppliers can pass eligibility and rights checks as easily as anyone, so qualified supply grows on schedule while buyer acceptance in those categories stays near zero. That mix shift is an assumption of this article, labeled as such, and the ledger later in the piece is how to test it.

The third channel is crowding. If preparation runs in arrival order, an unmatched file that qualified on Monday takes the slot a matched file would have used on Tuesday. Yash Kanoria and Daniela Saban's Management Science paper (2021) on matching platforms finds that a platform can reduce wasted search by restricting what agents may do, and that in unbalanced markets only the short side should initiate contact. In data licensing, buyer requests are often the short side, which suggests letting a buyer signal start the preparation queue.

The cohort model: three intake scales, two preparation rules (Illustrative)

The model follows one supplier cohort through screening, qualification, preparation and buyer acceptance at three intake scales and under two preparation rules. Every input is illustrative and describes no real company; the script and its formulas are in the methods note.

Three categories carry different buyer fit. Matched categories have an open buyer request with an acceptance test and a 70 percent acceptance rate for prepared datasets; adjacent categories, where buyers asked for nearby data, accept 30 percent; unmatched categories accept 5 percent. Qualification pass rates are 30, 30 and 35 percent, slightly higher in unmatched categories because their records are plentiful. Screening costs 2 staff hours per application, qualification 10 per qualified supplier and preparation 60 per dataset, with 2,400 preparation hours, or 40 datasets, available per cohort.

The base cohort has 100 applications, half from matched categories, 30 from adjacent and 20 from unmatched. Broader outreach adds 100 more at 20, 30 and 50 percent, then 200 more at 10, 30 and 60 percent. Under the first rule, every qualified supplier is prepared in arrival order until capacity runs out; under the second, matched suppliers are prepared first, then adjacent ones, and unmatched suppliers are qualified and held until a buyer asks. The results, computed by script:

The cohort model: three intake scales, two preparation rules (Illustrative)
Rule and applicationsAccepted datasetsPreparation hours (stranded)Staff hours per accepted datasetStranded share of preparation
Prepare on qualification, 10013.61,860 (1,047)174.956.3%
Prepare on qualification, 20013.42,400 (1,594)255.766.4%
Prepare on qualification, 40010.22,400 (1,788)440.774.5%
Prepare on buyer signal, 10013.21,440 (648)147.745.0%
Prepare on buyer signal, 20020.12,340 (1,134)167.948.5%
Prepare on buyer signal, 40022.82,400 (1,032)197.143.0%
Stacked columns of preparation hours split into accepted and stranded work for 100, 200 and 400 applications under two rules: prepare on qualification gives 13.6, 13.4 and 10.2 accepted datasets; prepare on buyer signal gives 13.2, 20.1 and 22.8.
Illustrative cohort model from this article; not SourceX data.

Why accepted datasets fell when applications quadrupled

Accepted datasets fell from 13.6 to 10.2 because the added applicants were mostly unmatched, and under arrival-order preparation they took capacity in proportion to their share of the queue. At 400 applications, unmatched suppliers filled 20.5 of the 40 preparation slots and yielded 1.0 accepted dataset, while matched suppliers got 8.3 slots, down from 15.0 at the base, and yielded 5.8.

Staff hours rose 89.7 percent, from 2,370 to 4,495, and hours per accepted dataset more than doubled, from 174.9 to 440.7. Nothing in the headline funnel would show the problem: applications quadrupled and qualified suppliers rose from 31.0 to 129.5. The damage is visible only in accepted output per category and in a queue of 89.5 qualified suppliers left waiting.

Two sensitivities show how strong the effect is. Unmatched categories would need a 21.3 percent acceptance rate, more than four times the input, before the 400-application cohort matched the base cohort's accepted output under arrival-order preparation. And doubling preparation capacity to 4,800 hours under the same rule yields 20.4 accepted datasets, still fewer than the 22.8 that the buyer-signal rule reaches with the original capacity.

The buyer-signal rule gets its result from where the slots go, not from more of them. With the same 400 applications and the same 40 slots, matched suppliers fill 27.0 and adjacent suppliers 13.0, yielding 18.9 and 3.9 accepted datasets, and the stranded share of preparation falls to 43.0 percent from 74.5 percent. Total staff hours are identical under both rules at that scale, 4,495, because both prepare 40 datasets; they differ only in which suppliers get the slots.

Testing the alternative: is more intake only more expense?

