Leadership and readiness
Hybrid data licensing business strategy: the rights-to-product test
By SourceX Editorial · Reviewed by Noah Loul ·
9 sections
Short answer
A data company can profit from both licensing and synthetic work only where three things line up: each product has its own grant of rights, the synthetic line raises demand for licensed records instead of replacing them, and evaluation stays independent of generation. Shutterstock and Getty Images show the two lines can coexist. Test product by product with a rights-to-product grid and a matched-budget experiment before crediting the combination.
License terms
Hybrid data licensing
Key takeaways
- A training license does not cover generating from records, passing generated output on or running a generator service; each product needs its own grant.
- Shutterstock and Getty Images sell generated output beside licensing, and both set up contributor compensation for training use rather than relying on the original grants.
- Synthetic data replaces bulk demonstrations more easily than outcome-labeled records, scarce judgments and sealed test sets.
- Supplier, buyer and evaluation conflicts each carry a cost that belongs in the model of a dual-line business.
- A matched-budget test on untouched real work yields three readings an investor can use: exposure, dependence and displacement.
Can one data company profit from licensing and synthetic work at the same time?
It can, but only product by product, and only where three conditions hold: each product rests on its own grant of rights from the people whose records feed it, the synthetic line raises demand for licensed records instead of replacing them, and evaluation work stays independent of generation work. Where any condition fails, the second line cannibalizes the first or exposes it to a rights dispute.
That is narrower than the usual hybrid pitch, which treats a library of licensed records as a platform that permits anything built on top of it. A license is a list of permitted uses. Permission to let buyers train models on records says nothing, by itself, about generating synthetic records from them, reselling the output or offering a generator as a service. A hybrid data licensing business strategy should be judged on the grants it actually holds, one product at a time.
The thesis this analysis tests: a combined licensing and synthetic business is worth more than its parts only when the synthetic line runs on grants the licensing line already secured, and sells checks or seeds that buyers cannot get from generation alone. The evidence comes from two content licensors that built generators, a published model report, a court judgment and an experiment design. No SourceX results are reported.
Two licensors that built generators: what the public record shows
Shutterstock and Getty Images show that a licensor can sell training access and generated output side by side, and that doing so took new contributor terms rather than a reading of the old ones.
Shutterstock's Form 10-K for 2024 reports Data, Distribution and Services revenue of $175.3 million, up 28 percent, and describes the offering as including access to its metadata for machine learning and generative AI model training. In January 2023 the company announced an AI image generation platform for all its customers. Its Contributor Fund, launched in October 2022, compensates artists for the role their content played in training its generative technology, and the company says artists also receive ongoing royalties tied to licensing of newly generated assets. In July 2023 it added indemnification for enterprise customers using AI images created and licensed on its site.
Getty Images launched its generative AI tool in September 2023, saying it was trained solely on its creative library, carried full indemnification for commercial use and would compensate contributors whose content was included in training. Its current model card describes a November 2025 release built on a Bria model and trained on images from Getty Images and Bria's other creative partners, and says different monetary levels of indemnification are offered based on the package purchased.
The lesson for an investor is about rights, not pictures. Both companies treated the generator as a new product that needed its own contributor arrangements, and Getty's generator changed its training basis and indemnity structure within about two years. The comparison below rates only what the cited documents state:
| Criterion | Shutterstock | Getty Images |
|---|---|---|
| Sells library access for AI model training | Yes: the 2024 10-K describes metadata access for machine learning and generative AI training | Not established in the sources cited here |
| Sells generated output to customers | Yes: AI image generation platform announced January 2023 | Yes: generative AI tool launched September 2023 |
| Pays contributors for training use | Contributor Fund since October 2022, plus royalties tied to licensing of generated assets | Said at launch that contributors would be compensated for inclusion in training |
| Indemnifies generated output | Enterprise indemnification added July 2023 | Full indemnification at launch; the current model card ties the level to the package bought |
| Training basis stayed the same | Not assessed in the sources cited here | No: Getty's library only at launch; the November 2025 release adds Bria's other creative partners |
The rights-to-product grid: which grant does each product need?
