Skip to content

Deal economics

Data licensing unit economics: trace one dataset from first call to cash

By SourceX Editorial · Reviewed by Noah Loul ·

10 sections

Short answer

Data licensing unit economics should be measured per accepted dataset, after three costs a deal margin usually omits: hours spent on candidates that never shipped, repeat preparation before acceptance, and the days between spending and collecting. In this article's illustrative model, a 66-point deal margin on net revenue falls to 40 once those are charged, and screening rights before preparation is the lever that moves contribution most.

Value drivers

Data licensing unit

This page covers data licensing unit economics: trace one dataset from first call to cash.Illustrative

Key takeaways

  • Measure contribution on net revenue, the buyer's payment minus the supplier's payout, because gross or net presentation under ASC 606 changes the margin denominator, not the economics.
  • Charge every hour spent on candidates that failed qualification, rights review or preparation to the accepted datasets of the same entry cohort.
  • In the illustrative waterfall, unsuccessful candidates, rework and cash timing remove 26 of 66 points of reported deal margin.
  • A partial acceptance hurts far more than a slow review, because preparation cost is already spent when the buyer accepts fewer records.
  • The largest modeled lever is catching a doomed candidate at rights review instead of after preparation, worth about 10 points of contribution.
  • SourceX has not published unit economics; it shares company and commercial metrics with investors directly.

What should data licensing unit economics measure?

Data licensing unit economics should measure contribution per accepted dataset: what one licensed delivery leaves after its supplier is paid, its direct work is charged, and its share of failed candidates, repeat preparation and waiting for cash is added. A deal margin that stops after direct work describes the winners and leaves out the cost of finding them.

The unit is one dataset, one buyer, one license. A dataset licensed a second time is a second unit that reuses earlier preparation, which is a separate reuse question; this article follows the first delivery, because every later license inherits its costs and because it is where diligence questions most often go unanswered.

Start with the denominator. Under ASC 606 as amended by ASU 2016-08, issued in March 2016, a company that controls the specified good or service before it is transferred to the customer is a principal and reports the gross amount as revenue, while an agent reports only its fee or commission. The same data transaction can therefore appear as large revenue with a thin gross margin or as small revenue with a high one. A comparable view of data company gross margin starts from net revenue: the buyer's payment minus the supplier's payout.

Every figure below is indexed to that net revenue, set at 100 points. The split between buyer payment and supplier payout varies by deal and is deliberately not modeled; the waterfall starts where the intermediary's own economics start.

Follow one dataset: where each cost lands between first call and cash

A dataset passes eight gates on its way to cash, and each gate adds cost, ends the candidate or moves money. An investor should be able to point to a dated event that closes every gate, because a gate without a closing event cannot be costed or audited.

Follow one dataset: where each cost lands between first call and cash
GateWork and cost that land hereEvent that closes the gateEvidence to request
1. Sourcing and qualificationConversations, a check of systems and volume, a fit reviewCandidate qualified or declined with a reason codePipeline log with entry and exit dates
2. Supplier contractingScope, terms and signaturesAgreement signed before any data movesSignature date for each dataset
3. Rights reviewCustomer contracts, privacy notices, third-party contentRights sign-off or rejectionReviewer hours and rejection reasons
4. PreparationExtraction, de-identification, linking, documentationDataset passes internal checksTime records and pipeline run logs
5. DeliveryTransfer, drive shipment or seller-held accessDelivery receiptTransfer logs and shipping records
6. Acceptance reviewBuyer questions, fixes and repeat cyclesAcceptance recorded, full or partialAcceptance notices and fix tickets
7. Invoicing and collectionInvoice, payment terms, disputesBuyer payment receivedReceivables aging by buyer
8. Supplier paymentPayment under the supplier's termsSupplier paidPayment dates set against buyer collection dates

How should rejected candidates be charged to the datasets that paid?

Charge them to a cohort: every candidate that entered qualification in the same period shares its unsuccessful hours across the datasets from that period that were accepted. Charging failed work to overhead makes each accepted dataset look cheaper than the process that produced it.

Define the cohort by entry date, not close date, so a quarter with many late rejections cannot push them into the next quarter's overhead. Use one allocation driver and state it. Hours are the simplest, because they can be audited against time records.

Illustrative cohort: 10 candidates enter qualification in one quarter and 4 become accepted deliveries. The 6 that fail consume 280 hours between them, so each accepted dataset carries 70 hours of work it never used. At this article's illustrative conversion of 5 person-hours per point of net revenue, that is 14 points.

