Deal economics
Large dataset transaction profitability: four costs that grow faster than price
By SourceX Editorial · Reviewed by Noah Loul ·
10 sections
Short answer
Large dataset transaction profitability often falls as deals grow, because four costs rise faster than price: bespoke preparation, months of delay before cash, concentration on one buyer and obligations that outlive delivery. In this article's illustrative model, a deal at four times the scope earns a 6.7 percent loaded margin against 25.2 percent at the base size, while its reported deal margin still looks healthy. Price each cost separately.
Value drivers
Large dataset transaction
Key takeaways
- Price grows roughly with scope, minus volume discounts; bespoke preparation, delay, concentration and service obligations can grow faster.
- Reported deal margin, price minus records and standard preparation, stays healthy as deals grow and hides all four costs.
- In the illustrative model, loaded contribution peaks at twice the base scope and turns negative at eight times.
- Public filings show the concentration risk: Innodata attributed about 58 percent of 2025 revenue to one customer, and Appen lost a Google contract worth $82.8 million of FY23 revenue.
- Structure rescues large deals: paid change orders, milestone payments, narrower exclusivity and capped service terms.
When does a bigger data deal earn less than a smaller one?
When the costs that grow faster than price outrun the fixed costs a larger deal spreads. A data licensing deal at several times the usual scope tends to bring a buyer-specific specification, a longer path to cash, a larger share of one buyer in the business and obligations that continue after delivery. Each of those grows faster than the headline price, while the main advantage of size, spreading fixed sales and contracting work, grows slowly.
That is why large dataset transaction profitability should be judged on loaded contribution, not on the reported deal margin most deal ledgers show. The reported margin subtracts records and standard preparation work and stops. The four faster-growing costs land elsewhere: in overhead, in working capital, in a renewal that does not happen or in support work two years later.
This article models the effect on one hypothetical deal at four sizes, tests each mechanism against public evidence and ends with the structural changes that make a large deal worth taking. Every model value is an illustrative index point, not a SourceX price, deal size or margin; those are not public.
The scale model: one hypothetical deal at 1x, 2x, 4x and 8x scope
The model holds the dataset type constant and scales its scope, indexing the base deal's headline price to 100 points. Price per unit of scope falls with size through illustrative volume discounts of 5, 12.5 and 20 percent, records and standard preparation scale in line with scope at 55 points per unit, and fixed deal costs for sales, contracting and the evidence pack rise slowly.
The four faster-growing lines follow rules a reader can change. Bespoke preparation is 4 points times scope to the power 1.5, because larger buyers write more of their own specification. Service obligations step up by tier. Delay cost is direct spend times months to cash (3, 5, 8 and 12) times 1 percent a month, assuming the spend happens at the start; halve it if spend is spread evenly. The concentration reserve is a probability-weighted concession, set at 2, 4, 7 and 11 percent of headline price as the buyer's share of the business rises.
Reported deal margin climbs from 45 to 200 points as the deal grows. Loaded contribution rises from 25.2 points to 38.3 at twice the scope, falls back to 23.3 at four times, and becomes a loss of 82 points at eight times. Per month of cycle, the base deal earns about 8.4 points and the 4x deal about 2.9 (our arithmetic). The figure plots the two lines.
