How DeepSeek changed the market's mind

两人或三人, 红色蓝色为主要的角色，剩下的颜色就互相兼一下吧。

(SOUNDBITE OF COIN SPINNING)

MALONE： Imagine us at PLANET MONEY rolling out of bed late on Monday morning, still in our cozy pajamas.

CHILDS： Yeah. We got our Teenage Mutant Ninja Turtle slippers on. We got our cup of coffee.

MALONE： Big yawn, turn on the old TV, and oh, my - what is happening in the stock market right now?

UNIDENTIFIED PERSON #1： This is moving so fast. It's stunning.

CHILDS： I mean, wow.

UNIDENTIFIED PERSON #2： It is shaking this entire industry to its core.

JIM CRAMER： Crushed the Nasdaq, which plunged 3.07%.

MALONE： Yeah.

MALONE and CHILDS：(Vocalizing).

CRAMER： A single largest loss in a day of market capitalization in history.

CHILDS： What was happening? It was apparently some kind of AI apocalypse.

UNIDENTIFIED PERSON #4： OK. So AI apocalypse. Not so sure about that.

CHILDS： OK, fine, whatever.

UNIDENTIFIED PERSON #4： But without a doubt, this is a monumental shift.

MALONE： Yeah. Because on Monday, AI-related stocks started plummeting and TV-related people started grasping for big metaphors.

UNIDENTIFIED PERSON #5： It was an earthquake today in the world of artificial intelligence.

MALONE： The seismic AI event? A new-ish AI model from a company called DeepSeek.

1'20" (SOUNDBITE OF MUSIC) 音乐起一小会儿入

MARY CHILDS： Hello, and welcome to PLANET MONEY. I'm Mary Childs.

KENNY MALONE： And I'm Kenny Malone. Today on the show, call it an artificial intelligence earthquake, call it an AI apocalypse, but Monday was not just a market freak out. I mean, it was that markets lost hundreds of billions of dollars but it was also a teachable moment.

MARY CHILDS： Oh, yes. Because if we look at the specific companies that got slammed on Monday and the few that benefited, we can see pretty clearly what people with money are betting our AI future looks like. And this week, the AI model from DeepSeek has them betting on a very different-looking future.

KENNY MALONE： Or as Jim Cramer so artfully put it...

另一个CV： Scream...

KENNY MALONE： This week has been a tectonic shift in assumptions about how the world is going to look. So let us first discuss how those assumptions became assumed. We shall visit a simpler, ancient time.

MARY CHILDS： Yes. Two years ago, roughly November 2022. This is when the world got its first look at ChatGPT. You will recall we all lost our minds. ChatGPT could write poetry. It could tell stories. Maybe it could take our jobs. We'd never seen anything like it.

KENNY MALONE： Now, that AI model was developed by an American company called OpenAI, and their AI model, ChatGPT, had taken a ton of time to develop. OpenAI had spent billions of dollars creating it. And as the model developed, it became clear that running better and better versions of GPT would be so expensive because it required the best semiconductors in the world, lots of them.

MARY CHILDS： The American AI arms race began. Google, Meta, Microsoft, all built giant, expensive AI models.

KENNY MALONE： And newer companies got more competitive, too, Anthropic, Perplexity also with gigantic AI models requiring unearthly amounts of compute, as they say, and money, as they also say.

MARY CHILDS： They do say that. And what seemed to be true in all these cases was that in order to compete in the AI revolution, these companies needed unimaginable scale, more and more computing power, more and more investment - billions and billions of dollars.

KENNY MALONE： If there was a way to win the AI arms race, it seemed pretty clear you needed the scale of a gargantuan company to do so.

MARY CHILDS： And then on Monday, all of those assumptions fell apart, as did the stock market.

KENNY MALONE： Was Monday a bummer of a day for you? In the grand scheme of days, how's that shape up?

ZINO： Yeah. I mean, listen, as far as kind of Monday morning is concerned, it starts off on a sour note.

MARY CHILDS： Angelo Zino is an equities analyst at a company called CFRA Research. Angelo's job, in part, is to look at the tech world and identify good and bad stocks for investors.

KENNY MALONE： And, he says, on Monday, there were bad signs even before the stock market opened in the United States.

ZINO： I was up at 4, 5 a.m., which is when I typically wake up, and I already had a number of inquiries in my inbox from investors out in Europe. So...

KENNY MALONE： And what are those investors - like, Angelo, you told us American AI was the future.

ZINO： Yeah. I mean, listen, you know, who are the winners? Who are the losers from this?

