Litigating and Monetizing Content Licensing to LLMs
Licensing data to LLMs is a potential revenue stream for streamers much like advertising on CTV platforms: it is an opportunity that didn’t exist until recently that has the potential to deliver dividends for years to come. But as with CTV advertising, its viability and profitability won’t happen overnight.
For now, there are numerous rights lawsuits in the works, and most companies have not come around to respecting content copyright.
Image source: ChatGPT Is Eating the World
One of the companies that media analyst Laura Martin tracks in her work as a Managing Director at Needham & Company is OTT subscription streaming service Curiosity Stream (CURI). CURI recently reported Q3 2025 revenue of $18.36 million, up 46% year-over-year. “What we like most about CURIs 3Q25 was its 425% LLM licensing revenue growth,” Martin says. “Today, CURI is licensing 2 million hours of premium TV, film, and video content to 19 LLMs. In 2026, CURI believes it can double its content supply and grow its LLM licensees by ~50% to ~30 LLMs.”
Of the 8 deals they have had in place for at least a year, CURI expects revenue from these licensing deals to double in 2026. The company owns ~300,000 hours of their own content and does a revenue share for 50% of any LLM fees they get on the other 1.7 million hours of content distributed through their platform.
“This has only happened a couple of times in my career in different companies when you achieve enough scale in one area, that a derivative service line can become potentially the main service line,” says Veritone CEO Ryan Steelberg.
Veritone started in 2012 and is now processing 250,000 hours of unstructured data and content per day. In Q3 2024 they entered into LLM licensing, facilitating deals for backend model training. “We represent a very large number of audio and video IP owners. The majority of the deals that we’ve done to date are datasets that I had no relationship with before.”
Content at Scale
In terms of training mediums, text has a much clearer provenance. “The through line is very clear,” says Brian McNeill, COO/CPO and Co-Founder, Stringr. Stringr has been in business for 12 years and has a videographer marketplace of 160,000 videographers. They have 2.5 million video assets where you can buy stock or have videographers go out and shoot specific content.
In addition to finished content, there is a large ratio of outs to good shots. “All of these news agencies have huge archives of images and video that are sitting there,” he says, “There is a large amount that either goes to archives or is even deleted.”
Right now, 100,000+ minutes seems to be the minimum for training. McNeill says library content “is only valuable when you’re talking about the 2 million video assets we have or the tens of millions of assets that sit over at Reuters or AP. That's when the conversation starts to get interesting for the model company, because there’s enough there to actually train on," says McNeill.
Do they use existing metadata? No. “Most of the models are using their own image detection algorithms to identify what is in there, as opposed to using the metadata that was provided by the by the contributor.”
News organizations have been approached by deal aggregators with a wide variety of terms. Currently one consultant in the field is recommending:
- $20 to $30 per hour (net to stations) of licensed content
- prohibitions against replication or generation of facsimiles of specific human beings or their voices, or anything we would consider a “derivative work”
- ability to withhold from specific buyers
- no grant of copyright or any other exhibition rights, replication rights, etc.
- one-year term with right to pull out upon notice
“What these firms are looking for seems to shift every three to six months based on the next model they’re training. For a while it was, we need faces straight on and then it was, we need shots taken in a particular style so we can replicate that style, whether it’s various camera angles, lighting techniques,” says McNeill.
Depth
There are different issues with using media content. Converting audio and video into tokens—the distillation of content into a linguistical unit, which is mapped to a numeric value and used to train the next generation of models both has more complicated rights clearances as well as a much deeper level of data for model training.
“It [was] a lot easier to tokenize text, which in effect is on a per-word basis, to train an LLM than it is to convert audio and video into tokens and then use that as training data,” says Veritone’s Steelberg. “Now we’re entering into the very messy world of all these other forms of unstructured data dominated by audio and video."
Steelberg notes that each frame of video could contain “as many as 2,000 tokens. How do you describe centricity of a face on screen, or movement as that head turns into shadows? You’re looking at the capability to generate description on anything and everything inside a frame: shadowing, lighting, movement, interaction between two objects, what they’re saying, etc.”
Veritone has seen the most demand for licensing reality content. Sports is another key area. Veritone is now representing the BCAA to turn their audio and video into training data. They have found that gaining clearances for sports are more straightforward than for scripted shows. “Scripted content may not necessarily represent a true representation of the real world,” says Steelberg.
Training Costs
ChatGPT sourced some data on True Geometry on training costs for the hyperscalers:
- Minimum: $70 million (a smaller, less capable model)
- Typical: $100 million-$200 million+ (for a model with comparable capabilities)
- State-of-the-art (e.g., GPT-4, Gemini): $200 million-$1 billion+ (and potentially much higher).
A 2025 analysis of model-training cost growth (“The rising costs of training frontier AI models”) shows: “The amortized cost to train the most compute-intensive models has grown at ~2.4×/year since 2016 … If the trend continues, the largest training runs will cost more than a billion dollars by 2027.”
How much will media be worth in the world of LLM training? “The hyperscalers that are competing to build LLMs are spending $400 billion in CapEx this year, and $500 billion, which is half a trillion next year,” says Needham & Company’s Martin. “The way to get competitive advantage over your competitors is to have data other people don’t have.”
This aspect of the business will give a whole new revenue stream to media and entertainment. Who knew all that longtail content or video b-roll and outs would be valuable to more than the storage companies one day?
Related Articles
On Wednesday, February 26, leading industry expert Brian Ring will moderate the Streaming Media Connect panel "LLMs on Air: Gen AI Use Cases for News, Sports, and Entertainment." Large language models (LLMs) are making inroads everywhere in the streaming world. As a text-generating subset of Gen AI that stands apart from audio and video creation but is nonetheless having a significant impact on how broadcasters deliver content and how viewers experience it, how will these LLMs get commoditized? This expert panel addresses a host of issues around LLMs and streaming, from aggressive data scraping to metadata-driven discovery and more. Confirmed panelists include experts from Sinclair, Play Anywhere, and evision.
14 Feb 2025
This article explores the current state of AI in the streaming encoding, delivery, playback, and monetization ecosystems. By understanding the developments and considering key questions when evaluating AI-powered solutions, streaming professionals can make informed decisions about incorporating AI into their video processing pipelines and prepare for the future of AI-driven video technologies.
29 Jul 2024