Ben Pouladian(@benitoz):Last one of the day I was in NVIDIA's Rubin room at Hot Chips. Same loop as this morning. An agent is 20 to 50 inferences, not one. Then they said no one uses one chip. A 100 MW factory. 2 zettaflops of 4-bit inference. 11 petabytes of memory. 800 petabytes a second. I called this the Agentic CPU. This is the GPU half. Token revenue, not a faster chip. They said 30x versus Blackwell Ultra, on real silicon. The slide is 2x versus GB300 until you push interactivity

Ben Pouladian

@benitoz

AI Infrastructure • Compute Economics • Semiconductors • Models • Power • $NVDA since 2016

加入 April 2009

1.3K 正在關注 25.2K 粉絲

Ben Pouladian@benitoz

2026.08.25 00:39

Last one of the day I was in NVIDIA's Rubin room at Hot Chips. Same loop as this morning. An agent is 20 to 50 inferences, not one. Then they said no one uses one chip. A 100 MW factory. 2 zettaflops of 4-bit inference. 11 petabytes of memory. 800 petabytes a second. I called this the Agentic CPU. This is the GPU half. Token revenue, not a faster chip. They said 30x versus Blackwell Ultra, on real silicon. The slide is 2x versus GB300 until you push interactivity