Meta has a revenue sharing agreement with Lama AI model host, filing reveals

in Blog post Last July, Meta CEO Mark Zuckerberg said “selling access” to Meta’s openly available Llama AI models “isn’t it.” [Meta’s] Business model. “But at least we make meta. Some Funds from the Lama through income sharing agreements, according to a newly unedited court filing.

SubmitKadreyv filed by an attorney for the plaintiff in a copyright lawsuit. In Meta, Meta has been accused of training the llama model with hundreds of terabytes of pirated e-books, revealing that Meta is “sharing a percentage of revenue.”

The filing does not indicate which particular hosts pay. However, Meta lists a variety of different llama host partners Blog PostsAWS, Nvidia, Databricks, Groq, Dell, Azure, Google Cloud, Snowflake, etc.

Developers do not need to use the llama model via their host partner. The model can be downloaded, tweaked and run on a range of hardware. However, many hosts offer additional services and tools, making the Llama model simpler and easier to run.

Zuckerberg mentioned the possibility of license access to the Llama model Revenue call last Aprilwhen he monetized the llama in other ways, such as ads and ads for “AI Interactions.” However, he did not outline the details.

“[I]You’re like Microsoft, Amazon, Google and you’re basically going to resell these services. That’s what I think we should get a portion of the revenue,” Zuckerberg said.

Recently, Zuckerberg Claim The fact that most of the value meta is derived from llamas comes in the form of improvements to the AI ​​research community’s model. Meta uses Llama Models to drive numerous products across platforms and properties, including Meta AI assistant Meta AI.

“I think it’s a good business to do this in an open way,” Zuckerberg said. Meta said in its third quarter 2024 revenue call. “[I]T is not only building models that no one in the industry has standardized, but also makes the product even better than if they were on an island. ”

The fact that Meta could potentially make revenue in a rather direct way from the llama is important. This is because the plaintiffs in Cadrey v. Meta argue that Meta not only developed llamas using pirated works, but also promoted infringement by “seeding” or uploading. The plaintiffs allege that Meta used secret torrenting methods to obtain the e-book for training, and in the process, they shared the e-book with other torrenters through the torrenting mechanism.

Meta plans to significantly increase its capital expenditure this year thanks to increased investment in AI. In January, the company said it would spend between $60 billion and $80 billion (about twice the meta CAPEX in 2025) on CAPEX in 2025, expanding its AI development team.

As it is likely to offset some of the costs, Meta is reportedly considering launching Meta AI subscription services that add unspecified functionality to the assistant.

Updated 3/21 at 1:54pm: Meta spokesman pointed to TechCrunch This revenue calls the transcript For additional context. I added a quote from Zuckerberg from there. In particular, it is a quote about meta’s intentions for revenue share with large hosts in the Llama model.