Deal economics
Partial dataset acceptance economics: five pricing terms, one subset
By SourceX Editorial · Reviewed by Noah Loul ·
10 sections
Short answer
Partial dataset acceptance economics turn on one ratio: the share of a prepared dataset the buyer selects, set against preparation costs that do not shrink when the buyer takes less. In this article's illustrative model, a deal earning 20 points at a full pick breaks even at a 60 percent pick and loses 10 at 40 percent. Pricing terms agreed before preparation decide who carries the records left behind.
Value drivers
Partial dataset acceptance
Key takeaways
- Most preparation is spent before the buyer chooses, while revenue priced per selected record shrinks with every record left out.
- In the illustrative model, contribution falls from 20 points at a full pick to minus 10 at a 40 percent pick, and break-even sits at a 60 percent pick.
- A minimum commitment protects margin only when it sits at or above the break-even pick; a 50-unit floor still loses money below 60 percent.
- A preparation charge or staged preparation lowers the break-even pick, to 37.5 and about 43 percent in the same model.
- Value the unselected remainder at zero until a second license is signed, because it is what one buyer chose not to take.
- Ask for contracted, prepared, selected and invoiced scope as separate fields on every deal, and for the distribution rather than the average.
Partial dataset acceptance economics: what changes when a buyer takes less?
Partial acceptance turns a fixed cost into a bet on the buyer's pick. Most of the work of preparing a dataset happens before the buyer chooses what to take, while revenue priced per selected record shrinks with every record left out. A deal that looks profitable on its full scope can lose money on a subset, and the size of the loss is set by terms agreed before preparation began.
Partial acceptance here means the buyer accepts and pays for part of a dataset that was prepared in full: some years, some record types, or the records that survive its filters. That is a different question from how many candidate datasets are accepted at all, which belongs in a separate acceptance-rate ledger. This article models the subset question, tests five pricing terms against it and lists the evidence an investor should request. Every value is illustrative, and no SourceX results are reported.
The central quantity is the selection share: selected units divided by prepared units, deal by deal. Anyone assessing data licensing unit economics should see its distribution across deals, not its average, because the losses sit in the low picks.
Three ways a buyer ends up taking a subset
Buyers end up with subsets for three different reasons, and each has its own fix: they choose a narrower scope, their filters remove records, or rights and privacy review removes records before the buyer sees them.
Scope choice is the cheapest to prevent. A buyer that wants only recent years, one language or one workflow can say so before preparation starts, provided someone asks.
Filtering is routine in AI data work, and it happens after the data exists. In the DataComp benchmark (2023), the flagship DataComp-1B set is a 1.4 billion sample subset filtered from a 12.8 billion sample candidate pool. The FineWeb paper (2024) describes FineWeb-Edu as a 1.3 trillion token subset of the 15 trillion token FineWeb corpus, picked by an educational-quality classifier. Deduplication removes more: the 2021 paper Deduplicating Training Data Makes Language Models Better reports a single 61-word sentence repeated more than 60,000 times in the C4 corpus, and train-test overlap touching over 4 percent of the validation sets of standard datasets. These are open web pools, not licensed business records, so their ratios do not transfer. The habit of filtering does.
Evaluation buyers filter for another reason. When researchers built GSM1k in 2024 as a fresh counterpart to the GSM8k grade-school math benchmark, they found accuracy drops of up to 8 percent for some models (arXiv 2405.00332, current version), a sign of contamination. A buyer assembling an evaluation set has reason to drop records that resemble public material, which shrinks the paid subset again.
Exclusions start on the supply side but cost the same. Under the HIPAA Safe Harbor method, identifiers of the individual and of relatives, employers and household members must be removed before health information counts as de-identified; free-text records where that cannot be done reliably leave the scope after the preparation money is spent.
Which preparation costs stay fixed when the buyer takes less?
Anything spent to make the dataset ready before the buyer picks is fixed with respect to the pick: rights review, contracting, schema mapping and pipeline setup, de-identification method design and validation, documentation and the sample. Only work done per delivered record, such as final checks and transfer, and any share of revenue owed to the data owner, falls when the buyer takes less.
The split is a design choice, not a law of nature. A step moves from fixed to variable when it is done only for the selected scope, which is the logic of staged preparation. Rights review cannot move, because nothing can be shown to a buyer before it is done. How the main cost lines usually behave against the buyer's pick:
Checklist · 6 items
- Rights review and contracting: fixed, and they must precede any sample.
- Schema mapping, linking and pipeline setup: mostly fixed, done once per source system.
- De-identification method design and validation: fixed; per-record checks on the selected scope are variable.
- Documentation and evidence pack: mostly fixed; the delivered manifest changes with the pick.
- Final quality checks and transfer: variable with selected units.
- The data owner's share, where it is a share of revenue: variable with revenue, so it falls with the pick and absorbs part of the loss.
0 of 6 ticked. Tick each item as you go.
The transaction waterfall at a full and a 40 percent pick
In the illustrative model, the same prepared dataset earns 20 index points at a full pick and loses 10 at a 40 percent pick, because 30 points of preparation do not shrink. Contribution equals headroom per selected unit times selected units, minus fixed preparation, so the break-even selection share is fixed preparation divided by headroom on the full scope: 30 divided by 50, or 60 percent.
