one year on
Zuckerberg: Meta will spend 'hundreds of billions' on multi-gigawatt superclusters
Prometheus (1 GW, online 2026) and Hyperion (scaling toward 5 GW) get names and dates — with tents going up over server racks because buildings take too long.
Mark Zuckerberg puts names, numbers, and dates on Meta’s compute ambitions: Prometheus, a 1-gigawatt supercluster coming online in 2026, and Hyperion, designed to scale to 5 gigawatts over several years — with “multiple more titan clusters” behind them and total spending in the hundreds of billions of dollars.
The detail that travels furthest comes from industry analysts: Meta is erecting tents over server racks at its data center sites, because conventional construction is too slow for the timelines superintelligence apparently demands.
For scale: 5 gigawatts is roughly the output of a large nuclear power station, dedicated to a single company’s AI training. Two weeks after the Superintelligence Labs memo, the message is consistent — Meta intends to buy its way back to the frontier, in researchers and in watts, and considers the price of missing it higher than the price of overbuilding.
The record
Five gigawatts is a nuclear plant's worth of AI. The grid, not the GPU supply, starts looking like the binding constraint.
Ask what product, exactly, justifies capex approaching an Apollo program — and whether 'superintelligence' is a business plan or a vibe.
One year later — open only if you can handle spoilers
The tents were real, the spend was real, and it kept escalating — Meta's 2026 capex guidance came in above $100B as the industry's compute buildout became a macroeconomic story. The 'is this a bubble' debate that started here never resolved; it just got bigger numbers.
The Weekly Replay · free by email
This week, one year ago — every Sunday.
One email each Sunday: the week's replayed AI news, with the one-year-later annotations included. Written like it's breaking — dated like it isn't.
Free · double opt-in · unsubscribe anytime · privacy