KENNY MALONE： Yeah.

ZINO： What exactly is happening?

MARY CHILDS： Great questions, European investors. What exactly was happening?

KENNY MALONE： Well, so, yeah, DeepSeek was happening. Here's the backstory. This Chinese company, a subsidiary of a hedge fund, actually, had been developing an AI model, just, you know, for fun, for its own hedge-fundy uses, I guess. And this was not a secret. Lots of people in the AI tech world knew about this, Angelo included, because the hedge fund had been sort of open sourcing what it was doing. After all, the parent company was not an AI company. It was a hedge fund.

MARY CHILDS： Right. So, people generally knew that this AI model was likely more useful than just for hedge-fundy things, but what seems to have happened, what seems to really have rocked the stocks, were a few key things.

KENNY MALONE： Yeah. No. 1, the DeepSeek AI had been training, getting better and better. And it seems that the newest version, released just 11 days ago, had got real good. It hit certain benchmarks that showed it was possibly, allegedly, as good, or nearly as good as, the gigantic, fancy, expensive AI models being built by the American AI companies.

MARY CHILDS： Except and here's the big thing No. 2 DeepSeek is not a big, fancy, expensive AI model. It was reportedly built for a fraction of the cost and reportedly did not need top-of-the-line chips and semiconductors and processors to run, like the models from the American AI companies need.

KENNY MALONE： And then big thing No. 3, according to Angelo Zino, news of all of this starts to spread. And over the weekend last weekend lots of people download a DeepSeek app, presumably to see what this buzzy new AI model is really like.

ZINO： DeepSeek topped the app store chart and kind of got ahead of OpenAI. I think it kind of, you know, put the technology right in the eye of the storm for investors out there.

MARY CHILDS： And then, five hours after Angelo wakes up on Monday, markets open, and whammo, a bunch of stocks start plummeting.

KENNY MALONE： I've been thinking about, like, should Monday have a name like Black Friday had a name, you know? And I've been trying to make this one work.

ZINO： Yeah.

KENNY MALONE： Mond-AI (ph) Apocalypse.

ZINO： It's not bad. It's not bad.

KENNY MALONE： Hey. You know...And so we shall now dissect and make meaning of the Mond-AI Apocalypse, starting with Angelo's specialty, the tech sector.

MARY CHILDS： So which tech stocks had an awful Mond-AI?

ZINO： So when you kind of look at some of the names that got hit the most, I mean, essentially, chipmakers that are heavily exposed to the data center market. And, you know, that would include Broadcom, Marvell, Micron.

KENNY MALONE： Is it Mar-vell (ph)? I've been saying Mar-vull (ph) this whole time.

ZINO： No...

KENNY MALONE： Like a nerd who reads comics. OK.

ZINO： Yeah.

KENNY MALONE： Whoops.

ZINO： So Marvell, specifically in Nvidia was the one that got the most attention out there.

MARY CHILDS： Ah, yes, Nvidia. Nvidia manufactures top-of-the-line processors that have become the not-very-secret sauce that American-made AI models need in order to do the unfathomably large amounts of computing required to train and run AI models. If AI is the gold, Nvidia is selling the picks and the shovels.

KENNY MALONE： So for a lot of the people who were interested in investing in the brave new AI future, Nvidia seemed like a good place to do it, especially because it has actually been quite hard to invest directly in the AI companies. Like, some of the biggest companies developing the models OpenAI, Anthropic they do not have shares you can just go and buy. They're not publicly listed - not yet, at least.

MARY CHILDS： All of this is why Nvidia, seemingly overnight, has become one of the most valuable companies on the planet. In 2020, you could buy Nvidia's stock for, like, six bucks. Last week, 142 bucks.

KENNY MALONE： That is, like, 23X growth, basically, because the only way the AI revolution can happen is with the fancy AI chips from Nvidia. In fact, Nvidia was seemingly so important that in 2022, the United States banned Nvidia's most powerful chips from being sent to China to preserve America's AI advantage for national security reasons.

MARY CHILDS： You could not overstate Nvidia's value. And then enter DeepSeek, this Chinese hedge fund, apparently building a top-of-the-line AI model, even though they weren't allowed to build it on the very best AI chips from Nvidia because Chinese companies aren't allowed to buy them.

KENNY MALONE： Which the markets seemed to think perhaps meant Nvidia was not quite as important to the AI revolution as they thought. And on Monday, Nvidia's stock price fell so much - nearly 20%.

