SemiAnalysis Reports Rubin Ultra Memory Cut to 192GB
SemiAnalysis says the chip will use 192GB HBM4, below earlier plans and regular Rubin.
TLDR
SemiAnalysis posted that Nvidia Rubin Ultra will ship with 192GB of HBM4 in an 8-hi stack. The account called the amount a massive downgrade from the 1TB originally previewed and noted it falls below the 288GB in regular Rubin. The post attributed the reduction to fewer memory cubes, a lower stack height, and the switch to HBM4 dies. AI commentator scaling01 replied with a bearish take on model scaling limits. Replies on X split between concerns over constrained supply and arguments that efficiency gains offset the cut. The packet contains no first-party Nvidia confirmation.
Combined views
964.4K
6 Sources, first seen ago
Sources
LA 
Lisan al Gaib@scaling01bearish as fuck we are going to be stuck on small ass models forever https://twitter.com/SemiAnalysis_/status/2091994110911557679
- likes: 627
- replies: 45
- bookmarks: 105
- reposts: 12
ZE Zephyr@zephyr_z9Nah, they are blowing up the scale up world size to compensate for lower HBM per GPU https://twitter.com/scaling01/status/2092000566905888955
- likes: 215
- replies: 8
- bookmarks: 52
- reposts: 5
DT DAY TRADER@SedarZan52448@SemiAnalysis_ The headline is a downgrade, but the bigger story is efficiency. If Nvidia can deliver comparable AI performance with significantly less HBM, that changes the economics of scaling AI infrastructure. Less memory. More efficiency. Bigger scale.
PU Puck!@puhcko@scaling01 They’re literally doing it to stabilize futures pricing because their compute backs loans now gdi
JB Johny Bigood@JohnyBeGood26@SemiAnalysis_ 🤡🤡🤡🤡🤡🤡🤡🤡🤡🤡🤡🤡🤡🤡🤡🤡🤡🤡🤡🤡🤡🤡🤡🤡🤡🤡
BE Benthecarman@benthecarman@scaling01 This is bullish, the supply is extremely constrained for memory.