Meta and Google are the latest to trade blows in the arms race for AI data centers, and are pledging billions of dollars in new investments to build large-scale infrastructure.
Meta CEO Mark Zuckerberg said on Monday that the first in the series of AI data, Superclusters, will be online next year as part of a plan to invest “multiple millions of dollars” to secure a spot as the dominant player in artificial intelligence. Supercluster is a large AI data center with the most advanced AI accelerators.
In a post on the Facebook social media platform, Zuckerberg said Prometheus will be one of several sites offering multi-gigawatt computing power. Another planned supercluster, Hyperion, can scale to 5 gigawatts of power.
Meta is not just about building data centers. It builds negotiation power.
Pradeep SanyalAI and data reader, Capgemini
“The meta isn't just about building data centers,” said Pradeep Sanyal, data leader at AI and Capgemini, in an email interview. “It's building negotiating power. … Computing is the new battlefield for Frontier AI. With gigawatt-scale clusters, Meta is taking part in a class of infrastructure that only a handful of players can match.”
Big technology's appetite for AI advances has led to hyperscalers and AI companies competing to build such clusters.
Meta's AI Research SuperCluster has been in operation since 2022. The cluster uses 16,000 NVIDIA A100 GPUs for large-scale language model (LLM) training. This year, Amazon and Anthropic announced a joint supercluster (Project Ranier) using AWS Trainium2 chips. Elon Musk's Xai effort, Colossus is powered by 200,000 Nvidia chips, which require 300 megawatts of power.
Google's Power Play
The drive for AI computing requires enormous amounts of electricity, and large tech companies are competing to strengthen their energy pipelines.
On Tuesday, Google said it will invest $25 billion in data centers and AI infrastructure within the US's largest U.S. electric grid (covering 13 states) over the next two years.
Google also said it would spend $3 billion on updating two Pennsylvania hydroelectric plants to help drive energy-hungry data centers. Pennsylvania's investment is part of an agreement with Brookfield Asset Management to purchase 3,000 megawatts of carbon-free hydroelectric power in the United States
Google's parent company Alphabet said earlier this year that it would spend $75 billion on data centers and infrastructure in 2025. In June, the company said it would partner with CTC Global Corp. to use new technology to free up grid capacity nationwide.
Sanyal said there is a stark contrast between Meta and Google's AI strategy.
“While Meta bets on a bespoke AI supercluster, Google is taking a more balanced route,” he said. “They combine aggressive data center expansion with long-term clean energy procurement. The $3 billion hydro trading ensures up to 3 gigawatts of renewable electricity, providing predictability as well as capacity.”
While Meta is taking part in the scale-first buildout, Google is threading computing, sustainability and grid coordination, he said, adding that “both are aiming for leadership.”
Meta's AI Gambling
Sanyal said Meta's strategy could be to show off their AI infrastructure and talent. The company has already poached top talent from competitors, including Openai and Apple, offering profitable signature bonuses. According to a published report, Meta hired former Safe Superintelligence CEO Daniel Gross.
“Meta is trying to get back to the heart of AI conversations using computing, capital and personnel,” Sanial said. “This also carries serious execution risks. Scaling such infrastructure is difficult. Human resource integration at this level is even more difficult.
It still needs to be determined whether META will be implemented in that strategy, but outcomes may be important no matter what.
“If the meta is run, this could reset the order in AI,” he said. “And if they don't, it could be the most expensive HR experiment in history.”
A veteran journalist with over 20 years of experience, Shane Snider covers IT infrastructure with Informa TechTarget.