MARY CHILDS： A near-double decimation, if you will. The company's value dropped by almost $600 billion, the single biggest one-day drop in American history.

KENNY MALONE： Now, historically, had you been bullish on Nvidia? Had you been telling these European investors like, hey, go long Nvidia?

ZINO： So listen, we went bullish on Nvidia actually in March of 2020. Yes, we have, you know, continued to pound the table on Nvidia. As recently as kind of a week or two ago, we do think they have the most important intellectual property probably in the world.

KENNY MALONE： Right. But if you were a week ago telling people bullish on Nvidia and then on Monday it plummets, how does that - like, what's that like when you're in your chair?

ZINO： It doesn't look good.

KENNY MALONE： (Laughter).

ZINO： And it's definitely not an easy conversation to have.

KENNY MALONE： It is, however, a conversation that many investors needed to have with themselves on Monday because DeepSeek's prevalence suddenly did seem to undermine the core assumption that in order to build god-tier AI, you needed god-tier AI chips. But DeepSeek had apparently pulled it off with cheaper chips and fewer of them.

MARY CHILDS： Which the markets interpreted as not good for Nvidia and the other chipmakers in their future.

KENNY MALONE： As for Angelo, what did he tell his angry European investors and other investors during the Mond-AI Apocalypse?

ZINO： We didn't say sell your Nvidia shares. We continue to have a buy recommendation on the shares.

KENNY MALONE： OK, so you didn't change that through the cliff.

ZINO： Right, especially on that day. I mean, when the stock was down I believe it was 18 or 19%, we did think that was an overreaction.

KENNY MALONE： Because, he says, the AI revolution will still need lots and lots of processing power. Just, you know, whose chips and how many and what kind? Well, the markets seem a bit less sure about all of that than they were one week ago.

MARY CHILDS： For our next lesson learned from the Mond-AI Apocalypse, we turn to Jennifer Hiller of The Wall Street Journal.

KENNY MALONE： I'm going to share my screen, and I'm just going to...

HILLER： OK.

KENNY MALONE： I'll explain in one second, all right? Let me...

HILLER： OK.

MARY CHILDS： Jennifer has been reporting on the energy industry for over a decade.

KENNY MALONE： I just want you to read a headline that you wrote from, like, about two weeks ago.

HILLER： Yeah. "Once Unwanted, Constellation Energy Is One of the Hottest Stocks."

MARY CHILDS： "Once Unwanted, Constellation Energy Is One of the Hottest Stocks." It's a story about how investors were pouring money into America's biggest provider of nuclear power. The value had been shooting up and Constellation Energy hit an all-time high stock price just last week.

KENNY MALONE： Well, Jennifer, it seems you know what I'm going to do next, which is two weeks later I'm just going to pull up a graph of their stock price.

HILLER： It looks like it sort of fell off of a cliff and then bumped along the bottom, and then dipped some more.

KENNY MALONE： Yeah.

HILLER： I'm obviously a big jinx. You don't want me to write a story about your all-time high.

MARY CHILDS： But this was not Jennifer's fault. This was, of course, DeepSeek's fault.

KENNY MALONE： Yes. OK. So in case you have not heard this, the AI revolution is going to require a lot of energy. And this goes back to the market assumption we just discussed, about how training AI models and running them requires really high-tech processors, which use loads of electricity. And then AI uses loads of those fancy processors using loads of electricity, and they put them all together, I guess, in giant big buildings.

HILLER： You know, really large data centers that are kind of often on the edge of town...

KENNY MALONE： Uh-huh.

HILLER： ...In great big buildings.

KENNY MALONE： Should we imagine it like having the electrical meter outside, and it's just spinning so fast you can't see the hand?

HILLER： (Laughter).

KENNY MALONE： ...Like (vocalizing).

HILLER： I like that idea. I don't know actually how they're metered, but it must be some very fancy version of that.

MARY CHILDS： Jennifer says that people have been talking about needing and building, like, Manhattan's worth of new power supply, as in enough energy to power Manhattan three, four, seven times over.

KENNY MALONE： And for reasons. Jennifer says some of the tech companies have become fixated on nuclear as a great option for the huge AI power needs.

HILLER： All of these big tech companies have climate commitments and climate targets that they've made.

MARY CHILDS： Yep. Nuclear works on that front. It is carbon emission-free.

HILLER： AI and data centers need power consistently, all the time, 24/7.

KENNY MALONE： Apparently, that is how nuclear power works - consistent energy, 24/7. And sure, you know, it has a notorious track record, but...

