Reply Estimates 100K GPUs for Token Capacity
Reply discusses token output commitment and GPU fleet estimates for Chinese AI.
@teortaxesTex replied to @ohlennart that people took a statement as a commitment for output tokens. The reply noted it would require about 100K GPUs and suggested that matches their fleet if other operations halted. @ohlennart stated China would serve more tokens in absolute terms but not relatively. US companies could produce 15x to 20x more quality-adjusted tokens. The reply left open when deployment advantages matter and how to apply them.
Combined views
3.1K
2 posts, first seen 2d ago
Reactions
2 postsT( 
Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex@ohlennart I think people (me included) took this as a commitment for output tokens. Note that it'd have taken only about 100K GPUs still. I think that's their fleet, and so theoretically they could have served as much by halting everything else
- likes: 4
- replies: 0
- bookmarks: 1
- reposts: 0
LH Lennart Heim@ohlennartExpect China to serve more tokens, at least absolutely not relatively. US companies could do 15x to 20x more quality-adjusted tokens. But for some threats, that won't matter. IMO the key unanswered questions: when does this deployment advantage matter and how do you use it?
- likes: 13
- replies: 1
- bookmarks: 1
- reposts: 3
