In a recent blog post from last July, Meta’s CEO Mark Zuckerberg stated that the company’s business model does not involve “selling access” to its openly available Llama AI models. However, a newly unredacted court filing reveals that Meta generates revenue from Llama through revenue-sharing agreements with certain companies.

The filing, submitted by the plaintiffs’ attorneys in the copyright lawsuit Kadrey v. Meta, discloses that Meta “shares a percentage of the revenue” earned by companies hosting its Llama models from users. Although the filing does not specify which hosts pay Meta, the company has listed several Llama host partners in its blog posts, including AWS, Nvidia, Databricks, Groq, Dell, Azure, Google Cloud, and Snowflake.

Developers have the option to use Llama models through a host partner or download, fine-tune, and run them on various hardware. Many hosts offer additional services and tooling that simplify the process of getting Llama models up and running. During an earnings call last April, Zuckerberg mentioned the possibility of licensing access to Llama models and exploring other monetization strategies, such as business messaging services and ads in “AI interactions.”

Zuckerberg noted that if companies like Microsoft, Amazon, or Google resell Llama services, Meta should receive a portion of the revenue. He stated, “[I]f you’re someone like Microsoft or Amazon or Google and you’re going to basically be reselling these services, that’s something that we think we should get some portion of the revenue for.” Meta has started making such deals, according to Zuckerberg.

More recently, Zuckerberg emphasized that the majority of the value Meta derives from Llama comes from improvements to the models made by the AI research community. Meta utilizes Llama models to power various products across its platforms, including its AI assistant, Meta AI. Zuckerberg believes that making Llama open-source is a good business decision, as it enhances Meta’s products and encourages standardization within the industry.

The revelation that Meta generates revenue from Llama is significant, given the plaintiffs’ allegations in Kadrey v. Meta that the company used pirated works to develop Llama and facilitated infringement by “seeding” these works. The plaintiffs claim that Meta used surreptitious torrenting methods to obtain ebooks for training and, in the process, shared them with other torrenters.

Meta plans to substantially increase its capital expenditures this year, primarily due to its growing investments in AI. The company aims to spend $60 billion-$80 billion on CapEx in 2025, roughly double its 2024 expenditure, mainly on data centers and expanding its AI development teams.

To offset some of the costs, Meta is reportedly considering launching a subscription service for Meta AI that will offer additional capabilities to the assistant.

Updated 3/21 at 1:54 p.m.: A Meta spokesperson directed TechCrunch to this earnings call transcript for further context. We have added a relevant quote from Zuckerberg, discussing Meta’s intent to revenue-share with large hosts of Llama models.