Meta to invest ‘hundreds of billions’ in new multi-gigawatt AI data centers

view original post

Meta Platforms Inc. will invest hundreds of billions of dollars in new data centers, Chief Executive Officer Mark Zuckerberg announced today.

The facilities will support the company’s artificial intelligence development efforts. Last month, it launched a new AI unit called Meta Superintelligence Labs to lead the effort. The group includes prominent machine learning researchers that Meta has poached from rivals such as OpenAI.

“Meta Superintelligence Labs will have industry-leading levels of compute and by far the greatest compute per researcher,” Zuckerberg wrote in a post on Threads today.

The first data center cluster that the company is building as part of the project is called Prometheus. According to Zuckerberg, it’s set to launch next year and will consume multiple gigawatts of power. One gigawatt corresponds to the electricity use of about one million households.

A second data center cluster, Hyperion, will come online at an unspecified later date. It’s expected to require up to 5 gigawatts. That’s more than three times the expected power consumption of OpenAI’s first Starlink data center campus, which is set to use 1.2 gigawatts.

The OpenAI site will reportedly host $40 billion worth of Nvidia Corp. graphics processing units. If Hyperion’s power consumption is any indication, it could be equipped with more than three times as many chips.

Meta may not use Nvidia accelerators alone to power its upcoming data centers. In 2023, the company debuted an internally-developed AI processor called MTIA v1. Last year, Meta detailed a new version with twice the performance and revealed that future iterations will support generative AI workloads. That would make them suitable for training large language models.

Using custom chips could help Meta reduce its reliance on Nvidia GPUs, which cost over $30,000 apiece. But that doesn’t mean it plans to completely phase out the chipmaker’s silicon from its data centers. The 2024 version of MTIA was specifically designed to work with “next-generation GPUs,” Meta detailed at the time of the processor’s debut.

The company’s Hyperion project is believed to be a $10 billion data center campus in Louisiana that it first previewed last year. The project will see Meta build up to nine facilities on a four million square foot site. Entergy Louisiana, a local utility, is reportedly constructing a 1.5-gigawatt power plant to support the cluster.

Meta plans to follow up Hyperion with several additional AI data centers. “We’re building multiple more titan clusters as well,” Zuckerberg wrote on Threads. “Just one of these covers a significant part of the footprint of Manhattan.”

Image: Meta