The Peter Zeihan Podcast Series - AI Drama: DeepSeek, Nvidia, and China || Peter Zeihan

Episode Date: February 7, 2025

Join the Patreon here: https://www.patreon.com/PeterZeihanChinese firm, DeepSeek, claimed to develop an AI model at a fraction of the cost used to develop competing models. This has caused quite the stir, but what is actually going on?Full Newsletter: https://mailchi.mp/zeihan/ai-drama-deepseek-nvidia-and-china

Hey, all, Peter Zion here. It's almost question time, which is when Patreon members at the analyst level can submit their awkward, painful, pointed questions to me directly. So if you haven't gotten your question in, do it now. And if you haven't signed up for our analyst level, do that now. Because anyone who signs up this month for the next quarter, any money that you would have given me goes to a charity that we've identified, the Los Angeles Regional Food Bank. So, you know, get the tax deduction and help me, help them help you keep the ports open so you can still get your crap. Everybody, Peter Zine here coming from Colorado. Lots of people have asked me to comment about recent things that are going on with artificial intelligence from a market point of view.

Let me begin by saying I never comment on market movements and this is not going to change today. What is going on is we've had a bit of a route in the United States for a company called Navidia, which designs the ultra-high-end GPUs. Those are graphics processing units, which are the advanced chips that are used in almost all artificial intelligence. training and applications. Specifically, a Chinese company called Deep Seek says that it managed to generate an artificial intelligence model at 120th at the cost of what Navidia has been able to do, or what companies have been able to do with Navidia chips.

And in doing so, we basically saw more market value bleed off of Navidia in about eight hours than most countries have economies. It was ridiculous. It's also probably completely fabricated. keep in mind that almost anything that comes out of U.S. high tech is proprietary, so the information is circumspect, and almost anything that comes out of China has a degree of fraud involved. And if you're using either of these two things to color your trading decisions, I can't help you at all. But what I can do is talk about where this leaves the industry and what it would mean one way or another.

So let's assume for the moment that this is all fraud, in which case, this just all blows away really quickly. and a lot of people who have been trading money and trading stocks feel really stupid in the not-too-distance future. I think that's the most likely outcome. But let's assume for the moment that it's real. And this Chinese firm working with subpart chips has been able to make superior large-language model-trained AI systems to what the American bigwigs like OpenAI can do. 120th of the cost with Dumber Chips, that would be significant. One of the big problems that we have with the chips right now is the supply chain is so great,

grossly vulnerable, that it doesn't take much of that disruption anywhere in the world to break the ability of the world to make these high-end chips at all. And if that happens, then we're going to have to decide how to ration our chip usage. If you can drop the cost of training these new models by 95%, obviously, that makes that concern a little bit less of a problem. But even within that, there's going to be some very clear breaks geopolitically. So again, assuming that this is for real. Let's look at the chips that were used. The high-end chips that are basically three, four, and five nanometers that provide the backbone of most AI operations. The ones that we think could be used to retraining AI models are all under sanctions as of 2022. And the chips

that were used by Deepseek to do their operation were custom-built chips. Basically, Navidia looked at the sanctions, said that they still wanted a market in China. So they worked with the Chinese company Deepseek to design a chip that would comply with sanctions, and they dumped it down. If Deep Seek's announcements are true, clearly Navidia did not dumb it down enough, but the Biden administration, in one of its last actions, decided a couple months ago that Navidia has not been playing in good faith and basically banned all of these chips as well. So if Deep Seek was able to do this with the dumbed down chips, those dumbedown chips, those dumbed down chips are no longer accessible to the Chinese economy writ large anyway.

Moreover, the most advanced GPU that the Chinese system is capable of manufacturing is about a 14 nanometer. So you're talking about a dozen generations behind where we are right now. It's worse than it sounds because even to make those 14 nanometers, they have to use imported equipment and a lot of imported labor to do the quality checks. Yes, yes, yes. We have had a couple Chinese firms make something in the 7 nanometer range, but it's been really clunky. It's been a massive power hog. they get something only like a 20% success rate, and most commercial chips,

if they have anything less than 90% are considered failures. So if they push and push and push and push and push and push and throw an exorbitant amount of resources into it, they can make a chip that is still half a dozen generations behind what would be necessary to do an AI model. So if Deep Seek is not lying, then we would have a breakthrough in AI that the Chinese would be utterly unable to participate.

participate in. And that would generate a very interesting world. But we're going to know how this is going to go either way in the not too distant future, because everything that DeepSeek says that it did is open source. And we're going to be able to try to recreate it in any number of companies in any number of countries in the not too distant future. One other reason that I'm really doubtful that Deepseek actually did this. There are basically two ways you train AI. You basically you teach it like in school, like, you know, X plus two equals Y, three plus four equals seven. And once you've done that and built this kind of catalog of basic information, you then let it loose on bigger problems. What DeepSeek is saying is they just skipped the first step,

saying that you don't need to teach basic math to the computers. You could just set them loose on calculus. If that's true, which I find exceedingly unlikely, this can be replicated in a matter a weeks to months, not years. So we're going to know in the not too distant future just what actually happened and how many people in the financial world jumped the gun in a really silly way or not. And even in the worst case scenarios, the right term, even in the situation where Deepseek really did generate a bit of a revolution, everyone's going to be able to play except for the countries that can't get the chips.

And if there's one thing that we understand about the Trump administration versus the Biden administration is they have no problem taking sledgehammers to specific companies. So if it is the understanding of the Trump administration that Navidia has been playing fast and loosed with sanctions, you can count on Navidia having much bigger problems than simply a Chinese company coming up with a more efficient model.