Licensed, collected or synthetic training data

There are three ways to source training data. You license material that already exists, you commission a vendor to collect or label new data to your specification, or you generate synthetic data from a model or simulator. Licensed data is fastest, collected data fits your task most closely, and synthetic data scales most cheaply once you can verify it.

Reviewed 7 November 2026.

How do the three routes compare?

Typical traits of each route. Individual deals vary.
Question	Licensed	Collected or labeled to order	Synthetic
What you get	Existing material from a rightsholder: archives, publications, recordings, logs	New data made to your specification: tasks, recordings, labels, expert answers	Data produced by a model, simulator or rules engine from seed data or a spec
Time to first delivery	Fast once terms are agreed; the data already exists	Slowest: recruiting, guidelines, pilots and quality review come first	Fast to generate; validation takes the time
What drives the price	Size, uniqueness, exclusivity and how long you may use it	Hours of skilled work, task difficulty, review rounds	Software seats or compute, plus the seed data
Rights you usually get	A license for a term, with limits on use and sharing	Often full ownership, if the contract says so	Depends on the generator's terms and the seed data's license
Fits best when	You need breadth or real-world variety that would take years to collect	Nothing suitable exists, or you need a specific behavior taught	Real data is scarce, sensitive or expensive, and you can check outputs
Main risk	Gaps in the chain of title; terms that end before the model does	Cost overruns and inconsistent quality across annotators	Errors and blind spots copied from the generator; drift from reality

When should you license existing data?

License when the value is in breadth: years of forum posts, a publisher's back catalog, a library of stock footage, a company's support archive. Nobody can commission that variety quickly. Since 2024 the largest licensing deals have been between platforms or publishers and AI labs; the deals post collects the figures that were made public.

The work is in the rights. Ask who created the material, whether the seller may sublicense it for training, and what happens to your model when the license ends. The provenance checklist lists the questions.

Where to look: the licensed data and marketplaces sections of the directory.

When should you commission collected or labeled data?

Commission when you need something that does not exist yet: answers written by domain experts, preference rankings for a particular model, speech in an accent no catalog covers, or labels on your own images. You set the guidelines, so the data matches your task, and the contract can give you ownership.

Budget for a pilot. A paid pilot of a few hundred items shows how a vendor's annotators read your guidelines before you commit to the full volume. Most vendors in the annotation and expert and RLHF sections price this kind of work by quote, though a few publish per-object or per-hour rates.

When does synthetic data make sense?

Synthetic data earns its place when real data is scarce, sensitive or slow to label: rare defects in manufacturing images, edge cases for driving perception, tabular records that cannot leave a secure environment, or instruction data generated by a stronger model. It is weakest when the generator shares the blind spots you are trying to fix.

Keep a real, held-out evaluation set. It is the only reliable check that synthetic training data improves the model rather than teaching it the generator's habits. Vendors are listed under synthetic data.

How do you decide between them?

Write down the behavior the model must learn and how you will measure it.
Check whether licensable data already covers it. If it does, price that first.
For the gaps, decide whether people must create the data (collected) or whether a generator you can check will do (synthetic).
Price each piece with the same unit: cost per usable example after review.

Questions about sourcing routes

Which route is cheapest?

Per item, synthetic data usually costs least once a pipeline exists, and collected expert data costs most. Per useful item the order can flip: a cheap set that needs heavy cleaning or fails evaluation is not cheap. The cost estimator works through published unit prices.

Can I combine the routes?

Most production training sets do. A common pattern is a licensed or collected core, expert-labeled examples for the hard cases, and synthetic data to fill rare classes, with a held-out real set for evaluation.

Is annotation of my own data a fourth route?

It is collection applied to data you already hold. You supply the raw material and the vendor adds labels, rankings or corrections. Rights questions are simpler because you own the inputs, but contributor terms still matter.

Does licensed data come with exclusive rights?

Rarely by default. Most licenses are non-exclusive. Exclusivity is negotiated and priced separately, often at a multiple of a standard license.

How do I know a synthetic set is good enough?

Test a model trained on it against real held-out data you trust. Distribution checks on the synthetic data alone do not tell you how a model will behave.