So what was one of your first clues that Monday was going to be an interesting day?
Yeah, it's funny. I was actually commuting in from Long Island on Monday morning. I was on the Long Island railroad and I looked at my phone. I was checking my emails, obviously checking the markets. That's my colleague, Gunjan Banerjee. She hosts WSJ's Take on the Week podcast and has covered markets and investing for the journal for years. And I see that futures contracts tied to the NASDAQ composite were down more than 4%.
which was really an eye popping move. We had not seen a move of that magnitude in quite some time. Something was going down in the stock market.
And by the time Gunjan made it to the office, things had gotten even wilder. I get to my desk and, you know, around the time the market opens at 9.30, I think everyone is kind of glued to their screens at that point and they see that this really ugly day for the stock market is beginning. All three major indexes are down a bunch. NVIDIA, of course, is down double digits.
Nvidia, the AI chip maker. Its stock was tanking. We could be looking at the biggest drop in market cap on record here for Nvidia when you take a look at the right across the screen specifically the NASDAQ 100 futures and the
So NVIDIA was down more than 10% shortly after the opening bell. It ended the day down 17%. And just to put that into context, that is a market value loss of almost $600 billion. So much money.
Right, that is just an insane amount of wealth, an amount of value that evaporated within hours. In fact, it is the biggest one-day market value drop on record. By the end of Monday, about a trillion dollars of value had been wiped from the stock market. But while the drop was historic, Gunjan also had a pretty good idea of why it was happening.
So what traders, people on Wall Street, Silicon Valley was pointing to was this upstart artificial intelligence company DeepSeek. DeepSeek. It's an AI company out of China. And over the last few days, its chatbot has been blowing people away.
Experts say DeepSeek's AI is just as capable, maybe even more capable, than leading AI chatbots like ChatGPT. But its creators claim it was made for much less money. And that's set off a major shakeup in Silicon Valley and on Wall Street. DeepSeek, this new artificial intelligence competitor, forced everyone
to take a look at their portfolios, take a look at their AI products, and really rethink who the winners and losers of this artificial intelligence trade were going to be. All of a sudden investors were going, hey, are the stocks that we own? Is Nvidia? Is it worth what we think it's worth? Welcome to the journal, our show about money, business, and power. I'm Jessica Mendoza. It's Wednesday, January 29th.
Coming up on the show, how deep-seek sank the stock market. So, what is deep-seek?
DeepSeq is an AI chat bot. If you've tried chat GPT, it's just like that. You go to the website, you log in, and you ask it a question, and it'll give you an answer as if a pretty smart human were answering it. That's my colleague, Stu Wu. He covers tech in Asia. And how did you first hear about DeepSeq?
I was doing this video interview with somebody in San Francisco who was the founder of an AI company. And we were talking about something else and he didn't know something. So he shared his screen with me and said, let me look it up. And what I thought was weird was that he didn't go to Google or chat GPT. He went to something I never heard of, deep seek. And what he said was that he'd been playing with it for the past couple of days. And he and his coworkers were just talking about how it was amazing and probably just as good as all the American competitors that he's been looking at.
So tell us a little bit about the company itself and the AI model. Who made it? So it's the brainchild of a Chinese guy named Liang Wenfeng. He co-founded this hedge fund in China. It's based in Hangzhou, which is also the same tech hub where the Chinese company Alibaba is based. So deep-seek grew out of that. So Liang's a pretty smart guy. He studied AI at one of China's top engineering programs.
And what I thought was really interesting about the company was that it had this really unusual hiring practice. Liang wants creative people, but he doesn't really care that much about experience. And he says his hiring principle is hire people with least amount of experience because his idea is that if you ask someone with work experience to solve a problem,
they're going to say, well, we should solve it like this because this is how I've done it in the past. But if you ask people without experience to solve that same problem, they'll have to sit down, think about the problem, and then they'll figure out the best and freshest and most efficient way to do it. So that's why a lot of people who work at DeepSeek are either fresh graduates or people with just a year or two of work experience.
And so this sort of takes us to the AI model that approach kind of created. So before DeepSeek, what was sort of the going assumption about how you make a cutting edge AI model? Yeah, so the conventional thinking was that if you wanted to make a world class AI chatbot or AI system, you needed a lot of the world's best AI chips that are super expensive as well.
