Peak Prosperity - Chinese Firm Disrupts Big AI With DeepSeek

Episode Date: January 28, 2025

There has been something of a massive bubble in AI stocks of late. DeepSeek's new ChatGTP competitor may have just burst the bubble!Click here for part 2.Click here for Peak Financial Investing.

Thank you. to watch the video and to find other insightful content such as articles, discussion forums, and exclusive subscriber-only content. Hi everyone, I'm Dr. Chris Martinson of Peak Prosperity. We're going to be talking about the bursting bubble. If you haven't noticed, big giant stock market rout today. And of course, we've been expecting this for a while just because things have been frothy but there's been something of a massive bubble in ai stocks of late right so if you haven't noticed well i guess you've been busy doing something more productive with your life than tracking the latest fed inspired bubbles of course the federal reserve are your serial bubble blowers they've been blowing bubbles

ever since well the dot-com boom and bust the housing boom and bust and here they are they've done it again now what's a bubble a bubble is when an asset price rises beyond what incomes can sustain a corollary to that is that in a bubble things, you need just two things to have a bubble run. One, you need ample credit. That's where the Fed comes in. Two, you need a good story. Swampland in Florida, railroads at one time, tulip bulbs, eyeballs during the pets.com, dot com bomb craze, all of that other stuff but uh now it's ai all things listen i think this isn't a indictment of ai itself i think ai is going to absolutely transform our future in ways that we're only beginning to appreciate all that however people humans can get

a little ahead of themselves and this is one of those times and we now have a dot-com style super bubble in ai ai stocks and it just got burst by some chinese techies researchers okay so let's go there um i think that's one of the funnier memes that would be sam altman's face dubbed onto uh the theranos gal uh holmes at any rate nice combination there so let's take a look at this. Now, remember, it was just on the, was this the worst timed presser, press conference ever? Trump brought out Mahayoshi San and Larry Ellison and then Sam Altman right there, who looks like he's been cast for a movie where somebody from the future comes back to tell him he's done something he really shouldn't do. Something like that. They come out and they announce this huge thing where, oh, Project Stargate.

It's such a sci-fi, such a cool name. And these are all the headlines on January 23rd, Trump announcing it. And these guys coming out and saying, wow, we're going to create cancer mRNA vaccines and all this amazing stuff. I now understand why this presser had a very odd flavor to me, a very odd smell. And so we reported on it. At the time, we said, something's wrong with this.

Now I know what was wrong. They absolutely knew that their lunch was about to get eaten by somebody else. Billions, hundreds of billions of dollars have flowed into the US AI, big, giant, huge, centralized data centers. We we're gonna have to build nuclear power plants we're gonna need whole seas of forty thousand dollar nvidia gpus um so uh that was on the 23rd but then this weird thing happened um you know and of course we had larry ellison coming out and just giving us grifty snake oily stuff um and let's talk about theranos sam getting in there and explaining how this is just going to

change everything all right so what happened was uh a little known in fact i didn't know anything about it and i doubt very few people did but but these people should have that there was a chinese firm out there working on something called deep seekek, which is, oh, no, they released it. And look what happened when I wake up this morning. This is what a liquidity bubble bursting actually looks like. OK, sea of red there and all kinds of everything from soybeans to gold to crude oil to NASDAQ. Just kind of a sea of red right there now. Let's talk about this. So it was chat gpt remember chat gpt came out not

that long ago right you know from this guy right here pretty cool took the world by storm i actually use it i think it's okay they raised six billion dollars in capital they hired thousands of very bright people they set about building bigger and larger ai models, right? And each new iteration of those models was exponentially larger than the last, right? The computing power required was exponentially larger for a variety of, you know, complicated reasons. The chipset company, NVIDIA, they had a corporate moat built around its chips combined with its software and a superior means of linking all those thousands of units together into cohesive whole. And so, you know, what did we see here? Well,

NASDAQ took a big hit today so far and obviously just a scratch. I mean, that's just sort of bumping along. You wouldn't see that there's anything too bad with that at this point. Dialing in a little. This is looking at it just over the last month and here you can see yeah you know nothing really to talk about the nasdaq has been bouncing around here even with today's fairly steep decline not that big of a deal but as we're going to talk about later for my subscribers we're going to see that this is how bubble burstings begin they begin they look just like this but here's how we can tell that something really dramatic actually happened, right? It turns out that among the worst performing stocks are independent power producers. Here's Oklo, which is making a next gen nuclear plant, Bloom Energy down a stunning 28%.