How should rejected candidates be charged to the datasets that paid?
Where the candidate stoppedCandidatesHours spent on eachHours carried by the 4 accepted datasets
Qualification: no buyer fit3412
Rights review: a customer contract bars outside use22448
Preparation: records lack the outcome fields buyers test1220220
Total unsuccessful64 to 220280, or 70 per accepted dataset

One dataset, every dollar: the transaction waterfall in points of net revenue

The waterfall turns one accepted dataset into eight lines, from net revenue to loaded contribution, and marks where a reported deal margin usually stops. In the illustration, the deal margin is 66 points and the loaded contribution is 40.

Lines 2 to 5 are what a deal margin normally includes. Lines 6 to 8 are what diligence adds, and here they remove 26 points, about two fifths of the reported margin. The single candidate that failed during preparation accounts for 11 of those 26 points on its own, more than all rework.

The metered line is small for a reason an investor can check: machine work is priced per page or per gigabyte, and the prices are published. On October 26, 2026, Google Cloud listed Document AI's Form Parser at $30 per 1,000 pages for the first million pages a month; its Sensitive Data Protection pricing example came to $34 for inspecting 10 GB and transforming 2 GB; internet data transfer from Cloud Storage to worldwide destinations outside Asia and Australia was $0.12 per GiB for the first 10 TiB a month; and a 40-terabyte Transfer Appliance was $300 per use with 10 days included. A margin gap rarely hides in lines billed by the gigabyte. It hides in hours.

Editable means each input is a cell an analyst replaces with the company's own records: hours by gate from time records, cohort counts from the pipeline log, carrying cost from receivables aging and the company's cost of capital. A line the company cannot fill from records is an estimate and should be labeled as one.

One dataset, every dollar: the transaction waterfall in points of net revenue
LineIllustrative pointsRunning totalWhat sets itEvidence to request
1. Net revenue: buyer payment minus supplier payout100100License price and supplier termsSigned license and supplier agreement for the dataset
2. Rights review, 40 hours-892Contracts, notices and third-party content to clearReviewer time records
3. Preparation, 100 hours-2072Extraction, de-identification, linking, documentationTime records and pipeline logs
4. Metered tools and shipping-270Pages processed, gigabytes moved, drives shippedProcessor and carrier invoices
5. Delivery and acceptance support, 20 hours-466: reported deal marginBuyer questions during reviewSupport ticket log
6. Unsuccessful candidates in the cohort, 70 hours-1452The gate at which candidates failCohort ledger by entry date
7. Rework before acceptance, 45 hours-943Repeat preparation cyclesRework ledger with triggers
8. Cash timing-340: loaded contributionDays between spending and collectingReceivables aging and payment dates
A horizontal waterfall in points of net revenue falling from 100 to a 66-point deal margin and then to a 40-point loaded contribution after failed candidates, rework and cash timing, with a 115-day cash gap marked below.
One dataset, every dollar, indexed to 100 points of net revenue. Illustrative: values are assumptions, not measured data.

When does cash arrive, and who carries the gap until it does?

Cash arrives last: preparation spending starts weeks before delivery, and the buyer pays only after acceptance and its invoice terms. Whoever pays out before the buyer pays in carries that gap, and the carrying cost belongs in unit economics, not only in the cash flow statement.

In the illustration, 115 days separate the first preparation hour from the buyer's payment. Two supplier structures change who funds that window. Paid-when-paid terms push the wait onto the supplier and lower the intermediary's working capital, at the price of a less attractive offer. A fixed payment schedule, independent of the buyer, makes the offer stronger and moves the gap, along with any dispute risk, onto the intermediary's balance sheet.

Ask for distributions, not averages: days from delivery to acceptance and from invoice to cash, by buyer, with cases still open shown as open. One slow buyer can set a company's working capital needs.

When does cash arrive, and who carries the gap until it does?
EventIllustrative dayCash movementWho carries it
First contact; qualification startsDay 0Sourcing hours, nothing inIntermediary
Supplier agreement signedDay 21NoneNeither side yet
Rights cleared; preparation startsDay 35Preparation spending beginsIntermediary
Delivery to the buyerDay 75Transfer or drive shipment paidIntermediary
Acceptance recorded; invoice issuedDay 105Receivable opensIntermediary until paid
Buyer pays the invoiceDay 150Cash inThe buyer's terms set the date
Supplier paidDepends on the contractCash outSupplier if paid after the buyer; intermediary if paid on a fixed schedule

Stress the baseline: a slow review and a partial acceptance

A partial acceptance does more damage than a slow review, because the buyer's smaller payment arrives after the full preparation cost is spent. Model each event as its own column beside an unchanged baseline, so the effect of each is visible on its own and the baseline stays comparable from one review to the next.

