Orbit · Explore

AI Lab · by Hamza Malik · 2026-10-20

Inside Prometheus: Meta's gigawatt-scale GB200/GB300 supercluster

engineering.fb.com

Meta's engineering blog has a deep dive on Prometheus, their gigawatt-scale supercluster built on GB200 and GB300 systems. Gigawatt is a power-plant number, which says a lot about where frontier training is headed. Posts like this are some of the best public info on how these clusters are actually built, from power to networking to reliability. What stood out to you, and does anyone think this scale actually translates into a model lead for Meta?

78 sparks · 18 comments

Comments

Ananya Rao · 14 likes

serious question, where is all the power coming from and who pays when the grid gets tight

Narin Chai · 12 likes

this is the part of the ai boom nobody wants to discuss at earnings calls

Chiara Romano · 9 likes

the local ratepayers, usually

Rafael Almeida · 12 likes

Cooling must be insane. Is it all liquid now at this density?

Noah Johansson · 10 likes

pretty much has to be for those racks from what I understand

Tiny Mango · 11 likes

engineering blog posts like this are honestly the best thing meta does for the community

Minji Kim · 3 likes

between this and the open hardware stuff, yeah, credit where due

Camille Dubois · 11 likes

my homelab draws 400W and my partner thinks thats too much

Urban Fern · 2 likes

show them this article, instant perspective

Bilal Raza · 10 likes

Mixing GB200 and GB300 in one cluster sounds like a scheduling headache. Would love to know how they handle the heterogeneity for a single training job

Liam Wilson · 1 likes

that'd be my guess too, nobody wants stragglers on a run that size

Rafael Almeida · 0 likes

probably split by job type rather than mixing in one run, but I'm guessing

Tola Ibrahim · 7 likes

the reliability part is what interests me. at this scale something is always failing, checkpointing strategy is everything

Daniela Torres · 11 likes

the section on failures was the best part for me

Mehdi Benali · 5 likes

compute was never the bottleneck for meta's models, imo it's been research direction and execution

Amelia Hughes · 1 likes

spicy but probably true

Yui Tanaka · 1 likes

the fact that we talk about clusters in gigawatts now is wild. that's a city's worth of power

Camila Rocha · 3 likes

and we're arguing about whether it can write a for loop properly