The cost per unit of AI intelligence is falling so fast (about 40x per year) that it's a scary exponent for infrastructure buildout, and like every past tech infrastructure cycle, some people will get badly burned by their compute contracts.

20:08intelligence let's say it's been averaging like 40x X for a given level per year. You know, that's like a very scary exponent from an infrastructure buildout standpoint. Now, again, we're taking the bet that there will be a lot more demand as that gets cheaper, but I have some fear that it's just like, man, we keep going with these breakthroughs and everybody can run like a personal AGI on

20:30their laptop and we just did an insane thing here. Some people are going to get really burned like has happened in every other tech infrastructure cycle at some points along the way. >> I think that's really well said and you have to hold those two simultaneous truths. We had that happen in 20201 and yet the internet became much bigger and produced much greater outcomes for society than anybody estimated in that

20:53period of time. >> Yeah. But I think that the one thing that Sam said is not talked about enough which is the current for example the optimizations that OpenAI has done on the inference stack for a given GPU. I mean I it's kind of like it's you know we talk about the MOS law improvement on one end but the software improvements are much more exponential than that.

21:14Someday we will make a incredible consumer device that can run a GPT5 or GPD6 capable model completely locally at a low power draw. And this is like so hard to wrap my head around. >> That will be incredible. And you know that's the type of thing I think that scares some of the people who are building obviously these large centralized compute uh stacks. And Satcha you've talked a lot about the

21:38distribution both to the edge as well as having inference capability distributed around the world. Yeah, I mean the way at least I've thought about it is more about really building a fungeable fleet. I mean when I look at sort of in the cloud infrastructure business, one of the key things you have to do is have two things. One is an effic like in this context in a very efficient token

21:59factory and then high utilization. That's that's it. There are two simple things that you need to achieve and in order to have high utilization you have to have multiple workloads that can be scheduled even on the training. I mean, if you look at the AI pipelines, there's pre-training, there's mid-training, there's post- training, there's RL. You want to be able to do all of those things. So, thinking about fungeibility