The slow review costs 6 points: 15 more hours of buyer support and 45 more days of carried balance. The partial acceptance costs 34 points, because revenue falls by 30 while every cost line except rework stays where it was, and rework rises by 20 hours to cut the accepted subset. Pricing per accepted record moves the risk of a partial acceptance to the seller; a written acceptance test agreed before preparation is the protection that addresses it.

Stress the baseline: a slow review and a partial acceptance
Line, in points of net revenueBaselineSlow review: 45 more daysPartial acceptance: 70 of 100 records
Net revenue10010070, priced per accepted record
Direct costs: rights, preparation, tools, delivery support343734, already spent
Unsuccessful candidates141414
Rework9913
Cash timing363
Loaded contribution40346
Contribution as a share of net revenue40 percent34 percentAbout 9 percent
Illustrative days from first contact to cash150195150

Which operating improvement moves contribution most?

In this model, catching a doomed candidate at rights review instead of after preparation moves contribution most, by about 10 points, more than halving rework or preparing 20 percent faster. Hours spent on a candidate that never pays carry no revenue at all, so moving a rejection earlier is worth more than making accepted work cheaper.

The ranking depends on the illustrative inputs, so run the same five levers on a company's own ledger. The pattern worth testing is whether the order of checks, rights and buyer fit before preparation, does more for contribution than any efficiency program inside preparation itself.

Faster preparation deserves a caveat. If speed raises repeat cycles, the 4 points saved on line 3 come back through line 7, so ask for both figures from the same period.

Which operating improvement moves contribution most?
LeverChange modeledLoaded contribution, baseline 40Why it moves
Screen rights before preparationThe candidate that failed after 220 preparation hours is caught at rights review after 24About 50Removes 196 unpaid hours from the cohort, 49 per accepted dataset
Halve repeat cyclesRework falls from 45 to 22.5 hours44.5Fewer passes over finished work
Prepare 20 percent fasterPreparation falls from 100 to 80 hours44Direct hours fall, if rework does not rise
Shorten acceptance review by 30 daysCash timing falls from 3 to 1 point42Less balance carried; cash arrives a month earlier
Agree the acceptance test before preparationAvoids the partial-acceptance case40 instead of 6Protects revenue rather than cutting cost

What could make these unit economics look better than they are?

Four recording practices can flatter a data licensing company's numbers, and one market risk can shrink the number of datasets however good each one is. Test each before trusting any margin.

  • Gross presentation. A company reporting as principal shows buyer payments as revenue, which makes revenue larger and percentage margin smaller; one reporting as agent shows the reverse. ASU 2016-08 sets the control test, and net revenue is the comparable base either way.
  • Overhead absorption. Unsuccessful candidates and repeat cycles booked to general overhead disappear from deal margin. Ask for the cohort ledger and for the share of preparation hours with no accepted dataset attached.
  • Survivor cohorts. Cohorts defined by close date drop candidates still in progress. Define them by entry date and show open cases.
  • Collection timing. Accepted but unpaid deliveries look like finished revenue. Ask for receivables aging by buyer and a count of partial acceptances.
  • Buyer budgets. Demand for licensed data depends on AI developers' spending holding up. In September 2023, Sequoia's David Cahn estimated that each year of GPU spending at then-current levels implied about $200 billion of lifetime revenue to pay it back, leaving a gap of $125 billion or more, and in June 2024 he put the figure at about $600 billion. If end revenue lags that spending, data budgets are an easy line to cut, and strong contribution per dataset offers no protection against fewer datasets.

Questions to ask a data licensing company, SourceX included

Ask for each line of the waterfall from records, with the cohort and dates stated, and put the same questions to SourceX.

SourceX is one application of this analysis. For businesses licensing operational data to AI developers, it handles every stage the waterfall charges (supplier sourcing and qualification, rights review, the contract, data preparation, delivery and payment), so every gate sits inside one operator. No data reaches SourceX until the company has signed the agreement, having first approved its scope and terms. For large deliveries, the records are kept in the seller's own storage or sent on encrypted drives rather than hosted by SourceX.