In the US, AI development has been dominated by a handful of big tech companies who've trained their AI models using tons of top-line AI chips. Those chips are largely made by, you guessed it, NVIDIA. The assumption was if you didn't have enough of the right kind of chips, you couldn't build a world-class AI model.
At the earlier assumption was that a Chinese company could never do that because the US government had set these restrictions on what kind of chips US companies could sell to China. The thinking was that China would never catch up. So let's take a look at those assumptions. The first of those assumptions is that, like you said, you need a lot of chips to create these high-powered AI models. How did deep-sea sort of undermine that assumption?
Yeah, so DeepSeek released this research paper and explained how it did what it did. And it said that it spent a fraction of the money developing its advanced chat bot. And it did so using less advanced chips. So how can we understand that? So I think a good analogy is that let's look at the first chat GPT that many of us have used. And let's try to understand how that was trained.
So imagine that chat GPT is like a librarian that's read all the books in the library. And when you ask it a question, it'll give you an answer because it's read that book. But the problem is that to read all those books, that requires a lot of time and a lot of electricity for those computer chips to read those books.
So DeepSeek didn't have those resources, so it tried a new approach. So imagine you're still in the library, and DeepSeek is a librarian, but it hasn't read all those books. What it does instead is that it's focused on being really good at figuring out what book has the answer after you ask it the question. And it turns out that's just as effective as what chat GPT originally did. It was just as good, but it used a fraction of the resources.
It makes me think a little bit about kind of expert versus journalist in some ways. It's like what we do is we know who to ask and what questions to ask instead of actually like, you know, getting the PhD. We like go to the experts ourselves versus the expert who has to like learn everything about that subject. Yeah, you know, let's get an example. There's very few of us who can just read all those books and just maintain all that information in their head. And then when we, when we have to figure out it, we just kind of like stress out and call everybody we know and try to answer that question within an hour.
But DeepSeek does that in just a few seconds. And then the second assumption here is that a Chinese company couldn't do this because they wouldn't have access to the best chips, to NVIDIA's chips. First of all, when and why did the US start restricting the export of AI chips to China?
So the thinking during the Biden administration was that AI is gonna be really important for military purposes. So just imagine you can use it for developing a nuclear weapon or biological weapon or helping a general make a decision on the battlefield. It could give one side an absolute advantage. So that's why they decided we gotta stay a couple years ahead of China on AI. We can't lose an edge with AI on the battlefield.
In 2022, the Biden administration put restrictions on the kinds of chips US companies could sell to China. So what they said was that if you're a US company that wants to sell these chips to China, you have to restrict this parameter called interconnect bandwidth. And the analogy that I would use is that if you were designing a race car, this restriction would constrict how much gasoline ran through the fuel line. NVIDIA followed that rule, but it also figured out a workaround for its Chinese chips.
It complied with that fuel line, the fuel line was constricted, but it increased performance in other parts of the car engine to compensate for that, to make the most out of the fuel it did have. The result was that the chips and video was selling in China were more powerful than the US government would have preferred. The Biden administration eventually cut off that workaround, but it took about a year.
So they gave DeepSeek and other companies a year to buy these pretty powerful chips. And if you look at one of DeepSeek's research papers, it said it used about 2,000 of these powerful China-only chips from NVIDIA to train one of its advanced AI models. Last week, DeepSeek released its most advanced AI model yet, called R1. And what is the reaction bin?
Well, I can't remember anything quite like this. I mean, I think the closest thing is when chat GPT came out three years ago, and that kind of changed the world. Everybody was trying to write poems on it, you know, immediately. But this had some serious financial consequences, right? That financial fallout is after the break.
For the past few years, before deep-seat crashed onto the scene, investors had been piling into AI stocks, betting on big returns. They called it the AI trade. Here's Gunjan Banerjee again.
Basically, investors had latched onto this idea that artificial intelligence was going to unleash this wave of productivity in the economy among US workers and lead to gobs and gobs of profits for a handful of big technology companies, including NVIDIA. At its most recent peak, the company was worth more than $3 trillion. NVIDIA wasn't the only company people were betting on.
Who else did people think were the winners of AI? Really the big technology stocks, even think of like Meta Microsoft, which also has a competitor to chat GPT. People were thinking of some of these huge technology companies in the US as the key winners from the AI boom.