The inputs are editable assumptions, not benchmarks: 100 units of offered scope, all prepared; a price of 1.0 per selected unit, so a full pick brings 100 points of revenue; a data owner's share of 40 percent of revenue; 0.10 per selected unit for final checks and delivery; and 30 points of fixed preparation. Replace each with a deal's own records to rerun it. Line by line, at both picks:
| Line | Formula | Full pick (100 units) | 40 percent pick (40 units) | What moves it |
|---|---|---|---|---|
| Revenue | Price per unit times selected units | 100 | 40 | Price and selection share |
| Data owner's share | 40 percent of revenue (illustrative) | -40 | -16 | Owner terms |
| Final checks and delivery | 0.10 times selected units | -10 | -4 | Per-record work |
| Headroom before preparation | Revenue less the two lines above | 50 | 20 | Headroom of 0.5 per selected unit |
| Fixed preparation | Rights review, contracting, pipeline, method validation, documentation | -30 | -30 | Scope agreed before preparation |
| Contribution | Headroom less fixed preparation | 20 | -10 | Break-even at a 60 percent pick |
Five pricing terms tested on a 40 percent pick
Pricing terms decide who carries the unselected scope, and every term in the test earns the same 20 points at a full pick. They differ only below it: a minimum commitment protects margin only when it sits at or above the break-even pick, while a preparation charge or staged preparation lowers the break-even itself.
The 50-unit minimum is the instructive failure. It caps the loss at 5 points but never turns contribution positive below a 60 percent pick, because the floor sits under the break-even line. A 60-unit minimum keeps contribution at zero or better at every pick. A flat license for the full scope would earn 26 points at a 40 percent pick, but it hands the whole selection risk to the buyer, who has little reason to accept that before testing the data.
The terms also change what reaches reported revenue. Pricing per selected unit is variable consideration under Topic 606, which admits an estimate into revenue only to the extent it is probable that a significant reversal will not occur, and the standard sends entities to paragraphs 606-10-55-85 through 55-88 to judge how a customer acceptance clause affects when control passes. Contract value at signing can therefore overstate what a subset deal recognizes. Each term at a 40 percent pick, with the pick at which it breaks even:
| Pricing term | How revenue is set | Contribution at a 40 percent pick | Break-even pick | Who carries the unselected scope |
|---|---|---|---|---|
| Per selected unit, no floor | 1.0 per selected unit | -10 | 60 percent | The data business |
| Per unit with a 50-unit minimum | 1.0 per unit, at least 50 units billed | -4 | 60 percent; the floor sits below break-even | Mostly the data business |
| Per unit with a 60-unit minimum | 1.0 per unit, at least 60 units billed | +2 | None: zero or better at every pick | The buyer, up to the floor |
| Preparation charge plus unit price | 30 up front plus 0.70 per selected unit | +0.8 | 37.5 percent | Shared: the charge recovers most preparation |
| Staged preparation | 1.0 per unit; half of preparation spent only on the selected scope | -1 | About 43 percent | The data business, on a smaller fixed base |
Is the unselected remainder inventory or a write-off?
Treat it as a write-off until a second license is signed. The remainder is, by construction, what one buyer chose not to take, so it is likely weaker than the dataset's average, and its value rests on rights the first deal may already have narrowed.
Three tests decide whether it is inventory. Rights: the data owner's agreement must allow another license, and the first buyer's terms must not grant exclusivity over the full scope. Fit: another buyer must want what the first one passed over, which is plausible when the reason was scope, such as another language or period, and unlikely when it was quality. Freshness: operating records age, and a remainder held for a year may need part of its preparation again.
An investor should ask how much prepared but unselected scope was later licensed, how soon and at what price relative to the first sale. Without that history, the remainder belongs at zero in any model of partial dataset acceptance economics.
Selection effect or something else? Four causes of a thin margin
A thin deal margin does not prove that partial acceptance caused it. Rework, shared overhead and slow collection can erase a reported margin just as completely, and each needs a different fix, so separate them before blaming the buyer's pick.
At a glance
Selection effect or something else? Four causes of a thin margin
The selection effect can be measured on its own: the unselected share times headroom per unit times prepared scope. In the illustrative model, a 40 percent pick forgoes 60 units at 0.5 points each, or 30 points, which is exactly the gap between contribution at the full pick and at the partial one. Ask for each cause as its own line on every deal:
- Selection effect: prepared units, selected units and headroom per unit, so the forgone headroom is computed rather than estimated.
- Rework: preparation repeated after a failed check, logged by trigger and hours and kept out of the selection line.
- Overhead: shared staff and tooling allocated by a stated rule, so a deal does not look thin merely because it absorbed more of the allocation.
- Collection delay: days from delivery to cash and the cost of financing that gap, which grows when a dispute over the subset holds up acceptance.
What could make the partial-acceptance analysis wrong
Five developments would shrink the problem or change its shape, and each is a reason to test the model against a company's own deal records rather than adopt it.
