The Peter Zeihan Podcast Series - Artificial Intelligence Isn't Ready for Mass Application || Peter Zeihan

Episode Date: January 3, 2025

Today's AI technology, while promising, isn't quite ready for widespread application. I'm not talking so much about AI's capabilities, but rather the hardware limitations and supply chain challenges that are getting in the way. Join the Patreon here: https://www.patreon.com/PeterZeihanFull Newsletter: https://mailchi.mp/zeihan/artificial-intelligence-isnt-ready-for-mass-application

Everybody, Peter Zine here, coming to from Rear, Beach, just north of Boston. A lot of you have written in asking for my opinions on AI, so here we go, pick it apart however you will. It's tantalizing. Chad GPT and the large language models are taking us forward. They're nowhere near conscious thought. Oftentimes they can't even associate their own work from previously in a conversation with itself. It's basically targeted randomness, if you will. That said, it is still providing insights

and the ability to search vast databases in a much more organized and coherent matter than anything we have seen from search engines before. So promising tech, we've had a taste. It's definitely not ready for what I would consider mass application, but the possibilities are there, especially when it comes to data management, which when it comes to things like research and genetics

is very important. However, I think it's important to understand what the physical limitations are of AI, and that is a manufacturing issue. So the high-end chips that were using, the GPUs, graphics processing units, were not designed to run AI models. They were designed to run multiple things simultaneously for graphics, primarily for gaming consoles. And the gamers among us, who have logged lots of time,

Doom and Fortnite and all the rest have been the primary economic engine for pushing these technologies forward until very recently. It's only with things like autonomous driving electric vehicles that we've had a larger market for high-end chips. But the GPU specifically because they run multiple scenarios and computations

simultaneously, that is what makes a large language model work. Wow, it got windy all of a sudden. Let me make sure this works. Okay, so GPUs. They generate a lot of heat because they're doing multiple things at the same time. And so normally you have a gaming console, you have a GPU at the heart of it, and multiple cooling systems, typically fans, blowing on them to keep your laptop from catching on fire.

So if you take these and put 10 or 20,000 of them in the same room in a server farm, you have a massive heat problem. And that's why most forecasts indicate that the amount of electricity we're using for data centers is going to double in the next few years. to compensate. That's why they're so power-intensive. Now, if you want to design a chip that is four large language models and AI systems, as opposed to that's just being an incidental use, you can. Those designs are being built now, and we're hoping to have a functional prototype by the end of calendar year 2025. If that is successful, then you can have your first mass run

of the chips, enough to generate enough chips for a single server farm by the end of 2026. And then you can talk about mass manufacture getting into the system by 2029, 2030. So, you know, even in the best case scenario, we're not going to have custom design chips for this anytime soon. Remember that a GPU is about the size of a postage stamp because it's designed to be put on a laptop where if you're going to design a chip specifically to run AI, you're talking about something that is bigger than a dinner plate because it's going to have the cooling system built in. And not to mention being able to run a lot more things in parallel. So even in the best case scenario, we're looking at something that's quite a ways out. Then you have to consider the supply chain just to make what we're making now.

The high-end chip world, especially sub-10 nanometer, and we're talking here about things that are in the 4-nometer and smaller range, closer to 2, really, is the most sophisticated and complicated and proprietary supply chain in human history. There are over 9,000 companies that are involved in making the stuff that goes into the stuff that goes into the stuff that ultimately allows TSM to make these chips in Taiwan. And then of course, 99% of these very high-end chips are all made in one town in Taiwan that faces the People's Republic of China. So it doesn't take a particularly egregious scenario to remove some of those 9,000 pieces from the supply chain system.

And since roughly half of those supply chain steps are only made by small companies that produce one product for one end user and have no competition globally, you can lose a handful of them and you can't do this at all until you rebuild the ecosystem. Based on what goes wrong, that rebuilding can take upwards of 10 to 15 years. So in the best case scenario, we need new hardware that we're not going to have for a half a decade. And a more likely scenario, we're not going to have the supply chain system in order to build the hardware for a decade or more. However, we've already gotten that taste of what AI might be able to do. And since with the baby boomer retirement, we're entering into a world of both labor and capital shortages,

the idea of having AI or something like it to improve our efficiency as something we can't ignore. The question is whether we're going to have enough chips to do everything we want to do, and the answer is a hard no. So we're going to have to choose. do we want the AI chips running to say crack the genome so that we can put out a new type of GMO in the world that'll save a billion people from starving to death in a world where agricultural supply chains fail?

Do we use it to improve worker productivity in a world in which there just aren't enough workers and in the case of the United States, we need to double the industrial plant in order to compensate for a failing China? or do we use it to stretch the investment dollar further now that the baby boomer money is no longer available and allow our financial system to be more efficient? Or do we use it for national defense and cryptography? You know, these are top-level issues. And we're probably only going to have enough chips to do one of the four.

So I would argue that the most consequential decision that the next American president is going to have to make is about where to focus what few chips we can produce and where do you put them? There's no right answer, there's no wrong answer. There's just less than satisfactory answers. And that leads us with the power question. Assuming that we could make GPUs at a stale that will allow mass adoption of AI,

which we probably can't anyway, you're talking about doubling the power requirements of what is used in the data space. Here's a thing, though, if we can't make the GPUs and we're not going to be able to make the more advanced chips anytime soon, we're still going to want to get some of the benefits from AI. So we're going to use older, dumber chips that generate a lot more heat per computation

in order to compensate, which means we're probably going to be seeing these estimates for power demand, not simply double, but triple or more, at the same time we get less computations, fewer computations, and generate an AI system that's actually less effect because we're not going to be able to make the chips at scale. So is it coming?

Yeah. But in the short term, it's not going to be nearly as fast. It's going to cost a lot more. It's going to require a lot more electricity. And we're probably going to have to wait until about 2040 before we can design and build in mass and apply the chips that we actually want to be able to do this for real. So, believe it or not, I actually see this as a borderline good thing

because it's so rare in the United States that we discuss the outcome of technological evolution before it's completely overwhelmed us. Here, I'd argue we've got another 15 years to figure out the fine print.