AI companies have run out of easy data. The open web has largely been scraped, publishers are blocking crawlers, and the copyright lawsuits have made "we found it online" a risky answer. So the biggest labs, and now robotics companies too, are paying for data they can legally defend.
That has created a new market, and a lot of noise around it. Social media is full of claims that a single "data stack" sells for six or seven figures. Some deals really are that large. Most are not. This guide collects every public price point we could verify as of October 10, 2026, so you can tell the difference. Most deals are private, so treat these as reported figures, not a rate card.
Key Takeaway
Three things set the price of a dataset: how much there is, how hard it is to get anywhere else, and how cleanly you can prove you own it. Miss any one and the number drops fast.
The headline deals: large archives
The largest reported deals all involve institutions licensing archives they already owned.
- Reddit and Google: about US$60 million a year, according to Reuters' sources. Neither company confirmed the figure.
- News Corp and OpenAI: reported to be worth about US$250 million over five years.
- Shutterstock: its first deals with Apple, Meta, Google, and Amazon were reportedly worth US$25 million to $50 million each, and have since been expanded, according to Reuters.
- Wiley: told investors it expected US$44 million from AI licensing deals.
- Taylor & Francis and Microsoft: more than US$10 million for initial access, plus recurring payments.
Source: Search Engine Land, Reddit and Google licensing deal
Source: VentureBeat, News Corp and OpenAI
Source: The Decoder, the AI training data shopping spree
Source: The Bookseller, Wiley AI rights deals
Source: The Bookseller, Taylor & Francis and Microsoft
These are not repeatable for most businesses. They reflect decades of unique content at enormous scale. They also wobble: Shutterstock reported its data licensing revenue fell sharply in the first quarter of 2026, which the company attributed partly to deal timing, a reminder that one-off training deals are lumpy.
Source: Shutterstock, Q1 2026 results
Unit prices: video, images, books, and text
For everyone else, unit prices are more useful. These are the figures reported in the press:
- Video: about US$1 to $4 per minute for unpublished footage, according to Bloomberg. Ordinary footage sits at the bottom of the range, while 4K, drone, and 3D footage earns more. Adobe was separately reported to pay about US$3 per minute.
- Images: Freepik signed deals at 2 to 4 cents per image. A data marketplace quoted US$1 to $2 per image for higher-value sets.
- Film: the same marketplace quoted US$100 to $300 per hour.
- Text: around US$0.001 per word for general text.
- Books: Microsoft paid US$5,000 per title for a three-year licence in its deal with HarperCollins, split evenly between the publisher and authors who opted in.
Source: The Decoder (citing Bloomberg), AI companies buying unpublished video
Source: Bloomberg, Adobe buying video at about $3 per minute
Source: The Decoder, HarperCollins and Microsoft book licensing
One more benchmark is worth knowing, though it is not a market price. In 2026 a US court gave final approval to Anthropic's US$1.5 billion settlement with authors, which works out to about US$3,000 per book. It set a public reference point for what a single work can be worth when rights were not respected.
Source: JURIST, Anthropic settlement approved
The fastest-growing segment: expert data and RL environments
The biggest money in AI data today is not in archives at all. It is in data made to order: experts writing and grading answers, and "RL environments", simulated versions of software or tasks that models practise in. Mercor reported more than US$2 billion in annualized gross revenue by mid-2026, and research group Epoch AI reports that environment contracts are often six to seven figures per quarter, with exclusive deals costing roughly four to five times more than non-exclusive ones.
Source: Epoch AI, the state of RL environments
Source: Startup Fortune, Mercor revenue
This work is engineering-heavy and run by well-funded vendors. It matters here because it shows where demand is heading: toward specialist, hard-to-make data, and away from generic bulk content.
So is "six figures per data stack" real?
Sometimes. We could not trace the phrase to any verifiable deal. Here is the honest version.
- It is realistic for large, unique, rights-cleared archives, for exclusive deals, and for engineered expert data.
- It is unrealistic for a few hundred hours of ordinary footage. Bloomberg reported individual creators earning deals in the thousands of dollars.
- Do the math: at US$1 to $4 per minute, US$100,000 of video means roughly 400 to 1,700 hours of footage a buyer actually wants.
Volume thresholds are real. One video licensing startup that pools creator footage said it needed at least 25,000 hours of content in its catalogue before AI companies would buy in. Small datasets usually sell best pooled with others rather than pitched alone.
Source: Tubefilter, video licensing volume minimums
What pushes your price up
- Uniqueness: data that is not already online, such as unpublished footage, internal manuals, or specialist recordings.
- Volume: buyers think in hundreds or thousands of hours, and tens of thousands of images.
- Clean rights: a clear chain of ownership and documented consent from anyone who appears in the data.
- Documentation: a data card that explains source, lineage, licence terms, and known gaps.
- Exclusivity: worth a large premium, but it locks you out of other buyers, so price it carefully.
- Specialism: bilingual, technical, or skilled-work data beats general content.
What pushes it down
Scraped content, missing consent, personal information you cannot remove, and anything a buyer could get elsewhere for free. Since the major copyright cases, provenance is a large part of what buyers are paying for. If you cannot prove where data came from, assume it is unsellable.
Key Takeaway
Before you spend anything packaging a dataset, get a rough valuation against these public benchmarks. Plenty of archives are worth selling. Plenty are not, and it is better to find out in a week than in six months.
If you run a Canadian business and think your archive might be worth something, our AI data licensing service starts with exactly that valuation, and our free data value calculator gives you a rough number in a minute. For the legal and practical side, read how to sell your data to AI companies from Canada, and if you work in the trades or robotics, see our guide to egocentric data for robotics.
Wondering what your data is worth? Book a free call. Tell us what you have, and we will give you an honest read on whether it is worth taking further.
Get Started