And it was just six days ago down there. You see that? Bloom Energy sees Wall Street price target raised amid data center demand. Just six days ago, everybody's raising their targets. Today, boom, down a smoking 28%. What happened? And yeah, that was then, this is now. So it's not something, something pretty big went down. Constellation Energy Corp., another big firm out there, down a whopping 20%. So what happened is this. This DeepSeek new AI model came out. And all of a sudden, what do we have here? Well, we see that Microsoft hits a pause on its $3.3 billion with a B data center for open AI supercomputer. Oh, that happened.

And so, hey, this whole thing broke down. Now, there's a really, really great, very deep dive summary of this whole thing. I'll give you the link to that in a second. If you really like context, took me all morning to read through it. I do that so you don't have to. We're going to get to the summary. And then somebody summarized that for all of us. And here's a really good explainer of what this deep seek thing is and why it is so transformative and so brilliantly different. Morgan Brown writing here. And by the way, he was. I'll show you the actual article that he's pulling off of in a few slides this is a summary of that really deep dive one he says quote first some context right now training top ai models is insanely expensive and electricity was a big part of that cost open ai anthropic etc spent a hundred

million dollars just on compute so anthropic and OpenAI have two models out there, two AI models. ChatGPT coming from OpenAI. They need massive data centers with thousands of $40,000 GPUs. And it's like needing a whole power plant to run a factory. DeepSeek just showed up and said, well, what if we did all this for five million instead and they didn't just talk they actually did it their models match or beat chat gpt4 in cloud which is the um anthropic model on many tasks the ai world is as my teenagers say shook how they rethought everything from the ground up. Traditional AI is like writing every number with 32 decimal places.

DeepSeq was like, hey, what if we just use 8? Still accurate enough, it turns out. Boom, 75% less memory needed. That's a huge deal when you're talking about billions and billions, trillions of, you know, in-memory RAM uh ram units there so way beyond giga you need terabytes of this stuff um not gigabits terabits anyway quote four then there's their multi-token system normal ai reads like a first grader the cat sat deep seek reads in whole phrases at once 2x faster 90% is accurate and when you're processing billions of words this matters so it found a way to be almost as good but wicked faster and with a lot less requirements and so these

requirements mean you don't need those big $40,000 GPUs. In fact, you can run DeepSeek on little like 4090s and other GPU cards like that where in graphics cards, I mean, it can run on fairly stripped down things. Carrying on, he says here in five quote, but here's the really clever bit. They built an expert system instead of one massive AI trying to know everything, like having one person be a doctor, lawyer and an engineer. They have specialized experts that only wake up when needed. Traditional models, all 1.8 trillion parameters are active all the time. DeepSeek, 671 billion total, but only 37 billion active at once only it's like having a huge team but only calling in the experts you actually need for each task so this is very clever so what happened the

united states decided that nvidia's technology was going to be too sensitive or would give too much of a competitive damage or somehow china would use it or other countries might use it in ways we wouldn't want them to. So we limited exports. So in theory, China didn't have access to these H-100 units that NVIDIA was producing. I bet they could still get their hands on them. But at any rate, these engineers were given a challenge. And the challenge was, can you do this without needing all of this really expensive hardware and its computing requirements. So they started from the ground up and they built something that was really,

it's very clever, if not very ingenious. All right. So, and by the way, China had decided that financialization wasn't his best overall interest. That's, I think that's very smart. That's a whole other story. The United States has some of our best and brightest minds working at financialization, like writing better derivative this and trading programs and quant that and all this, just sort of gaming around for the fantasy digits that we call money. And instead, China said, you know, we don't need all that much financialization.

Let's not get clever there. Let's get clever in engineering. Let's put our engineers towards things like the DeepSeek project. And they are working on lots of things like this. Okay, so one of China's quant funds decided to deploy their talent, not in financialization efforts, but they deployed them at this and said, hey, go ahead, God, go ahead and start all over. See if you can build AI from the ground up.

All right. Now, constrained by that H100 chipset limitation, they tried all these other really clever approaches, right? And so we find out that necessity is indeed the mother of invention and in quite a startling fashion. So what were the results according according to Morgan Brand here? Mind-blowing. Training costs reduced from $100 million

down to $5 million. GPUs needed down from $100,000 to just $2,000. The API costs 95% cheaper, and they've modeled this out. Like if you want to, their API that you can use to get into their model is 95% cheaper than open AIs. Listen, people, at the end of the day, all of this has to be about making money.

