Deal economics
Data supplier acquisition cost: count the suppliers that never signed
By SourceX Editorial · Reviewed by Noah Loul ·
10 sections
Short answer
True data supplier acquisition cost is all the effort spent on a cohort of supplier candidates, divided by the datasets that cohort produced at each gate: eligible, signed, accepted and paid. Counting only hours logged against successful deals understates it. In this article's illustrative cohort, the deal ledger shows 49 hours per paid dataset while the full cohort cost is 112.5 hours. Investors should ask for the waterfall, not the ledger.
Value drivers
Data supplier acquisition
Key takeaways
- Measure acquisition cost per cohort, not per deal: the numerator is all effort on every candidate, the denominator is the supply that cohort produced.
- Report cumulative cost at four gates, eligible, signed, accepted and paid, with denominators tied to the same cohort.
- Keep candidates that are still open at the cohort close as censored observations rather than failures or omissions.
- Separate a category's cost from one unusually easy or difficult supplier with medians, counts and an outlier note.
- Test qualification changes prospectively: hours saved count only if accepted supply holds.
What is the true data supplier acquisition cost?
It is the total effort a data company spends on a cohort of supplier candidates, including every candidate that declined, failed qualification or never delivered, divided by the datasets that cohort produced. The cost of the successful deals alone is a different and much smaller number, and confusing the two is how supply economics get overstated.
The distinction matters because a selective qualification process loses many candidates along the way, and each one still costs something: outreach, an intake review and, for those who get further, rights and data review. If a company reports cost per signed supplier from a ledger that records time only against signed deals, that effort drops out of the unit economics and comes back later as overhead nobody can explain. The cohort number is also the one that tells an investor whether scaling data operations will make supply cheaper or merely larger.
Federal contracting rules deal with a version of this problem directly. The Federal Acquisition Regulation defines bid and proposal costs as the costs of preparing, submitting and supporting bids and proposals on potential contracts, and allows them as indirect expenses on contracts to the extent they are allocable and reasonable. The cost of a losing bid does not vanish; it joins the cost base that the winning work carries. A supplier pipeline deserves the same treatment: the candidates that failed are part of the cost of the ones that succeeded.
Define the cohort, the unit of supply and the hours that count
Fix the cohort by entry date and source, fix the unit as a dataset accepted by a buyer against a written acceptance test, and fix which hours belong before any number is computed. A company that changes any of the three between periods can make its acquisition cost improve without any operational change.
At a glance
Define the cohort, the unit of supply and the hours that count
Shared costs need a stated rule too. The FAR's indirect cost principle requires indirect costs to be grouped logically and allocated on a base common to all the cost objectives they serve, in line with the benefits each receives, and it forbids fragmenting an accepted base by removing individual elements. Applied to supply, the same logic says a company should not drop its hardest category from the base to make the average look better.
A public eligibility rule makes the first gate auditable. SourceX's published rule, for example, covers U.S., Canadian and EU companies whose full-time headcount reached at least 50 at its peak, contractors aside, and qualifying starts with a short company assessment, no data upload required. Whatever the gate, write it down before the cohort starts and count the candidates who failed it.
- Cohort entry: the date a candidate first received outreach or submitted an intake form, grouped by month or quarter.
- Cohort source: inbound applicants and outbound targets tracked separately, because their conversion and effort differ.
- Category: the data family, such as support tickets or claims files, because review effort varies by category.
- Unit of activated supply: one dataset accepted by a buyer, with a paid flag added when cash arrives.
- Hours that belong: outreach, intake review, rights and data review, contracting, preparation, supplier support and review of datasets that failed acceptance, including work on candidates that never signed.
- Hours allocated separately: platform and tooling development, buyer acquisition and general overhead, each on a stated basis.
- Cohort close: a fixed date after which no new effort is attributed; candidates still open are censored.
The true supplier acquisition cost model: a cohort waterfall
The waterfall follows one cohort through each stage, records the hours spent on every candidate that entered the stage, and keeps the hours spent on candidates that later succeeded in a separate column. The example is illustrative: its counts and hours are invented to show the arithmetic, use no SourceX data and describe no real company.
Read the last two columns together. The 8 datasets that were accepted and paid account for 392 of the cohort's 900 hours. The other 508 hours went to the 192 candidates that dropped out at some stage or, in one case, were accepted but not yet paid when the cohort closed. A deal ledger shows the first number; the cohort shows both.
| Stage (illustrative cohort) | Candidates entering | Hours per candidate | Stage hours | Of which on the 8 paid |
|---|---|---|---|---|
| Outreach and first contact | 200 | 0.5 | 100 | 4 |
| Intake assessment review | 80 | 1.5 | 120 | 12 |
| Qualification: rights and data review | 30 | 6 | 180 | 48 |
| Contracting | 15 | 8 | 120 | 64 |
| Preparation and delivery | 12 | 25 | 300 | 200 |
| Buyer acceptance support | 10 | 8 | 80 | 64 |
| Cohort total | 200 | Varies by stage | 900 | 392 |
Cost per eligible, signed, accepted and paid dataset
Divide cumulative cohort hours by the candidates that passed each gate, and keep every denominator tied to the same cohort. Cost per unit rises at each gate because the numerator keeps growing while the denominator shrinks, which is exactly the information a per-deal ledger throws away.
