In a blog post last July, Meta CEO Mark Zuckerberg said that selling access to Metas openly available Llama AI models isnt [Metas] business model. Yet Meta does make at least some money from Llama through revenue-sharing agreements, according to a newly unredacted court filing.The filing, submitted by attorneys for the plaintiffs in the copyright lawsuit Kadrey v. Meta, in which Meta stands accused of training its Llama models on hundreds of terabytes of pirated ebooks, reveals that Meta shares a percentage of the revenue that companies hosting its Llama models generate from users of those models.The filing doesnt indicate which specific hosts pay Meta. But Meta lists a number of Llama host partners in various blog posts, including AWS, Nvidia, Databricks, Groq, Dell, Azure, Google Cloud, and Snowflake.Developers arent required to use a Llama model through a host partner. The models can be downloaded, fine-tuned, and run on a range of different hardware. But many hosts provide additional services and tooling that makes getting Llama models up and running simpler and easier.Zuckerberg mentioned the possibility of licensing access to Llama models during an earnings call last April, when he also floated monetizing Llama in other ways, like through business messaging services and ads in AI interactions. But he didnt outline specifics.[I]f youre someone like Microsoft or Amazon or Google and youre going to basically be reselling these services, thats something that we think we should get some portion of the revenue for, Zuckerberg said. So those are the deals that we intend to be making, and weve started doing that a little bit.More recently, Zuckerberg asserted that most of the value Meta derives from Llama comes in the form of improvements to the models from the AI research community. Meta uses Llama models to power a number of products across its platforms and properties, including Metas AI assistant,Meta AI.I think its good business for us to do this in an open way, Zuckerberg said during Metas Q3 2024 earnings call. [I]t makes our products better rather than if we were just on an island building a model that no one was kind of standardizing around in the industry.The fact that Meta may generate revenue in a rather direct way from Llama is significant because plaintiffs in Kadrey v. Meta claim that Meta not only used pirated works to develop Llama, but facilitated infringement by seeding, or uploading, these works. Plaintiffs allege that Meta used surreptitious torrenting methods to obtain ebooks for training, and in the process due to the way torrenting works shared the ebooks with other torrenters.Meta plans to significantly up its capital expenditures this year, largely thanks to its increasing investments in AI. In January, the company said it wouldspend $60 billion-$80 billion on CapEx in 2025 roughly double Metas CapEx in 2024 primarily on data centers and growing the companys AI development teams. Likely to offset a portion of the costs, Meta is reportedly considering launching a subscription service for Meta AI thatll add unspecified capabilities to the assistant. Updated 3/21 at 1:54 p.m.: A Meta spokesperson pointed TechCrunch to this earnings call transcript for additional context. Weve added a Zuckerberg quote from it specifically a quote about Metas intent to revenue share with large hosts of Llama models.