Each product a dual-line company sells needs a specific permitted use from the companies or creators behind the records, and the uses do not imply one another. The grid lists six products, the grants each requires, how each relates to the licensed line and the conflict an investor should price.
Read the relation column first. Rows marked complement add revenue without shrinking demand for licensed records. Rows marked substitute risk can replace a licensed sale with a cheaper synthetic one, and a company whose growth story rests on them needs the strongest grants and the clearest supplier economics, because it is selling the product most likely to make its own supply less necessary.
The grid, with the evidence to request for each row:
| Product | Uses the supplier agreement must grant | Relation to the licensed line | Conflict to price | Evidence to request |
|---|---|---|---|---|
| Licensed training set | Train, sublicensed to named buyers | Core line | Exclusivity promised to one buyer limits every other row | Permitted-use clause, exclusivity terms, buyer count per dataset |
| Sealed test set | Test only; training and generation barred | Complement | Loses its value if the same records seed generation | Access log and a written bar on seeding |
| Seed pack for a buyer's own generator | Generate from, with outputs kept by the buyer | Complement | The buyer's outputs can later replace repeat purchases | Clause on downstream outputs and resale |
| Synthetic set made in-house | Generate from, and pass generated records on | Substitute risk | Competes with the licensed set it came from | Generation grant and the supplier share of synthetic revenue |
| Generator as a service | Train, generate from, and pass outputs on | Substitute risk | Hardest grant to obtain and the strongest pull away from licensing | Model training grant, output terms, indemnity limits |
| Verification against real records | Test, with results reported to third parties | Complement | Independence fails if the verifier also sells the generator | Separation of teams and data between the lines |
Where synthetic data substitutes for licensed records, and where it complements them
Synthetic data replaces bulk demonstrations most easily and complements the scarce judgments that tell a model which outputs are good. The clearest public example puts both in one training run.
NVIDIA's Nemotron-4 340B technical report (June 2024) says over 98 percent of the data used in its model alignment process was synthetically generated. The same report trained its reward model on 10,000 human-annotated preference examples from HelpSteer2. Generation supplied the bulk; a small set of human judgments stayed as the reference that decided what the bulk was worth.
For business records the same split marks where each line earns money. Routine demonstrations, such as well-formed ticket replies or standard invoice handling, are the rows most exposed to substitution. Records that carry an outcome, a disputed judgment or a rare exception, and sealed sets that score models against real work, are rows a generator needs rather than replaces. A dual-line company profits when its synthetic line consumes the second kind under a proper grant and sells the first kind cheaply, without undercutting a licensed sale of the same records.
The strongest alternative deserves its full weight: synthetic methods can substitute for some licensed data outright. If a buyer can reach its target on real work with public seeds and a capable generator, the licensed line loses that sale whether or not the same company also sells synthetic data. The plan two sections down is designed to show which case a given task family is in.
Three conflicts a company selling both lines has to price
The conflicts run to suppliers, to buyers and to the company's own evaluation work, and each has a cost that belongs in a model of the business.
Supplier conflict: a generator built on suppliers' records can replace the next purchase of those records. Shutterstock's answer was payment tied to generated output, with royalties on licensing of newly generated assets, which keeps contributors paid when the synthetic line sells. A company with no such mechanism is asking suppliers to fund their own substitute, and the bill tends to arrive as refused renewals or narrower grants.
Buyer conflict: a synthetic set derived from records licensed exclusively to one buyer can undercut that exclusivity, and a cheap synthetic product can reset the price buyers expect for the original. Exclusivity terms have to name derived outputs, or the company ends up selling the same value twice and inviting a dispute.
Evaluation conflict: a company that generates training data and also holds test sets has an incentive problem and a contamination problem. Records used as seeds must never reach a sealed evaluation set, and a verification product loses its value the moment buyers suspect the verifier grades its own generator kindly. Separate teams, separate data and access logs are the evidence to ask for.
