In a blog post from July, Meta’s CEO, Mark Zuckerberg, stated that “selling access” to Meta’s publicly accessible Llama AI models is not part of the company’s business strategy. However, a newly unredacted court document indicates that Meta generates some income from Llama through revenue-sharing arrangements.
The court filing, submitted by the plaintiffs’ attorneys in the copyright lawsuit Kadrey v. Meta, where Meta is accused of using pirated ebooks to train its Llama models, discloses that Meta shares a portion of the revenue generated by companies hosting its Llama models with those hosts.
The specific hosts that transfer payments to Meta are not disclosed in the filing. Nonetheless, Meta has mentioned several partners hosting Llama models in its blog posts, which include AWS, Nvidia, Databricks, Groq, Dell, Azure, Google Cloud, and Snowflake.
Developers are not mandated to utilize a host partner to access Llama models, as the models can be downloaded, customized, and executed on various hardware. Despite this, many hosts offer added services and tools that facilitate easier deployment and operation of Llama models.
Zuckerberg previously discussed the potential for licensing Llama model access during an earnings call last April, suggesting other monetization strategies such as through business messaging services and advertisements in “AI interactions,” though he did not provide detailed plans.
Recently, Zuckerberg has emphasized that the majority of the value Meta gains from Llama is through enhancements to the models contributed by the AI research community. Llama models are employed by Meta to power various products across its platforms, including Meta’s own AI assistant, Meta AI.
During the Q3 2024 earnings call, Zuckerberg remarked, “I think it’s good business for us to do this in an open way,” highlighting that it improves their products more effectively than if Meta developed a proprietary model in isolation.
The revelation that Meta might earn direct revenue from Llama holds significance as it contradicts Meta’s alleged actions in the Kadrey v. Meta case. Plaintiffs claim Meta used pirated content to create Llama and facilitated infringement by “seeding” these works. They allege Meta utilized covert torrenting to acquire ebooks for training, inadvertently sharing them due to torrenting’s nature.
Meta anticipates a substantial increase in capital expenditures this year, driven largely by advancing investments in AI. The company has announced plans to spend between $60 billion and $80 billion on capital expenditures in 2025, doubling its 2024 spending, to focus on data centers and expanding AI development teams.
To potentially counterbalance these expenses, Meta is reportedly considering introducing a subscription service for Meta AI, promising additional features.
Meta has yet to respond to a request for comment, and updates will follow upon receiving further information.