- Buyers might fix scope up front. If most selection is agreed before preparation, partial acceptance becomes rare and fixed cost is spread over a known pick.
- Preparation might get cheaper. Automation that halves fixed preparation to 15 points halves the break-even pick to 30 percent in the same model.
- Few buyers may set the terms. Reddit's FY2025 Form 10-K says that, to date, "substantially all of the contract value associated with our licensing revenue is derived from two of our partners." A supplier facing that few buyers may not win minimum commitments, which keeps the selection risk on its side.
- The remainder might resell well. A company with a record of licensing prepared but unselected scope deserves a softer assumption than a write-off.
- The inputs are illustrative. The owner's share, preparation cost and price per unit vary by deal, and the break-even pick moves with each; the formula is the durable part.
Questions on partial acceptance for any data licensing company, SourceX included
Ask for the distribution, not the average. A company can show a healthy mean selection share and still lose money on its bottom third of deals, and a data company gross margin figure built on averages will hide it.
SourceX has not published selection-share or deal results. It is raising from accredited investors under Rule 506(c), shares investor materials on request through its investors page and discusses company and commercial metrics with investors directly; the questions apply to it exactly as they apply to any other company in the market. Questions to put to any data licensing company:
- What is the distribution of selection share across deals, with contracted, prepared, selected and invoiced scope kept as separate fields?
- Which preparation costs are fixed with respect to the buyer's pick, and what share of total deal cost do they make up?
- Which pricing terms are in use, and where does each put the break-even pick?
- How much prepared but unselected scope was later licensed, how soon and at what relative price?
- When did buyers specify scope: before preparation, after a sample or after delivery?
- How does reported revenue treat per-unit pricing and acceptance clauses, and how often has recognized revenue reversed?
Methods note: how the selection model was built
The model is a single-deal contribution calculation with five illustrative inputs: 100 units of prepared scope, a price of 1.0 per selected unit, a data owner's share of 40 percent of all revenue, 0.10 per selected unit for final checks and delivery, and 30 points of fixed preparation. Contribution equals 0.5 times selected units minus 30, and the break-even pick is fixed preparation divided by full-scope headroom. Pricing variants change only the revenue line, except staged preparation, which moves 15 of the 30 fixed points onto the selected scope; the preparation charge is assumed to carry the same owner's share as other revenue.
Public sources supply context, not inputs. The DataComp, FineWeb and deduplication papers show that AI teams filter data after it exists; GSM1k shows why evaluation buyers screen for contamination; the HIPAA Safe Harbor rule shows one route by which records drop out; Topic 606 sets how variable consideration and acceptance clauses reach revenue; and Reddit's filing shows how concentrated licensing demand can be. Sources were checked on October 26, 2026.
Not measured: real selection-share distributions, the fixed share of preparation cost at any company, resale rates for unselected scope and how often buyers set scope before preparation. Whole-dataset rejections are excluded and belong in a separate acceptance-rate ledger; the unselected remainder is treated as rejected scope whose preparation is charged to the selected part. No SourceX data is used.
Frequently asked questions
What is a selection share in data licensing?
It is the share of a prepared dataset the buyer accepts and pays for: selected units divided by prepared units, deal by deal. It differs from an acceptance rate, which counts how many candidate datasets are accepted at all.
Why does a partial pick hurt margin more than proportionally?
Most preparation is spent before the buyer chooses, so revenue falls with the pick while preparation does not. In the illustrative model, cutting selected units by 60 percent turns a 20-point contribution into a 10-point loss.
Does a minimum purchase commitment solve partial acceptance?
Only if the floor sits at or above the break-even pick. In the illustrative model, a 50-unit minimum still loses money at every pick below 60 percent, while a 60-unit minimum keeps contribution at zero or better.
Should unselected records count as inventory?
Not until a second license is signed. The remainder is what one buyer chose not to take, and its value depends on rights, fit with another buyer and how quickly the records age.
What should investors ask a data company about partial acceptance?
Ask for the distribution of selection share by deal, the fixed share of preparation cost, the pricing terms in use and the break-even pick each implies, and the history of relicensing unselected scope.
Sources
- The DataComp benchmark's DataComp-1B set is a 1.4 billion sample subset filtered from a 12.8 billion sample candidate pool. Source
- FineWeb-Edu is a 1.3 trillion token subset of the 15 trillion token FineWeb corpus, selected by an educational-quality classifier. Source
- Lee et al. (2021) found a 61-word sentence repeated over 60,000 times in C4 and train-test overlap affecting over 4 percent of standard validation sets. Source
- The GSM1k paper's current version reports accuracy drops of up to 8 percent against GSM8k for some models, consistent with contamination. Source
- The HIPAA Safe Harbor method requires removing listed identifiers of the individual and of relatives, employers and household members. Source
- Topic 606 includes variable consideration only to the extent a significant revenue reversal is probable not to occur, and directs acceptance-clause analysis to paragraphs 606-10-55-85 through 55-88. Source
- Reddit's FY2025 Form 10-K states that substantially all of the contract value associated with its licensing revenue is derived from two partners. Source
Request SourceX investor materials
SourceX shares company and commercial metrics with investors directly. Any offering is made only to verified accredited investors.