Used by approximately 3.6 billion people meta The social networking company plans to operate 32 data centers to handle the load with the completion of a new data center in Oklahoma. But that’s not enough.
Amazon’s The cloud division announced Friday that Meta has agreed to use Amazon’s generic Graviton chip in a deal that will last at least three years.
The deal shows Meta, under CEO Mark Zuckerberg, is willing to splurge to meet high computing demands. alphabet and microsoft. In recent weeks, Meta has signed deals with CoreWeave and CoreWeave worth a combined $48 billion. Neviusboth rent access to Nvidia graphics processing units (GPUs) that run AI models.
Amazon has not disclosed the value of the meta transaction.
Meta is balancing infrastructure expansion with headcount reduction. The company announced Thursday that it plans to lay off about 8,000 employees, or 10% of its workforce.
Unlike Nvidia GPUs, armBased on the top cloud Amazon Web Services’ Graviton processor, it can handle a variety of computing tasks. Intel’s or AMD’s central processing unit, or CPU. However, Graviton is still useful for AI workloads, especially for tuning and post-training tasks after training a model on large amounts of data using large compute clusters.
“Graviton is one of the most popular platforms for pre-training by many foundational model companies, and Meta is one of the newest today,” said Nafea Bshara, AWS Vice President and Distinguished Engineer.
Bshara co-founded Annapurna Labs, a chip company that Amazon acquired in 2015. Since then, Amazon has developed specialized chips for training and running AI models, among other components. Graviton became a huge hit. adobe, apple and snowflake. Earlier this week, Amazon-backed AI model builder Anthropic also announced plans to use Graviton processors.
According to AWS, Graviton can reduce energy consumption by 60% while providing the best performance at a given price among all computing options available through the EC2 Compute service.
Bushara said Meta has been using Graviton chips on a small scale, but now plans to leverage hundreds of thousands of chips and become one of Graviton’s top five customers. He said the company has been renting Nvidia GPUs from AWS since 2017.
On Thursday, Intel CEO Lip Vu Tan told analysts that demand for the company’s Xeon server chips is outstripping supply.
“For the past few years, the conversation around high performance computing has been exclusively about GPUs and other accelerators,” Tan says. “In recent months, we have seen clear signs that CPUs are once again establishing themselves as the essential foundation of the AI era.”
But Meta didn’t choose Graviton because no other types of CPUs were available, Bshara said.
“Our expansion to Graviton allows us to run CPU-intensive workloads behind agent AI with the performance and efficiency we need at our scale,” Santosh Janardhan, head of infrastructure at Meta, reportedly said in a statement.
WATCH: Amazon and Anthropic announce new deal to provide up to 5 gigawatts of compute for Claude AI models