Partly. More applications raise expense without improving accepted supply when the added intake shifts toward unmatched categories and preparation is not gated. At 200 applications under arrival-order preparation, staff hours rose 44.9 percent while accepted datasets slipped from 13.6 to 13.4.

Intake growth is not the problem in itself. In a variant where the added 100 applications keep the base mix, half matched, the same arrival-order rule lifts accepted datasets from 13.6 to 17.5, and the buyer-signal rule reaches 24.0. The model therefore separates two claims an investor should not let a company merge: that the funnel is growing, and that it is growing where buyers are.

The buyer-signal rule has its own cost. Qualified suppliers who are not prepared form a queue of companies waiting for an answer, and at 400 applications it holds 89.5 under either rule; the difference is who waits. Under arrival order, 18.7 matched suppliers are among them; under the buyer-signal rule none are, and 66.5 of the waiting are unmatched. A stricter version stops unmatched suppliers at screening, saving the qualification hours, at the price of not knowing what they hold if a buyer appears later.

The category cohort ledger an investor should request

Ask for the funnel by category and cohort, with preparation hours split by whether a buyer signal existed when the work started. A single blended conversion rate hides every effect in this model. The ledger below defines six measures, the denominator each needs, and the reading that should prompt a follow-up question:

The category cohort ledger an investor should request
MeasureDefinitionDenominatorReading that needs a follow-up
Buyer-fit share of intakeApplications in categories with an open buyer request carrying an acceptance testAll applications in the cohortFalling while applications rise
Preparation started without a signalPreparation hours on datasets with no buyer request when work beganAll preparation hours in the cohortRising share as outreach widens
Stranded preparationPreparation hours on datasets no buyer accepted within a fixed windowPreparation hours, by cohort and categoryHighest in the fastest-growing categories
Accepted datasets per 100 applicationsDatasets a buyer accepted, counted onceApplications, by category and cohortFlat or falling as intake grows
Matched queue waitDays a matched-category supplier waits from qualification to preparation startMatched suppliers qualified in the cohortLengthening while unmatched files are prepared
Qualified and waitingSuppliers qualified but not prepared, by category and ageQualified suppliers in the cohortLarge, old and concentrated in unmatched categories

What could make the buyer-fit thesis wrong

The thesis fails if unmatched categories turn out to have buyers, if preparation transfers between buyers, or if capacity is cheap to add. Each is testable, and an investor should look for the evidence before accepting either side of the argument:

At a glance

What could make the buyer-fit thesis wrong

  • Demand arrives later: if buyers emerge for unmatched categories, prepared datasets let a company answer quickly, and stranded hours become inventory with option value. The 21.3 percent break-even above is the acceptance rate those categories would need.
  • Portable preparation: if most preparation work, such as extraction, de-identification and documentation, serves any buyer, stranded hours are overstated; the test is how much of a prepared dataset survived the next buyer's specification unchanged.
  • Elastic capacity: if preparation can be added quickly and cheaply, crowding fades, though the model shows doubled capacity under arrival-order preparation still trails the gated rule.
  • Discovery starvation: a strict gate never learns about categories nobody has asked for yet; a small, explicit exploration budget is the usual answer, reported separately from the gated funnel.
  • Supplier goodwill: companies qualified and then held may not return when a buyer appears, so a gated funnel can lose supply it needs later.
  • Borrowed evidence: the matching studies cited here come from a holiday rental platform and matching-market theory, not data licensing, and the model's inputs are invented for illustration.

Reading SourceX's published process against the model

SourceX's public pages describe a funnel gated on buyer programs, in words; whether its cohorts behave that way is not public. Its screening page says requirements vary with SourceX's current data needs, that the material has to match a real program with AI developers worldwide, and that SourceX applies the criteria of the specific program, so a company that does not fit one program may fit another later. Its investors page lists focus areas matched against documented demand from AI labs and model developers.

The order of work points the same way. The how-it-works page says SourceX seeks commitment on the approved terms and signs its own onward agreement, then instructs processing, and the request-data page says datasets are sourced on request, not held in stock. In this article's terms, that is closer to the buyer-signal rule than to preparing every qualified supplier.

None of that shows the ledger's readings. SourceX shares company and commercial metrics with investors directly, and it is raising from accredited investors under Rule 506(c), with materials available on request through sourcex.si/investors. The questions in the next section are the ones to bring to that conversation, or to any company describing its supply growth.

