Meta has presented Llama as an openly available family of AI models, but a newly unredacted court filing shows there is also a direct commercial layer around some hosted access. The filing says Meta shares in revenue generated by companies that make Llama available to users through their own services.
The disclosure matters because it connects Meta’s open AI strategy with a money flow from hosted Llama usage. It also arrives inside Kadrey v. Meta, a copyright lawsuit focused on how Meta allegedly trained Llama and obtained e-books used in that training.
What the filing says about Llama revenue
The filing, submitted by attorneys for the plaintiffs in Kadrey v. Meta, says Meta “shares a percentage of the revenue” that companies hosting Llama models generate from users of those models.
The filing does not name the specific hosts that pay Meta. However, Meta has identified a range of Llama host partners in blog posts, including AWS, Nvidia, Databricks, Groq, Dell, Azure, Google Cloud, and Snowflake.
That distinction is important. The filing points to revenue sharing with companies that host Llama, not a requirement that every developer must pay Meta to use the models. According to the source, developers can still download, fine-tune, and run Llama models on different hardware.
Hosting partners can still play a major role because they offer services and tooling that make deployment easier. For developers and companies that do not want to manage the full technical setup themselves, hosted access can be the practical path to using Llama.
Why this complicates Meta’s open model message
Last July, Meta CEO Mark Zuckerberg wrote in a blog post that “selling access” to Llama “isn’t [Meta’s] business model.” The newly unredacted filing does not necessarily say Meta sells access directly in the same way as a closed AI service. It does, however, show that Meta can receive money when host partners generate revenue from Llama users.
Zuckerberg had already signaled that large cloud and infrastructure companies could be part of a commercial model for Llama. During an earnings call last April, he discussed the idea that companies such as Microsoft, Amazon, or Google reselling these services could owe Meta “some portion of the revenue.”
More recently, Zuckerberg has emphasized a different kind of value from Llama: improvements that come from the AI research community. He said during Meta’s Q3 2024 earnings call that working openly helps Meta’s products become better than if the company were building alone without broader industry adoption.
Both ideas can exist at the same time based on the facts in the source. Meta can benefit from outside research, adoption, and standardization while also receiving revenue from some hosting arrangements. The filing makes that commercial side harder to treat as only theoretical.
The lawsuit raises the stakes
The revenue-sharing detail is significant because of the claims in Kadrey v. Meta. Plaintiffs allege that Meta trained Llama models on hundreds of terabytes of pirated e-books.
The plaintiffs also claim Meta facilitated infringement by “seeding,” or uploading, those works. Their allegation is that Meta used surreptitious torrenting methods to obtain e-books for training and, because of how torrenting works, shared those e-books with other torrenters in the process.
The source does not say the filing proves those claims. It says the filing reveals revenue-sharing agreements while the broader lawsuit accuses Meta of using and sharing pirated works in connection with Llama development.
That pairing is why the business model question matters. If Llama creates direct revenue for Meta through hosted usage, the debate is not only about open access, research benefits, or product improvement. It also touches the commercial value Meta may receive from models whose training is being challenged in court.
AI spending adds another pressure point
Meta is also preparing for much larger AI-related spending. In January, the company said it would spend $60 billion-$80 billion on CapEx in 2025, roughly double Meta’s CapEx in 2024.
The source says that spending is primarily tied to data centers and the growth of Meta’s AI development teams. Those investments show how central AI has become to Meta’s plans across its platforms and properties.
Llama already supports products including Meta’s AI assistant, Meta AI. The more Meta invests in AI infrastructure and teams, the more pressure there may be to offset some of those costs through services, partnerships, or products connected to its AI systems.
The source also says Meta is reportedly considering a subscription service for Meta AI that would add unspecified capabilities to the assistant. No specific capabilities are described, and the source does not say whether such a service will launch.
What to watch next
The key unanswered point from the filing is which Llama hosts are paying Meta and under what terms. The source says the filing does not identify the specific hosts involved.
For now, the clearest takeaway is narrower but still meaningful: Llama is not only an openly available model family that developers can download and run. It is also part of a hosting ecosystem where Meta can receive a share of revenue from companies serving Llama to users.
That does not erase Meta’s argument that open distribution can improve its models and products. It does make the company’s Llama strategy look more layered: open availability, partner-hosted convenience, product integration, and direct revenue can all be part of the same picture.