I really like your use of “advanced compute,” Arjun. AI has become an overused buzzword.
In my view, though, you are not cautious enough on AI data centers intended to run LLMs. LLMs are discussed as the next great transformative technology.
Well, so were railroads, and the initial group of companies that tried to build and operate railroads all went bankrupt. And so were fiber optics, and the initial group of companies that tried to build and operate fiber optics at scale at least mostly went bankrupt. There are many other historical examples of the same phenomenon.
But today’s situation is worse, because LLMs have been vastly overhyped and cannot by their nature deliver the alleged capabilities. Advanced compute will deliver a lot of those over time, but is highly unlikely to need computing on the scale the LLMs would.
Finally it is worth remembering that the “secure, vast earnings” of the hyperscalers come primarily from advertising. This is a fundamentally ephemeral source.
It seems to me that there is an excellent chance of very negative outcomes here that take awhile to resolve.
Hi Paul, Thank you on "advanced compute". completely agree. On power demand, I feel good about a 2%-3% CAGR in U.S. and 4% globally. AI or advanced compute is maybe 1/3 of that growth. Whether tech is overvalue, overhyped, etc., may well be true. The hysteria over data center pushback was more the point, than how hyperscalers or LLM companies are value.
I really like your use of “advanced compute,” Arjun. AI has become an overused buzzword.
In my view, though, you are not cautious enough on AI data centers intended to run LLMs. LLMs are discussed as the next great transformative technology.
Well, so were railroads, and the initial group of companies that tried to build and operate railroads all went bankrupt. And so were fiber optics, and the initial group of companies that tried to build and operate fiber optics at scale at least mostly went bankrupt. There are many other historical examples of the same phenomenon.
But today’s situation is worse, because LLMs have been vastly overhyped and cannot by their nature deliver the alleged capabilities. Advanced compute will deliver a lot of those over time, but is highly unlikely to need computing on the scale the LLMs would.
Finally it is worth remembering that the “secure, vast earnings” of the hyperscalers come primarily from advertising. This is a fundamentally ephemeral source.
It seems to me that there is an excellent chance of very negative outcomes here that take awhile to resolve.
Finally, enjoy 250!
Hi Paul, Thank you on "advanced compute". completely agree. On power demand, I feel good about a 2%-3% CAGR in U.S. and 4% globally. AI or advanced compute is maybe 1/3 of that growth. Whether tech is overvalue, overhyped, etc., may well be true. The hysteria over data center pushback was more the point, than how hyperscalers or LLM companies are value.