Diligence questions for any company scaling data supply

Seven questions turn the model into diligence requests; each asks for a cohort, a category split and a denominator rather than a total:

  • What share of last quarter's applications came from categories with an open buyer request carrying an acceptance test, and how has that share moved as intake grew?
  • How many preparation hours began before any buyer signal, and what share of them led to an accepted dataset within a fixed window?
  • By category, how many accepted datasets did each cohort of 100 applications produce?
  • How long do suppliers in matched categories wait between qualification and the start of preparation, and is that wait growing?
  • How many qualified suppliers are waiting, in which categories, and for how long?
  • When a buyer specification changed, how much prepared work was reused unchanged?
  • What exploration budget is spent outside requested categories, and how is it reported?

Methods note: inputs, formulas and what the model leaves out

The model is bin/xs0191_fit_model.py in the writer's scratch folder (W09-main-016), with output saved beside it. It uses exact fractions and rounds each printed value once, half up. For each category, qualified suppliers equal applications times the pass rate; under arrival-order preparation, prepared datasets equal all qualified suppliers if they fit within 40 slots, otherwise 40 times each category's share of the qualified total; under the buyer-signal rule, matched suppliers fill slots first, then adjacent, and unmatched suppliers are not prepared. Accepted datasets equal prepared datasets times the category acceptance rate, as expected values rather than draws, so no randomness is involved.

Staff hours are 2 per application plus 10 per qualified supplier plus 60 per prepared dataset. Stranded preparation hours are prepared minus accepted datasets, times 60; the stranded share divides them by all preparation hours; suppliers waiting are qualified minus prepared, by category; the staff-hour changes compare each scale with the 100-application cohort under arrival order. The break-even acceptance rate solves for the unmatched rate at which the 400-application cohort's expected accepted datasets equal the base cohort's under arrival-order preparation. Two variants rerun the same formulas, one with 4,800 preparation hours (80 datasets) at 400 applications and one in which the second 100 applications keep the base mix of 50, 30 and 20 percent.

Not modeled: supplier withdrawal while waiting, rework after a buyer rejects, preparation reused across buyers, differences in dataset value, timing within the cohort, and any money amounts. The published sources support the mechanisms, not the inputs. No SourceX figure enters the model.

Frequently asked questions

What is demand fit for a data supplier category?

A category has demand fit when a buyer has asked for that kind of data with a written acceptance test the prepared dataset can be checked against. Market enthusiasm or a general interest in a sector is not the same thing, because neither tells the company what a buyer will accept.

Why can more supplier applications reduce accepted datasets?

When preparation capacity is limited and runs in arrival order, suppliers from categories no buyer wants take slots that matched suppliers would have used. In this article's illustrative model, quadrupling applications that way cut accepted datasets from 13.6 to 10.2.

Is preparing data ahead of a buyer always a mistake?

No. Preparation that serves any buyer, or categories where buyers are likely to appear soon, can shorten delivery later. The test is how much prepared work was accepted or reused, by category and cohort, and whether that work is reported apart from the gated funnel.

What should investors ask a data company about supply growth?

Ask for applications, qualified suppliers, preparation hours and accepted datasets split by category and cohort, with preparation hours marked by whether a buyer signal existed when the work began. Blended conversion rates hide the effects described here.

Does SourceX publish its supply funnel metrics?

No. Its public pages describe qualification against current data needs and processing after commercial commitment, and its investors page says it shares company and commercial metrics with investors directly.

Sources

  • Li and Netessine (Management Science, 2020): on a holiday rental platform, doubling market size cut traveler confirmation rates by 15.4 percent and host occupancy by 15.9 percent, attributed to search friction. Source
  • Arnosti, Johari and Kanoria (Manufacturing and Service Operations Management, 2021): when applying is cheap, applications go unread and congestion lowers welfare; platforms should not focus only on attracting users and making contact easy. Source
  • Kanoria and Saban (Management Science, 2021): a matching platform can reduce wasted search by restricting agents' actions, and in unbalanced markets only the short side should initiate contact. Source
  • SourceX's screening page: requirements vary with SourceX's current data needs; material must match a real program with AI developers; criteria of the specific program apply; a company that does not fit one program may fit another later. Source
  • SourceX's investors page lists focus areas matched against documented demand from AI labs and model developers and says it shares company and commercial metrics with investors directly. Source
  • SourceX's how-it-works page: SourceX seeks commitment on the approved terms and signs its own onward agreement, then instructs processing. Source
  • SourceX's request-data page says datasets are sourced on request, not held in stock. 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