Meta has revenue sharing agreements with Llama AI model hosts, filing reveals

In a blog post last July, Meta Mark Zuckerberg CEO said that “selling access” to Meta’s Llama Ai is openly available. [Meta’s] Employment model. However, Meta is made at least some Funds from Lama through revenue sharing agreements, according to the newly classified court.

The deposit, submitted by lawyers for prosecutors in a lawsuit for copyright. Meta, where Meta is accused of training Llama models on hundreds of terabytes from pirated e -books, reveals that “Meta” shares a percentage of revenue “that hosts Llama forms of users of these models.

The deposit does not indicate the specific hosts pay dead. But Meta recounts a number of Llama host partners in various bloggers, including AWS, Nvidia, Databrics, GROQ, Dell, Azure, Google Cloud and Snowflake.

The developers are not required to use the Llama model through a host partner. The models can be downloaded, adjusted, and operated on a group of different devices. But many hosts provide additional services and tools that make Llama models more simple and easier.

Zuckerberg reported the possibility of licensing access to Llama models during the profit call last April, when he also reinforced Lama in other ways, as is the case through commercial correspondence services and ads in “AI Interactions”. But he did not specify the details.

“[I]”You are a person like Microsoft, Amazon or Google, and you will mainly resell these services, and this is something we think we must get part of the revenue.”

Recently, Zuckerberg stressed that most of the Llama’s Meta values ​​come in the form of improvements to models from the artificial intelligence research community. Meta uses Llama models to play a number of products through their platforms and features, including Meta’s Ai, Meta Ai.

“I think it is good for us to do this in an open manner,” Zuckerberg said during the Meta Q3 2024 profit call. “[I]T makes our products better instead of being on the island of building a model that no one was uniting in this industry. “

The fact that Meta may generate revenues directly from Llama is important because the plaintiffs in Kadrey V. Meta claims that Meta does not use pirated business only to develop Lama, but facilitate violation by “seed” or loading, these works. Prosecutors claim that Meta used hidden torrent methods to obtain e-books for training, and in this process-because of the way the torrent works-e-books shared with other torrents.

Meta plans to increase its capital expenses significantly this year, thanks to its increasing investment in artificial intelligence. In January, the company said it would spend $ 60 billion-80 billion dollars on CEPEX in 2025-near Capex Meta in 2024-in the first place on data centers and the development of the company’s AI development teams.

It is possible to compensate for part of the costs, and according to what Meta thinks about launching a subscription service for Meta AI, which will add unlimited capabilities to the assistant.

Updated 3/21 at 1:54 pm: A Meta Techcrunch spokesman indicated this profit call version for an additional context. We added a Zuckerberg quote from him – specifically a quote about Meta’s intention to the revenue share with a large host of Llama models.