As sexy as AI is and all the things it could do, at the end of the day, it's still about money in to money out. And if you're going to put huge amounts of capital into something, you're going to need to see lots of money coming back out of that at a profit. That was never really resolved how that was going to happen but of course that is the great feature of a bubble remember i said bubbles have two things ample credit and a good story the good story can sit there for a while tulip bulbs eyeballs railroads whatever the story is because people are imagining that all these fun yeah i just i

don't not clever enough to understand, Chris. I just know AI is going to spit out huge amounts of cash in the future. Well, it's got to do that. Well, your company can't spit out cash when a competitor comes along and finds a way to do it 95% cheaper. You're not going to be spitting out cash, right? So all those people standing with Trump right there are all just massive losers in this game anybody who's investing in that stargate thing i don't know what kind of hard conversations are having today but they're going to be pretty hard all right but wait he writes you might say

there must be a catch that's the wild part deep seek it's all open source anyone can check their work the code is public The technical papers explain everything. It's not magic. It's just incredibly clever engineering. Now, why does this matter? Because it breaks the model of only huge tech companies can play in AI. Turns out you don't need a billion dollar data center anymore. A few good GPUs might do it.

So here's the article, if you want to read it, by Jeffrey Emanuel, the short case for Nvidia Stock. It's a very long explainer, goes through tons of detail, more than frankly I could understand. Maybe you can get more out of it. But if you cut to the end, you cut to the chase, you find things like these gems all the way down deep in this article. And he writes here, quote, perhaps most devastating is deep seeks recent efficiency breakthrough, achieving comparable model performance in approximately one forty fifth the compute cost. They didn't shave 10 percent off of this one forty fifth, just a little over two percent of the compute cost. This suggests the entire industry has been massively over provisioning compute resources end quote i've been talking to you about these data centers and all the energy

and where's that going to come from running all these data centers they're popping in everywhere oh no we're going to need more cooling more data centers we're going to need more power because we have all these data so uh oh i guess we didn't need that after all. Turns out there was a giant transformative disruption that was bearing down on this particular industry. All right. Carrying on. Quote, combined with the emergence of more efficient inference architectures through chain of thought models, a lot of complexity there, but it's very cool, the aggregate demand for compute could be significantly lower than current projections assume.

The economics here are compelling. When DeepSeat can match chat GPT-4 level performance while charging 95% less for API calls, it suggests either NVIDIA's customers are burning cash unnecessarily, or margins must come down dramatically. End quote. And NVIDIA's story just got popped. Their moat turned out to be something of a muddy puddle so uh and by the way editorially open ai was named open ai because it was all going

to be open open source it was this was for the people and then oops sam altman took it all nope now it's all hidden and they hide it and a lot of it uh and the chinese came along and said dude we can totally do that here it's open truly actually open source they made it freely available to everybody um and by the way uh this is ridiculously cool what their model can actually do it apparently can self-check and self-improve and uses uh this 45 times less computational power, and it matches the very best, the largest, the most expensive US AI models. Okay, so this is kind of cool. I like how they did this because it's kind of like they put evolutionary life principles,

they coded that in, and wouldn't you know it, cool things emerged. So again, from this awesome article down here, quote, the technical breakthrough here was their novel approach to reward modeling. Rather than using complex neural reward models that can lead to reward hacking, you know, where the model finds a bogus way to get that reward that don't actually lead to better real-world model performance, they developed a clever rule-based system that combines accuracy rewards, which is verifying your final answers, with format rewards, which encourage structured thinking.

This simpler approach turned out to be more robust and scalable than the process-based reward models that others have tried. Now what's particularly fascinating, they say here, is that during training, they observed what they called an aha moment, a phase where the model spontaneously learned to revise its thinking process midstream when encountering uncertainty. This emergent behavior. I love that we have these emergent behaviors. It's kind of cool. A little scary, but also cool.

