Licensing your video archive to AI companies: how it works and what affects the price

If you own a library of video, audio or text, AI companies may want to license it for training. That can turn footage sitting on hard drives into a new revenue line without selling the archive itself. This guide explains how archive licensing for AI works, what buyers look for, and the factors that push the price of a deal up or down.

What licensing to an AI company actually means

In a typical archive deal, you grant an AI company the right to use copies of your content to train and evaluate machine learning models. You keep ownership of the content and the copyright. The buyer gets a license, usually non-exclusive, defined by a contract that sets out what they can do with the material, for how long, and in which territories.

This is different from a traditional stock or broadcast license. The buyer is not going to publish your clips. They want the material as training signal: how people move, how scenes change, how speech sounds in a particular accent or room. That shifts what matters. A beautifully graded hero shot may be worth less than hours of ordinary, varied, well-labelled footage.

Who buys archive content, and why

Buyers range from large model developers to small robotics and computer vision teams. Common reasons include:

• Video models need large volumes of varied, real-world footage covering motion, lighting, camera movement and scene types.
• Robotics and embodied AI teams look for footage of people doing physical tasks, especially from first-person or close-up angles. See our note on egocentric video for why.
• Speech and audio models need recordings across languages, accents, ages and acoustic environments.
• Language models value well-edited text in specific domains or under-represented languages.

Buyers also care more than ever about provenance. In the EU, the AI Act requires providers of general-purpose AI models to publish a summary of the content used for training, with those obligations applying from August 2025. Licensed content with a clean paper trail is easier for a buyer to account for than material of unclear origin.

What affects the price

There is no public rate card for AI archive licensing, and deals vary widely. Rather than quote figures that would not apply to your library, here are the factors that most often move the number.

Rights clarity

This comes first. Can you show you own or control the rights to license the content for AI training? Commissioned work, contributor agreements, music, third-party clips and talent releases all need checking. Content where the chain of title is clean is worth more, and content where it is unclear may not be licensable at all.

People in the footage

Footage of identifiable people raises privacy questions under laws such as the GDPR in the EU and Indonesia's UU PDP. Releases that cover broad future use help. Without them, buyers may require face blurring, may restrict use, or may pass. Material involving minors is usually excluded. Our note on consent for AI training data covers what buyers check.

Volume and diversity

Buyers typically price by duration, file count or item count. Larger libraries attract more interest, but diversity matters as much as size: many locations, subjects, seasons, activities and recording conditions are more useful than thousands of hours of near-identical material.

Rarity

Content that is hard to find elsewhere carries a premium. That includes footage from regions under-represented in existing datasets, uncommon activities, specialist domains, and audio in low-resource languages. Our note on geographic bias explains why buyers are actively looking for data from outside North America and Europe.

Technical quality

Resolution, frame rate, stable audio, minimal compression artifacts and the absence of burned-in text, watermarks and logos all help. Raw or lightly edited footage is often more useful than heavily edited packages.

Metadata

Descriptions, timestamps, locations, camera details, transcripts and shot-level tags make content far easier to use. Good metadata can lift the value of an otherwise ordinary library.

License terms

Exclusivity, duration, territory, permitted model types and whether the buyer may keep training on the data after the term ends all change the price. A non-exclusive license is the common starting point and lets you license the same material to several buyers.

How the process usually works

1. Inventory. List what you have: formats, total duration or item count, subjects, locations, years and any existing metadata.
2. Rights review. Identify what you can license cleanly and what needs releases, blurring or exclusion.
3. Sample. Prepare a representative sample, usually a small slice with metadata, so buyers can assess fit without taking the whole library.
4. Buyer matching and terms. Approach buyers whose needs fit your content, then negotiate scope, exclusivity, duration and price.
5. Preparation and delivery. Clean, blur where needed, standardise formats, attach metadata and a manifest, and deliver through a secure transfer.
6. Reporting and renewal. Keep records of what was delivered to whom under which license, so you can renew, extend or license again.

Protecting yourself

• Keep ownership. A license should not transfer copyright unless you deliberately choose to sell.
• Read the use clause closely. Know whether the buyer can use your content for any model or only specific ones, and whether they can share it with affiliates.
• Avoid broad exclusivity unless the price justifies it.
• Agree what happens at the end of the term, including whether trained models may continue to be used.
• Ask for a clause preventing the buyer from republishing or redistributing your raw content.
• Get your own legal advice before signing.

Masana helps archive owners through this process, from inventory and rights review to buyer matching and delivery, while the owner keeps ownership of the library.