None of SourceX's unit economics, cohort ledgers or cash timing is public, and company and commercial metrics are shared with investors one to one. Its raise is open to accredited investors under Rule 506(c); investor materials can be requested through the investor page, where a reader can bring any of the questions below.

Checklist · 6 items

  1. Which cohort definition do you use, and what share of preparation hours in the last four cohorts went to candidates that were never accepted?
  2. At which gate do most unsuccessful candidates stop, and how many hours had been spent on them by then?
  3. How many repeat preparation cycles did the average accepted dataset need, and who triggered them?
  4. What were the days from delivery to acceptance and from invoice to cash, by buyer, with open cases shown?
  5. How often has a buyer accepted only part of a delivery, and how was the price set when that happened?
  6. Do you report revenue gross or net, and what was contribution on net revenue for the same period?

0 of 6 ticked. Tick each item as you go.

Methods note: inputs, sources and what was not measured

The gate table, cohort ledger, waterfall, cash timeline, scenarios and levers are an editorial model built for this article on 2026-10-26. Every number in them is illustrative: net revenue is indexed to 100 points, person-hours convert at 5 hours per point, and the cohort, day counts and carrying costs are assumptions chosen to show the mechanics, not measurements of SourceX or any other company.

Public inputs come from FASB's ASU 2016-08 on principal and agent reporting, Google Cloud's published price lists for Document AI, Sensitive Data Protection, Cloud Storage data transfer and Transfer Appliance as listed on 2026-10-26, and Sequoia's 2023 and 2024 essays on AI revenue. Not measured: actual hours by gate, cohort conversion, acceptance or collection days, buyer concentration, or contribution at any company. Each requires records that have not been published.

Frequently asked questions

What is a good gross margin for a data licensing company?

There is no reliable public benchmark, and the reported figure depends first on whether the company presents revenue gross, as a principal, or net, as an agent, under ASC 606. Compare contribution on net revenue instead, after unsuccessful candidates, repeat preparation and carrying cost are charged. A high margin with those lines left out says little.

Should rejected datasets count in unit economics?

Yes. Charge the hours spent on candidates that failed qualification, rights review or preparation to the accepted datasets from the same entry cohort, using a stated driver such as hours. Otherwise each accepted dataset looks cheaper than the process that produced it.

Is the supplier payout a cost or a reduction of revenue?

It depends on the principal or agent assessment. If the intermediary controls the dataset before it is transferred to the buyer, it reports the buyer's payment as revenue and the supplier's share as a cost; if not, it reports only its net amount. For unit economics, compare companies on net revenue either way.

Why does cash timing belong in unit economics?

Because an intermediary often spends on preparation months before the buyer pays, and may pay the supplier before collecting. Carrying that balance has a cost of capital, and a slow or partial acceptance raises it. In this article's illustration, a review that ran 45 days longer cut contribution from 40 to 34 points.

Does SourceX publish its unit economics?

No. SourceX has not published unit economics, cohort ledgers or cash timing, and it shares company and commercial metrics with investors directly. Materials can be requested through sourcex.si/investors, where an investor can ask for the definitions behind any unit-economics figure; any offering is open only to verified accredited investors.

Sources

  • FASB's ASU 2016-08 (March 2016) says an entity that controls the specified good or service before it is transferred to the customer is a principal and recognizes the gross amount, while an agent recognizes only its fee or commission, and that one contract can make an entity principal for some goods or services and agent for others. Source
  • Google Cloud's Document AI price list showed Form Parser at $30 per 1,000 pages for up to 1,000,000 pages a month and $20 above that, as listed on October 26, 2026. Source
  • Google Cloud's Sensitive Data Protection pricing example shows inspecting 10 GB at $3.00 per GB and transforming 2 GB at $2.00 per GB, a total of $34.00. Source
  • Google Cloud's Cloud Storage price list showed data transfer to worldwide destinations excluding Asia and Australia at $0.12 per GiB for the first 10 TiB a month, as listed on October 26, 2026. Source
  • Google Cloud's Transfer Appliance price list showed the 40-terabyte appliance at $300 per use with 10 days included and $30 for each further day, with U.S. freight typically starting around $120 round trip. Source
  • In September 2023, Sequoia's David Cahn estimated that each year of GPU capital spending implied about $200 billion of lifetime revenue needed to pay it back, leaving a gap of $125 billion or more per year at then-current levels. Source
  • In June 2024, Sequoia's David Cahn reran the analysis and put the revenue question at about $600 billion. 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