The scale model in index points, with every value illustrative:
| Line (index points) | 1x scope | 2x scope | 4x scope | 8x scope |
|---|---|---|---|---|
| Headline price | 100 | 190 | 350 | 640 |
| Records and standard preparation | 55 | 110 | 220 | 440 |
| Reported deal margin | 45 (45.0%) | 80 (42.1%) | 130 (37.1%) | 200 (31.3%) |
| Fixed deal costs: sales, contracting, evidence pack | 10 | 11 | 12 | 14 |
| Bespoke preparation | 4 | 11 | 32 | 90 |
| Service obligations after delivery | 2 | 6 | 18 | 44 |
| Delay cost at 1% a month on direct spend | 1.8 | 6.1 | 20.2 | 63.6 |
| Concentration reserve | 2.0 | 7.6 | 24.5 | 70.4 |
| Loaded contribution | 25.2 | 38.3 | 23.3 | -82.0 |
| Loaded margin on headline price | 25.2% | 20.2% | 6.7% | -12.8% |
Why large buyers rarely take the standard cut of a dataset
Large buyers usually buy to a specification of their own, and work done to one buyer's specification is hard to reuse. That is the core of the bespoke line: custom schemas, buyer-defined labels, handling for the edge cases one buyer cares about and repeated sample rounds, none of which the next buyer will pay for.
Evaluation purchases push the effect further. When researchers built GSM1k, a fresh set of grade-school math problems designed to mirror the public GSM8k benchmark, they observed accuracy drops of up to 8 percent for leading models, with several model families showing signs of systematic overfitting, according to the current arXiv version of the paper. The authors said they did not intend to release GSM1k publicly at that time, to avoid the same contamination. A buyer that wants an uncontaminated test set has the same incentive: keep the data unpublished and often exclusive, which removes the seller's chance to spread bespoke cost across later licenses.
The input that matters is the exponent. At 1.5, bespoke work at eight times the scope costs about 22 times the base amount; at 1.0 it would cost 8 times. Ask a company for bespoke hours by deal size. If they grow faster than scope and are not billed separately, the larger deals are being subsidized by the smaller ones, and the lost resale is a second cost that the exclusivity worksheet linked below prices on its own.
How much do the months between spending and cash cost?
In the model, delay is the third-largest drag on the 8x deal: 63.6 points, almost a third of its reported margin. Larger deals take longer to negotiate, deliver in more tranches, sit in longer acceptance reviews and often run over a multi-year term, so the seller funds more work for longer before cash arrives.
Public filings show how long large licensing commitments run. Reddit's quarterly report for the period ended March 31, 2026 put its remaining performance obligations on contracts longer than one year at $120.6 million, mostly from long-term content licensing contracts. When obligations stretch over years, cost, revenue recognition and cash can fall in different periods, and a deal ledger that books the whole margin at signing hides the gap.
Two numbers capture the effect better than margin. Contribution per month of cycle shows what a deal earns for each month of the team's attention: about 8.4 points for the base deal and 2.9 for the 4x deal in the model. Peak funded spend shows how much work is outstanding before the first payment. The delayed-payment comparison linked below covers discounting and payment structures in depth; this model needs only a monthly cost of capital, which should come from the company's own financing.
One buyer, one decision: pricing concentration into the deal
A large deal raises the share of the business that one buyer controls, and that share has a cost even if the buyer never leaves: it moves negotiating leverage toward the buyer and turns a single renewal decision into a material event. The concentration reserve prices it as an expected concession, the probability of a cancellation, renegotiation or partial acceptance times the value at risk.
Public AI data companies show the exposure. Innodata's annual report for 2025 says one customer in its Digital Data Solutions segment generated approximately 58 percent of total revenues in 2025 and 48 percent in 2024. Appen's February 12, 2024 market update said Google, which had given notice to end its contract with projects stopping by March 19, 2024, contributed $82.8 million of FY23 revenue at a 26 percent gross margin. Reddit's Form 10-K for fiscal 2025 states that "substantially all of the contract value associated with our licensing revenue is derived from two of our partners."
None of these figures says large customers are bad business; each company chose those relationships for reasons the filings explain. They show why a model needs a line for the downside. In the illustrative model the reserve climbs from 2 to 11 percent of headline price as the deal's share of the business grows, and at 8x it reaches 70.4 points, most of the base deal's entire price. A company that sets this line to zero for its largest deal is assuming the renewal.
Which obligations outlive delivery, and who pays for them?