A matched-budget plan that shows whether the two lines add up
One matched-budget experiment per task family shows whether the synthetic line cannibalizes the licensed line, depends on it or adds to it. Fix the training budget, vary only where the examples come from, and score every arm on real work none of them has touched.
At a glance
A matched-budget plan that shows whether the two lines add up
The plan produces three readings, each tied to a business decision: exposure (does the licensed line survive public-seed generation?), dependence (does the synthetic line need the licensed records?) and displacement (how many licensed records a hybrid can drop). The steps:
- Choose one task family with a measurable outcome, such as routing support escalations or catching invoice exceptions, and fix the number of training examples and the compute for every arm.
- Freeze the holdout first: real records from a supplier never used for seeds or training, held by a team with no access to the generation pipeline.
- Arm A, licensed only: real records used under a training grant.
- Arm B, public-seed synthetic: generated examples seeded only from public or self-written material.
- Arm C, licensed-seed synthetic: generated examples seeded from the same licensed records, under a generation grant.
- Displacement ladder: hybrids that keep 75, 50 and 25 percent of arm A's licensed records and fill the rest with arm C's examples.
- Read exposure as B against A, dependence as C against B, and displacement as the smallest licensed share on the ladder that still matches A on the holdout.
- Report scores with confidence intervals, the holdout's size and selection rule, and the date each arm was frozen; readings from one task family do not transfer to another.
What could make the dual-line thesis wrong
The thesis fails if generation stops needing licensed seeds, if rights leverage over generators turns out weak, or if suppliers refuse the grants the synthetic rows need. Each risk already has a public marker.
None of these risks argues against licensing as such; they argue against paying a premium for the combination before the grants and the experiment are in hand. The specific ways the thesis could break:
- Generation outgrows its seeds. Nemotron-4 340B's alignment data was over 98 percent synthetic; if business tasks follow, arm B will match arm A for more task families and the licensed line shrinks whatever the strategy.
- Leverage over outside generators proves narrow. In Getty Images v Stability AI ([2025] EWHC 2863 (Ch), handed down 4 November 2025), it was accepted that the model's training did not take place in the United Kingdom, so the training-based claim there fell away. Licensing premiums that lean on litigation risk can thin out.
- Suppliers decline generation grants or demand shares that erase the synthetic margin; royalties on generated assets, as in Shutterstock's model, are a cost a pure licensor does not carry.
- Rights bases move. Getty's generator changed training sources and indemnity tiers between its 2023 launch and its November 2025 release, and agreements written for one product version may not cover the next.
- Seeds leak into holdouts. If they do, every reported hybrid gain is overstated and the company's own experiment cannot be trusted.
Questions for a company that sells both, SourceX included
Ask for documents rather than descriptions: the clause behind each product, the money that flows back to suppliers when derived products sell, and the experiment that shows how the lines interact.
Applied to SourceX, the grid's questions land on its licensing agreements. Its public description covers sourcing and qualification, rights review, contracting, data preparation, delivery and payment for companies licensing operational data to AI developers; it does not describe a synthetic data product. Nothing is delivered until the supplying company has approved scope and terms and signed; afterward SourceX may use, sublicense and distribute the de-identified dataset only within what that agreement allows. Which grid rows an agreement covers is therefore a question the agreement itself answers.
SourceX does not publish deal, revenue or supplier figures. It shares company and commercial information with investors directly, and it is raising from accredited investors under Rule 506(c), with investor materials available on request at sourcex.si/investors. Questions to put to any company selling both lines:
- For each product line, which permitted use in the supplier agreement covers it? Show the clause, not a summary.
- When a synthetic product sells, what do the suppliers whose records seeded it receive, and on what basis?
- Do exclusivity terms name derived and synthetic outputs?
- Who holds the evaluation sets, and what stops their records from becoming seeds?
- For which task families has a matched-budget test run on untouched real work, and what were the exposure, dependence and displacement readings, with denominators?
- Which grid rows earn revenue today, and which are plans?