And when we're talking about sort of people piling on, how big did this get? The AI trade completely ate the stock market. It just took over almost every corner of financial markets that you can imagine. Like energy stocks, people bought up shares of energy and utility companies because training AI models uses a lot of power. Did it feel like a bubble?
It's interesting. There has been no shortage of investors the past few years saying we think this artificial intelligence trade is a bubble. And one of the reasons for that is just the amount of exuberance we've seen surrounding this trade and the levels of speculation
There was a lot of yoloing out there. You know, you only live one. Let's go for it. This was another flavor of like, let's get really, really rich from trading these AI stocks. Let's pile into their options, which are super risky and can provide kind of these boomer bus returns. Let's pile into really risky, exchange traded products.
There was just this mountain of speculation building and building and building while the AI craze continued. But it kept growing. It kept growing and growing and growing. It kept snowballing. And then came DeepSeek, a cutting-edge AI product that wasn't built by a US tech giant and seemingly didn't require a ton of chips. Investors who thought they knew who the winners of AI were suddenly weren't so sure.
I think the deep-seek news really spooked a lot of people about the valuation that they were assigning to some of these technology giants. It was a moment that made people question where they had been putting their money the past few years. And why were investors backing away from NVIDIA specifically?
The one thing that a lot of investors were fixated on is that it seemed like deep-seek needed a lot less computing power. So that would mean that the AI models of the future might not require as many high-end Nvidia chips as investors have been counting on. I mean, the way one investor put it to me was we've been banking on Nvidia being the disruptor. Are they being disrupted now?
In a statement Monday, NVIDIA praised DeepSeek's advancements. It added that serving up these kinds of AI models to users requires large numbers of its chips. Since Monday's chaos, the market seems to have stabilized. Tech stocks rebounded on Tuesday, with NVIDIA up 9%. But my colleague Stu Wu says that the AI industry is just beginning to wrestle with DeepSeek's model and its implications.
For example, there's still a lot of questions about how deep-seek pulled this off.
So DeepSeek published some research papers to explain how it accomplished, what it accomplished, but it hasn't revealed all of its secrets. So we don't know exactly what the training data it used. We don't know what that looks like. And there's a lot of people in Silicon Valley who are wondering aloud without evidence, I might add, but this is informed speculation that maybe DeepSeek actually had even more powerful NVIDIA chips than it's letting on. So there's still a lot to figure out.
DeepSeek disclosed some of its secrets, but not all of them. OpenAI, the maker of ChachiPT, has said it's looking into whether DeepSeek used large volumes of OpenAI data to help develop its model. DeepSeek didn't immediately respond to requests for comment. I'm curious, I mean, were all of those people also as surprised as the rest of us about this? I'm just trying to figure out, like, how did everyone, you know, Silicon Valley, those people in Wall Street, how did everyone miss this?
Yeah, so how did everybody miss this? Okay, that's a good question. So after Deepsea came out last week, a lot of prominent people in Silicon Valley, whether they're AI researchers or venture capitalists went on X or some other platform said this is really innovative, right? Like they just found a new way of doing this. And one of the guesses was that resource constraints breeds creativity, right? If you think about the book or the movie Moneyball,
How did the Oakland A's 20 years ago compete with the richest baseball teams despite having a fraction of the budgets? Well, they looked at undervalued strategies in baseball, and they figured out how to win despite this handicap. So that's one theory that resource constraints breeds creativity. Do you think part of the problem here was that people underestimated China?
I've been thinking a lot about this question. So I think a lot of people are in general surprised at how far China has come in technology. But in America, you don't actually get to see a lot of this because of effective bans on Chinese technology in America. A lot of Americans have never touched a Chinese cell phone made by Huawei or an electric car made by BYD, which is one of the world's biggest car companies, right? These things basically don't exist in America. So I think what happened was that when Deepsea came out,
anybody could download it, ask it a question in English, and see the answer in English. And they're like, wow, this wasn't supposed to happen. How did this happen?
That's all for today, Wednesday, January 29th. The journal is a co-production of Spotify and the Wall Street Journal. Additional reporting in this episode by Ace of Fitch, Rafael Huang, Karen Langley, and Sam Schechner. Thanks for listening. See you tomorrow.