Meta's multi-gigawatt AI supercluster flex

While most tech giants are still counting GPUs, Mark Zuckerberg is preparing to plug entire cities worth of electricity directly into his AI clusters.

Zuckerberg just announced that Meta will spend hundreds of billions of dollars on a fleet of multi-gigawatt AI supercomputing data centers. The massive buildout is aimed at fueling Meta's race toward "superintelligence" under its newly minted Meta Superintelligence Labs, an elite unit stocked with top researchers poached from rivals like OpenAI.

Zuckerberg claimed the new lab will boast industry-leading compute and "by far the greatest compute per researcher."

Titan clusters the size of Manhattan

The first multi-gigawatt facility in this offensive, dubbed Prometheus, is slated to come online in 2026 in New Albany, Ohio. A single gigawatt can power around one million homes, giving an idea of the sheer scale of the electricity involved.

A second site named Hyperion will follow in Louisiana, scaling up to an eye-watering 5 gigawatts by 2030. That is more than triple the 1.2-gigawatt draw of OpenAI's first Starlink data center campus, which reportedly houses $40 billion worth of Nvidia silicon. Local utility Entergy Louisiana is already constructing a dedicated 1.5-gigawatt power plant just to keep Hyperion fed.

Zuckerberg noted on Threads that Meta is planning multiple other "titan clusters," adding that just one of these massive sites covers an area comparable to a significant footprint of Manhattan.

Piles of cash and custom silicon

Meta generated over $160 billion in revenue in 2024, and Zuckerberg is clearly using that massive ad-fueled war chest to outspend the competition. As Cambrian AI Research analyst Karl Freund put it, Zuckerberg intends to spend his way to the top of the AI heap by handing poached talent the best hardware on the planet.

Meta is also trying to keep suppliers honest. While it will continue buying high-end Nvidia accelerators, Meta is pairing them with its custom in-house MTIA processors. The latest MTIA silicon is built to run generative AI workloads alongside next-gen GPUs, blunting the sting of Nvidia chips that run over $30,000 apiece.

All this horsepower comes with steep environmental overhead. AI facilities require massive amounts of energy and cooling water, with studies warning global data centers could swallow 1.7 trillion gallons of water by 2027. A single query can consume as much cooling water as a standard disposable water bottle.

Why it matters

For developers and the broader open-source ecosystem, Meta's mega-investment is a huge deal. As long as Meta continues its strategy of releasing open-weight models like Llama, independent builders will continue to reap the benefits of frontier compute without being trapped inside closed proprietary APIs.

For the AI landscape, the goalposts have fundamentally shifted from software to heavy industrial infrastructure. AI dominance is no longer just about prompt engineering or clever training tricks; it has become a brute-force contest over land, gigawatts, and grid capacity. When Meta commits to 5-gigawatt sites, it forces Microsoft, Google, and Amazon to match that infrastructure spend, making it virtually impossible for standalone AI startups to compete at the foundation model tier without deep-pocketed sponsors.

Meta's aggressive strategy looks smart on paper: use core advertising cash flow to lock down long-term energy and silicon capacity while balancing Nvidia with internal MTIA chips. The obvious risk, however, is capital efficiency. If frontier scaling laws plateau before "superintelligence" materializes, Meta could find itself saddled with hundreds of billions in specialized, depreciating infrastructure that makes the metaverse pivot look like pocket change.