Large buyers ask for obligations that continue after the data is delivered: refreshes, answers to questions about the data, documentation updates, audit cooperation, deletion on request and confirmation that deletion happened. Each is small per event and easy to leave out of a deal model, and together they move the service line from 2 to 44 points across the four sizes.
Regulation gives buyers reasons to ask. California's AB 2013 requires developers of generative AI systems to post documentation about their training data, including dataset sources or owners and whether the data was purchased or licensed, on or before January 1, 2026. The European Commission's template of July 24, 2025 asks general-purpose model providers to summarize training content, private datasets included, and to update the summary at six-month intervals when further training requires it. A buyer with those duties will want its supplier to keep answering questions for as long as the model is on the market.
Deletion duties can reach beyond the data. In May 2021 the Federal Trade Commission finalized an order requiring the photo app developer Everalbum to delete models and algorithms developed in whole or in part using biometric information from its users' photos and videos. A buyer that has seen remedies of that kind will ask large suppliers for warranties, indemnities and deletion support, and the seller carries the cost of honoring them for years. Capping that work in the agreement, or pricing it as a separate service, is how the service line stops outgrowing price.
Is deal size the problem, or the way it was priced?
Often the pricing. The strongest objection to this model is that bigger deals look worse only because they were priced as if they were small ones, and the symptom is a reported deal margin that disappears once rework, overhead and collection delays are counted. If that is the cause, the remedy is structure, not avoiding large buyers.
Test it by comparing deals of different sizes at similar scope complexity. If loaded margin falls only where bespoke hours, cycle months or service events rise, the cause is unpriced work rather than size itself. Then check that loaded costs across all deals reconcile to the hours and cash the company actually spent in the period; costs that appear in no deal are where large-deal losses usually hide.
In the model, four structural changes rescue the 4x deal. Together they lift loaded contribution from 23.3 to about 94.9 points, a 24.8 percent loaded margin on 382 points of revenue once bespoke change orders are billed (our arithmetic), close to the base deal's margin on almost four times the revenue. The levers, in order of effect:
- Bill bespoke work as change orders against a written specification: adds 32 points.
- Narrow exclusivity by field, use or time, or keep a second buyer, so the concentration reserve stays at the base deal's 2 percent: saves 17.5 points.
- Cap refreshes, support and audit cooperation at the 2x level, or price them as a separate service: saves 12 points.
- Pay on milestones tied to accepted tranches, cutting months to cash from 8 to 4: saves about 10.1 points.
Where the scale model could overstate the problem
The model is built to expose a mechanism, and four facts could make it overstate the problem for a particular company. Each is worth testing before concluding that a large deal is a bad one.
- Economies of scale could be larger. If evidence work, contracting and quality assurance are a bigger share of cost than assumed here, spreading them favors large deals more than the model shows.
- Large buyers may pay for what they ask. A buyer that funds its own specification, pays on milestones and accepts non-exclusive terms removes three of the four drags.
- An anchor buyer can justify a thin margin. A large first deal can validate a dataset type and open later licenses; the model scores one deal, not the option value of what follows.
- Public evidence is thin. The concentration figures come from three public companies with different businesses, and no public source reports bespoke hours or service costs by deal size, so the exponent and the reserve rates are assumptions to replace with a company's own data.
Diligence questions on deal size, SourceX included
Ask for the loaded contribution of the largest deals beside the median deal, with the four faster-growing lines shown separately. The questions apply to any company that licenses or brokers datasets.
At SourceX, the levers above are negotiated deal by deal: each supplying company must approve scope and terms and sign before it delivers any data, and whether a license is exclusive, and for how long, is written into that deal's own signed agreement. SourceX has not published deal sizes, prices or margins; company and commercial metrics are shared with investors directly through the investor materials request.
Questions to ask:
- What share of revenue and of specialist hours did your largest deal take last year, and what was its loaded contribution?
- How do bespoke preparation hours grow with deal scope, and are they billed separately?
