Post

Justin Garrison
@justingarrison.com

They upgraded to GB300 NVL72 racks with a bullet point that it has over 1 exoFLOP of FP4 compute. The GB200 rack says it has 1.4 exoFLOP of FP4 compute they just wrote it as "1,440 PFLOPS"

Justin Garrison
@justingarrison.com

Are you telling me that last year with GB200 12% of your GPUs were "interrupting" every day??? The stat says 1000 failures per 8000 GPUs per day. That's 1 in 8. IDK about you, but to me that sounds like a completely broken product

0 reposts4 likes

Justin Garrison
@justingarrison.com

This kinda kills me. They buy racks of 72 GPUs but only use 64 (8 spares) because the jobs work better with powers of 2. An 11% stand by capacity is a lot of money They also have 2% extra racks just waiting around

Justin Garrison
@justingarrison.com

Their goal is 1M GPUs and let's just use 72 GPU/rack as a round number. That means they have 13,888 active racks and 277 SPARES!!! A GB300 NVL72 rack costs about $5,000,000 which means they have (list price) $1.3B of spare capacity

Justin Garrison
@justingarrison.com

This really makes me think how much the frontier labs are spending when they rent the entire infrastructure from hyperscalers Datacenters are expensive. Not owning the infrastructure is expensive and risky

Hazel Weakly
@hazelweakly.me

Yup all these numbers seem about right to me. Although I would’ve expected more hardware failure and a higher reserve rate to try and account for hardware shortages. Would’ve expected 22-25% reserve maybe to factor in growth as well. I’m guessing the board shot that down?

CreatorCoreForge
@ccfteam.bsky.social

Some of it's failure-rate redundancy, sure — but $1.3B of 'just in case' GPUs is still a bill someone pays. Idle capacity is exactly the spend I built CoreForge Meter to catch: it never shows up in the keynote, only the invoice.