In the illustrative cohort the deal ledger understates the cost of a paid dataset by a factor of about 2.3: 49 hours against 112.5. The ratio means nothing outside the example. The point is that an investor cannot tell which number a company is reporting unless the company shows the waterfall behind it.
Hours become money through a loaded hourly rate for each role. Ask for the roles and the rates separately, so that a shift in staffing mix is not mistaken for an efficiency gain, and ask whether founders' and contractors' hours are logged at all; early cohorts look cheap when founders' hours go unlogged.
| Gate | Passed the gate | Cumulative hours | Hours per unit | What it tells an investor |
|---|---|---|---|---|
| Eligible | 30 | 220 | 7.3 | Cost of finding candidates worth reviewing |
| Signed | 12 | 520 | 43.3 | Cost of turning review into a signed agreement |
| Accepted | 9 | 900 | 100 | Cost of supply a buyer actually took |
| Paid | 8 | 900 | 112.5 | Cost of supply that produced cash |
| Deal ledger only | 8 | 392 | 49 | What a successful-deals view reports |
How should open candidates and late payments be counted?
As censored observations: kept in the cohort with their hours, counted neither as failures nor as successes, and resolved in a later report. Dropping them flatters whichever gate they sit at, and counting them as failures punishes recent cohorts for being recent.
Survival analysis exists for exactly this. The documentation of lifelines, an open-source Python library for survival analysis, warns that a common mistake is to ignore right-censored individuals, those whose event has not yet happened when the data are examined, and shows how doing so severely underestimates the true average duration. The same holds for time from first contact to accepted delivery: report a median with its censored count, or a Kaplan-Meier curve, rather than an average of finished cases.
In the illustrative cohort, one dataset was accepted but not yet paid at the close. It stays in the accepted count, stays out of the paid count, and its hours stay in the cohort total. When the payment arrives, the next report moves it to paid and restates the cohort.
Is it the category or one unusual supplier?
Report the median hours per accepted dataset for each data category, show the spread, and name any single supplier that moves a category's average by itself. One cooperative supplier with clean exports can make a category look cheap; one with tangled customer contracts can make it look unworkable.
Rights review is a likely source of outliers, because the paperwork varies so much from one supplier to the next. The Data Provenance Initiative, reviewing more than 1,800 text datasets, found licenses missing on widely used hosting sites in over 70 percent of cases and wrong in over 50 percent. If public datasets carry license information that unreliable, review effort on private records depends heavily on the quality of each supplier's documents, and a category average built on two or three suppliers says little.
Show each category with and without its most extreme supplier, state the count of accepted datasets behind every median, and treat a category with only a few accepted datasets as a set of anecdotes rather than a unit cost.
Which qualification change has the best measured cost-to-value effect?
The one that saves the most hours per candidate without reducing accepted datasets, measured in a prospective comparison rather than inferred from a good quarter. Cheaper qualification that rejects suppliers buyers would have accepted is not an improvement; it moves cost into lost supply, where no ledger records it.
Run the comparison on alternating weeks or randomized candidates within the same categories and sources. Fix the primary metric before starting, hours per accepted dataset, with accepted datasets per hundred candidates as a guardrail, and keep going until enough candidates have resolved or been censored to report an interval. Publish the change that lost as well as the one that won.
Outside pressure makes the test worth running. California's AB 2013 requires generative AI developers to post, by January 1, 2026, a high-level summary of their training datasets, including their sources or owners, and NIST's Generative AI Profile (AI 600-1, July 2024) suggests contracts that specify content ownership, usage rights, quality, security and provenance expectations. Both raise the documentation buyers expect from each supplier, which adds review hours that only show up in the unit cost if the cohort model is in place.
- A rights pre-check at intake: a few questions about customer contracts and platform terms before any full review.
- A structured export test before contracting, so preparation failures surface before legal time is spent.
- Category-specific intake questions that screen out datasets with too little history or volume.
- A standing acceptance test agreed with buyers before sourcing, so suppliers are qualified against it.
What could make a low acquisition cost misleading?
A falling cost per accepted dataset is good news only if the accepted datasets are worth as much as before and buyers keep coming back. Check each of these before crediting it.
At a glance
What could make a low acquisition cost misleading
- Demand concentration: Reddit's annual report for 2024 puts substantially all of its licensing contract value in the hands of two partners. Cheap supply does not offset losing a buyer that takes most of the output.
- Rights shortcuts: a lighter review that lets a defect through can cost far more later. The FTC's 2021 order against photo-app developer Everalbum required it to delete models and algorithms developed from users' photos and videos.