Wasn't explicitly programmed. It arose naturally from the interaction between the model and the reinforcement learning environment the model would literally stop itself flag potential issues in its reasoning and then restart with a different approach all without being explicitly trained to do this end quote so now we're on to something it's faster it's cheaper it's better that's pure disruption okay and i know that this is being downplayed by some people on twitter we're seeing people saying oh you know it's clever but it's not really new they're just building on top of our inventions and all this listen this is a level of cope um that i think is unwarranted here you know

given that none of this behavior right here is actually in any of the Silicon Valley, super expensive models, I'm going with it's clever and new at the same time, whole story there. I think that the East is now not saddling itself overly with certain limitations like dei and um green energy and uh you know various other delusions like that it just like best engineers what can we do and so i we're already seeing that china is very much advanced beyond the united states in terms of its development in infrastructure and other things so at any rate I'll leave more of that for later. Now, as if all this wasn't, like, scary enough, wasn't bad enough, they just announced today the same thing with an image generation AI tool called Janus Pro 7B,

which is besting OpenAI's D's doll e and stable diffusion uh here so the kobasi letter writing about this you can see here that um here circled in let me get my little laser pointer you can see here rectangled in red this is the performance of their janice and also their janice pro 7b janice is the light one 7 pro 7B is the darker one here. Having actually superior results, superior results in both accuracy percent and average performance, much, much better. So, and what does it do?

You can see here, if you said, give me the face of a beautiful girl, Janus sort of clunked along at that. Janus Pro 7B, doing a lot better at that. A steaming cup of coffee on a wooden table is a prompt. Kind of clunky. Much, much better with the Janus Pro 7B. Same thing for a red glass of wine on a reflective surface. This is much more sophisticated. How about this? A clear image of a blackboard with a clean, dark green surface and the word

hello written precisely and legibly at the center with bold white chalk letters no bueno as we've known ai's had a lot of trouble generating words not progena 7b that's got that one absolutely correct so at any rate um this is going to also not just in the ai part but the generative image side they've both of these things they've started to solve. So race is on, kind of interesting. Now, again, a bubble exists when asset prices rise beyond what incomes can sustain. And by that definition, we have a bubble in AI stocks. And it is not small. It is very, very large. And the shocks to the system, listen, they took a little while to develop. I've been posting about back at peakprosperity.com, been posting about this deep seek thing since last week, but it was only kind of over this weekend that I think, you know, general, more general

dawning awareness broke out. And of course, we've been talking about this in great depth and what the impacts would be on stock market portfolios. What does it mean when a bubble bursts? Because, hey, when a bubble does burst, it's very painful. I'm old enough to remember the last two. Nobody should live through more than one in their life. I've been through a dotcom bubble and a housing bubble, and I think I'm now going to see the AI bubble come to its final conclusion. And so when we think about this, listen, NVIDIA had a more than three and a half trillion dollar market capitalization that all rested on the belief that it had this impenetrable corporate moat. Nobody was going to touch their profit

margins. Hey, Hosanna forever and ever nothing but endless profits nobody would be able to beat them not for a long while if ever right because they had a practice nvidia and they still do pouring their earnings back into r&d okay that's going to keep them ahead of the pack forever but oops along comes a chinese team and the mo again turns into something of a puddle right okay welcome to the turbulent world of technology. It's nothing if not a long, unbroken tale of disruptors disrupting things. And while this tale fits that mold, we have to consider as well the extreme bullishness that had driven NVIDIA to a valuation at $3.5 trillion that was larger in valuation than the entire stock

market capitalizations of France and Germany combined. Now, listen, bigger AI models, we're going to need more chipsets, more data centers, more power, more, more, more. And then all of a sudden, it's been turned on its head, right? We saw that Microsoft pulled the $3.3 billion data center. We saw that the energy company has already taken this huge hit. Now, Grok at AI, the AI at X tells me that the cumulative market cap of the leading AI companies is around $16 trillion. Now, what if that gets cut in half, which would be about right for a bursting bubble?

That $8 trillion loss would match the total market cap meltdown of the great financial crisis back of 2008 and 2009. Now, could another such great financial crisis be brewing here? Short answer, yes. And that brings us to part two, where we're going to have to talk about what happens when your story gets shredded. And that's what just happened. The story absolutely got shredded, right? And so we're going to have to discuss that and discuss the features of bursting bubbles, what this means for the incoming administration's plans. And

listen, the short version is bursting bubbles they really suck okay especially for the unprepared now if your wealth isn't adhering is not adhering to a risk managed strategy then i wish you the very best of luck with everything bubbles make mincemeat out of passive strategies they always have they always will know this time will not be different but we're going to have to talk about this now. And for that, I would invite everybody to come back with me to peakprosperity.com, where we will continue this conversation in more depth for my subscribers. In the meantime, thank you for listening. And please pay attention to this.

And guess what? Bubble burstings take time. I think we've just entered a bubble burst. With that, thank you very much. Be prepared, be resilient, and we'll see you next time. Thank you.