- How many months pass between first spend and cash on your largest deals compared with your median deal?
- Which post-delivery obligations do your large deals carry, and are they capped or priced?
- What concession, renegotiation or non-renewal assumption sits in your plan for the largest buyer?
- Do loaded costs across all deals reconcile to total hours and cash spent in the period?
Methods note: how the scale model was built
All model values are illustrative index points chosen to show the mechanism, with the base deal's headline price set to 100. They are not SourceX figures, do not describe any real transaction and are not forecasts. The rules are those stated in the model section: volume discounts of 5, 12.5 and 20 percent, records and standard preparation at 55 points per unit of scope, bespoke preparation at 4 points times scope to the power 1.5, service obligations stepped by tier, a 1 percent monthly cost of capital on direct spend assumed to occur at the start, and concentration reserves of 2 to 11 percent of headline price.
The public evidence (Innodata's 2025 annual report, Appen's February 2024 market update, Reddit's fiscal 2025 annual report and March 2026 quarterly report, the GSM1k paper, the FTC's 2021 Everalbum order, California's AB 2013 and the European Commission's July 2025 template) was checked on October 26, 2026 and supports the mechanisms, not the model's values. Not measured: bespoke hours, cycle times, service costs or concession rates at any real company, taxes, revenue recognition rules and the option value of follow-on deals.
Frequently asked questions
Are larger data licensing deals more profitable?
Not automatically. Price grows roughly with scope, minus volume discounts, while bespoke preparation, delay, buyer concentration and post-delivery obligations can grow faster. Judge a large deal on loaded contribution with those four costs shown separately, not on the reported deal margin.
What is a concentration reserve in a data deal model?
It is an expected-cost line for depending on one buyer: the probability of a cancellation, renegotiation or partial acceptance times the value at risk. It rises as the buyer's share of the business grows, and setting it to zero quietly assumes the renewal.
How should bespoke preparation for a large buyer be priced?
Separately, against a written specification, as change orders or a preparation line the buyer pays for whether or not it renews. Folding it into a per-record price hides work the next buyer will not pay for, especially when the deal is exclusive.
Can a large anchor buyer justify a thin margin?
Sometimes. A first large deal can validate a dataset type and lead to later licenses, which a one-deal model does not capture. The test is whether the company can name the follow-on value it expects and the date by which it should appear.
Does SourceX publish its deal sizes or margins?
No. SourceX has not published deal sizes, prices or margins, and the scale model in this article is illustrative. Company and commercial metrics are shared with investors directly through the investor materials request at sourcex.si/investors.
Sources
- Reddit's Form 10-K for fiscal 2025 states that substantially all of the contract value associated with its licensing revenue is derived from two of its partners. Source
- Reddit's quarterly report for the period ended March 31, 2026 put remaining performance obligations on contracts longer than one year at $120.6 million, mostly long-term content licensing contracts. Source
- Appen's February 12, 2024 update said Google's terminated contract, with projects ending by March 19, 2024, contributed $82.8 million of FY23 revenue at a 26 percent gross margin. Source
- Innodata's 2025 Form 10-K says one customer in its DDS segment generated approximately 58 percent of total revenues in 2025 and 48 percent in 2024. Source
- The GSM1k paper reports accuracy drops of up to 8 percent for leading models on a fresh benchmark mirroring GSM8k, and the authors withheld GSM1k to prevent contamination. Source
- In May 2021 the FTC finalized an order requiring Everalbum to delete models and algorithms developed using biometric information from users' photos and videos. Source
- California AB 2013 requires generative AI developers to post training data documentation on or before January 1, 2026, including dataset sources or owners and whether datasets were purchased or licensed. Source
- The European Commission's July 24, 2025 template asks general-purpose AI providers to summarize training content, including private datasets, updated at six-month intervals when further training requires it. Source
Request SourceX investor materials
SourceX shares company and commercial metrics with investors directly. Any offering is made only to verified accredited investors.