- Synthetic substitution: Anthropic's transparency hub lists synthetic data generated by other models in the training mix of recent Claude models, such as Claude Opus 5.5. If synthetic data meets a buyer's acceptance test for a category, cheaper acquisition of real records in that category will not help.
- Mix shift: cost falls because the company moved toward easier categories that buyers value less.
- Volume illusion: more applications can raise total effort without raising accepted supply, so cost per application falls while cost per accepted dataset climbs.
Supply economics questions for any data company, SourceX included
Ask for the waterfall by cohort, the coded failure reasons and the censored counts, with definitions you set. SourceX is one application of the question, since each stage of the waterfall above is one it operates for businesses licensing operational data to AI developers: finding and qualifying candidates, reviewing their rights, contracting, preparing the data, and handling delivery and payment.
SourceX has not published cohort, cost or supply metrics; company and commercial figures go to investors directly. Its Rule 506(c) raise is open to accredited investors, with materials available on request. Bring these questions to that conversation, or to any other data company's.
- For your last four closed cohorts, how many candidates entered, and how many became eligible, signed, accepted and paid?
- Which roles' hours are in the cost base, and how are founders' and contractors' hours recorded?
- What share of cohort hours went to candidates that never signed, and what were their coded failure reasons?
- Which category has the highest hours per accepted dataset, and is that the category or one supplier?
- Which qualification change did you test prospectively, and what happened to accepted supply?
Methods note: what the model counts and what it leaves out
The cohort waterfall and the gate table are an editorial model built for this article on 2026-10-26. The cohort, its counts and its hours are illustrative and invented to show the arithmetic. The model's rules borrow from FAR 31.203 and 31.205-18 on cost allocation and the lifelines documentation on censoring; the risk evidence comes from the Data Provenance Initiative audit (2024), Reddit's annual report for 2024, California's AB 2013, NIST AI 600-1, the FTC's 2021 Everalbum order and Anthropic's transparency hub, each cited on this page.
Not measured: real acquisition costs at SourceX or any other data company, how costs vary by category in practice, and the effect of any qualification change. Those require time records and cohort data that are not public, and no SourceX results are reported here.
Frequently asked questions
How do you calculate data supplier acquisition cost?
Add all effort spent on a cohort of supplier candidates, including those that declined or failed review, and divide by the datasets the cohort produced at the gate you care about. Report eligible, signed, accepted and paid separately, with the cohort's entry dates and the number of candidates still open.
Why is cost per signed supplier misleading?
Signing is not supply. A signed supplier can still fail preparation or buyer acceptance, so cost per signed supplier understates the cost of supply a buyer actually took. Cost per accepted dataset, with failures included, is the number that connects to revenue.
Should founders' time count in supplier acquisition cost?
Yes, whenever founders do sourcing, review or contracting work. Leaving it out makes early cohorts look cheap and makes costs appear to rise when the company hires people to do the same work.
What is a good data supplier acquisition cost?
Judge it against the company's own trend and definitions rather than an outside figure, because comparable cohort data from other data companies is not published in a standard form. A cost that falls while accepted datasets per hundred candidates hold steady is a real improvement; a cost that falls because the denominator changed is not.
Does SourceX publish its supplier acquisition cost?
No. Its cohort, cost and supply figures are not public, and SourceX provides company and commercial metrics to investors directly. A materials request at sourcex.si/investors is the place to ask how its acquisition cost treats suppliers that never signed; any offering is limited to verified accredited investors.
Sources
- FAR 31.205-18 defines bid and proposal costs as the costs of preparing, submitting and supporting bids and proposals on potential contracts and allows them as indirect expenses on contracts to the extent they are allocable and reasonable. Source
- FAR 31.203 requires indirect costs to be grouped logically and allocated on a base common to all cost objectives in line with the benefits received, and bars fragmenting an accepted base by removing individual elements. Source
- The lifelines survival analysis documentation says ignoring right-censored individuals is a common mistake that severely underestimates the true average duration. Source
- The Data Provenance Initiative audit of more than 1,800 text datasets reports license omission rates above 70 percent and error rates above 50 percent on widely used dataset hosting sites. Source
- California AB 2013 requires generative AI developers to post by January 1, 2026 a high-level summary of training datasets, including their sources or owners. Source
- NIST AI 600-1, the Generative AI Profile published in July 2024, suggests contracts and service level agreements that specify content ownership, usage rights, quality standards, security requirements and content provenance expectations. Source
- Reddit's annual report for 2024 says substantially all of the contract value associated with its licensing revenue is derived from two of its partners. Source
- The FTC's 2021 final order against photo-app developer Everalbum required it to delete models and algorithms developed using users' photos and videos. Source
- Anthropic's transparency hub says Claude Opus 5.5 was trained on a mix including public and private datasets and synthetic data generated by other models. Source
Request SourceX investor materials
SourceX shares company and commercial metrics with investors directly. Any offering is made only to verified accredited investors.