Methods note: how the grid and plan were built, and what they leave out
The grid's six products and four grants (train, test, generate from, pass outputs on) are an editorial framework built on 27 October 2026 from the sources cited: Shutterstock's 2024 Form 10-K and three of its press releases, Getty Images' 2023 launch announcement and current model card, NVIDIA's Nemotron-4 340B technical report and the High Court judgment in Getty Images v Stability AI. The relation labels (core, complement, substitute risk) are judgments from the mechanism, not measurements.
Not measured: product-level revenue or margins at either comparable company, contract terms beyond what the companies published, how much each company's generator affects its licensing sales, and any SourceX deal, revenue or supplier figure. The comparison table rates only what the cited documents state, and a gap in those documents is reported as not established rather than as a no. The experiment is a design; no results are reported, and its readings would apply only to the task family tested.
Frequently asked questions
Can a data licensing company also sell synthetic data?
Yes, if its supplier agreements grant the uses each synthetic product needs, typically generating from the records and passing generated output on. Permission for buyers to train on the records does not imply those uses. Shutterstock and Getty Images both sell generated output alongside licensing and set up contributor compensation for training use.
Does a license to train on records cover generating synthetic data from them?
Not unless it says so. Training, testing, generating from records and passing outputs to others are separate permitted uses, and agreements often grant some without the others. Read the permitted-use clause for each product; this is general information, and counsel should review any specific agreement.
When does a synthetic line cannibalize a licensing line?
When buyers can reach the same score on real work with synthetic examples that do not need the licensed records, or with fewer licensed records than before. A matched-budget experiment with an untouched real holdout measures this per task family, and the licensed records a hybrid can drop while matching the licensed-only arm are the displaced volume.
How do Getty Images and Shutterstock pay contributors for generative AI?
Shutterstock says its Contributor Fund, launched in October 2022, compensates artists for the role their content played in training and pays royalties tied to licensing of newly generated assets. Getty Images said at its 2023 launch that contributors would be compensated for inclusion of their content in training. Both descriptions are the companies' own.
What should an investor ask a company that sells licensed and synthetic data?
Ask which clause covers each product, what suppliers receive when synthetic products sell, whether exclusivity names derived outputs, how evaluation sets are kept away from generation, and which task families have matched-budget results on untouched real work. Ask for the documents rather than summaries.
Does SourceX sell synthetic data?
SourceX's public description covers licensing operations, from sourcing and rights review to contracting, preparation, delivery and payment; it does not describe a synthetic data product. Investors evaluating SourceX can request materials at sourcex.si/investors, where company and commercial information is shared directly.
Sources
- Shutterstock's Form 10-K for 2024 reports Data, Distribution and Services revenue of $175.3 million, up 28 percent, and describes the offering as including access to metadata for machine learning and generative AI model training. Source
- Shutterstock announced an AI image generation platform available to all its customers in every language its site offers. Source
- Shutterstock says its Contributor Fund, launched in October 2022, compensates artists for the role their content played in training its generative technology, with ongoing royalties tied to licensing of newly generated assets. Source
- Shutterstock offered enterprise customers indemnification against potential claims related to their use of generative AI images created and licensed on shutterstock.com. Source
- Getty Images launched a generative AI offering in September 2023, trained solely on its creative library, with full indemnification for commercial use and compensation for contributors whose content was used in training. Source
- Getty Images' model card describes a November 2025 release built on a Bria model, trained on images from Getty Images and Bria's other creative partners, with indemnification levels that depend on the package purchased. Source
- NVIDIA's Nemotron-4 340B technical report says over 98 percent of the data used in its model alignment process was synthetically generated, and that its reward model was trained on 10K human-annotated HelpSteer2 preference data. Source
- In Getty Images v Stability AI ([2025] EWHC 2863 (Ch), 4 November 2025), it was accepted that the model's training and development did not take place in the United Kingdom, so the training-based claim fell away. Source
Request SourceX investor materials
SourceX shares company and commercial metrics with investors directly. Any offering is made only to verified accredited investors.