Kairos(@KairosPraxis):Cool chart from @SemiAnalysis_ showing that the frontier labs have now shifted most of their compute from pre-training to post-training (mostly RL). Unlike pre-training, the bottleneck moves from compute to networking - how fast you read from memory/access CPU. An interesting implication is that you don't need to stack memory (DRAM dies) anymore as this doesn't improve bandwidth so we've gone from 12 dies ---> 8 dies ---> 4 dies? So, I'm guessing companies that do scale up and scale in win here?

Kairos

@KairosPraxis

The first rule of investing is to not take yourself too seriously, and actually have fun.

参加 February 2025

463 フォロー中 5.9K ファン

Kairos@KairosPraxis

2026.09.19 15:09

Cool chart from @SemiAnalysis_ showing that the frontier labs have now shifted most of their compute from pre-training to post-training (mostly RL). Unlike pre-training, the bottleneck moves from compute to networking - how fast you read from memory/access CPU. An interesting implication is that you don't need to stack memory (DRAM dies) anymore as this doesn't improve bandwidth so we've gone from 12 dies ---> 8 dies ---> 4 dies? So, I'm guessing companies that do scale up and scale in win here?