How Many Will Actually Get Built & Is Energy AI's BIGGEST Bottleneck? | Positron AI Co-founder
Thomas Sohmers, co-founder of Positron AI, joins Harry Stebbings to discuss why energy and regulation, not compute, are the real limits on AI growth.
The gist
-
Inference hardware is memory-bound, not compute-bound, because memory bandwidth only improved 17x over a decade while compute improved 120x.
-
Token prices have dropped 60x in five years, from $60 to $1 per million, but each token delivers far more intelligence than before.
-
Sohmers argues anti-data-center campaigns in the West are built on false claims and effectively hand strategic ground to China.
-
Smarter on-device models will trigger more calls to frontier cloud services, pushing total token consumption up by orders of magnitude.
"I was at the ChatGPT launch at NeurIPS in 2022. When I went back to the hotel, I loaded it up and I was up for four or five hours straight just giving random prompts — this was the most magical experience that I've ever had with a computer."
▶ Play at 55:43159 creators recommend it, including John Coogan, Riley Brown, and Peter Yang.
See all on the ChatGPT pageWhat they recommend · 2 picks
"I was at the ChatGPT launch at NeurIPS in 2022. When I went back to the hotel, I loaded it up and I was up for four or five hours straight just giving random prompts — this was ..."
"I got very close when Sora 2 came out — I had a similar experience, short amount of time, but just mind blown by the quality of the videos, especially the weekend that Sora 2 la..."
Also mentioned
Related episodes
Episodes covering similar topics
More from 20VC with Harry Stebbings