HILLER： I think tech people just also kind of like the technology of nuclear.

MARY CHILDS： Nuclear is the tech bro of power. That makes total sense to me.

KENNY MALONE： Oh, hundred percent. But anyway, the point is that we have a similar story here to what we had with Nvidia and other chipmakers. People wanted a way to invest in the AI future, and so they were pouring money into nuclear stocks, including, of course, our nation's biggest nuclear provider, Constellation Energy.

MARY CHILDS： Over the last three years, Constellation went from, like, 40 bucks a share to, like, $300 because the markets thought our AI future needs all the nuclear energy.

KENNY MALONE： And then on Monday, because of DeepSeek, the markets were forced to perhaps rethink that assumption. You had this high-quality AI apparently needing less energy. The market starts selling Constellation off of the cliff, and at one point on Monday, the stock was down 20%.

HILLER： It's brought up this question of, how much power does the AI industry really need?

KENNY MALONE： Yeah. So that cliff you described the stock price of Constellation dropping off that is a market collectively saying, oh, crap, maybe the future doesn't require as much electricity as I was betting on.

HILLER： Right. I think it's pretty safe to say that there's a future where we're using a lot more electricity than we use now, but are we...

KENNY MALONE： Yeah.

HILLER： ...Using it at this, you know, extremely higher level?

KENNY MALONE： And to be clear, nuclear power has been getting lots of headlines, but the markets had also been pouring into really, like, any company that makes and sells any kind of electricity.

MARY CHILDS： And those companies, they got hit on Monday, too, including a company called the Texas Pacific Land Corporation, which is basically just some giant chunks of land in Texas that have oil and natural gas. Even that got whacked by the DeepSeek news on Monday.

HILLER： So, yeah, that shows you, like, how the tentacles (laughter) of this stretch out.

KENNY MALONE： The Mond-AI-apocalypse (ph) was not about whether or not there will be an AI revolution. If anything, the introduction of DeepSeek means more AI, lowering the barrier to AI, making it cheaper to use for- I don't know - whatever your AI mind can dream up.

MARY CHILDS： All your great ideAIs (ph).

KENNY MALONE： There you go.

MARY CHILDS： (Laughter) Getting into it. And there are two categories that you'll see companies sorted into AI enablers and AI adopters. Enablers are the companies that are basically the supply chain to make AI models chipmakers, power companies, the AI model companies themselves.

KENNY MALONE： And then the AI adopters, those are all the companies that stand to benefit from using AI. And really, what this week was about was, like, a shift away from the enablers and arguably a bit towards the adopters.

MARY CHILDS： Yeah, the shift away from the chipmakers and the energy companies was dramatic. You have to look a little bit harder, squint a little bit to see the market moves towards the adopters. But one example people point towards - Salesforce. They have basically made their whole thing, their identity adopting and using AI. And on Monday, their stock was up 4%.

KENNY MALONE： And I think there is one other huge assumption that was challenged this week. Up until Monday, the markets seemed to be confident that American AI companies had a moat around this technology, that, like, the barriers to entry were just so enormous that no one else was going to win this arms race.

MARY CHILDS： But that moat - that was maybe the biggest assumption that the markets were scrambling to rethink on Monday 'cause if a Chinese hedge fund that doesn't even make AI for a living was able to make DeepSeek as cheaply as they say, using fewer and less fancy processors, and if it's even close to as good as these American AI models, yeah, that probably does change everything.

KENNY MALONE： After the break, we sit down with someone actively trying to build DeepSeek from scratch to see how much of all of this is real, and did the world really change on Monday?

19'17" (SOUNDBITE OF WILLIAM HUTCHISON AND JACOB BACKER'S "WITH IT")

MARY CHILDS： There is a company called Hugging Face. Their logo is, like, that smiley face emoji that's also giving you a hug with those two big emoji hands. Hugging Face, you know?

KENNY MALONE： I can see that it is a very cute logo. But it is a kind of AI company where Leandro von Werra works, and he describes the basic business model this way.

VON WERRA： So you can imagine it a little bit like GitHub, if you're familiar with GitHub...

KENNY MALONE： Yeah.

VON WERRA： ...Where people share code and everything's free.

KENNY MALONE： But there's an enterprise edition that costs money, and that's how they make money. But, like, the point is, Hugging Face, cute logo, like an AI sharing platform - they do not build gigantic proprietary AI models to compete with OpenAI or Anthropic or Google or whatever.

MARY CHILDS： And the reason we got in touch with Leandro is that he heads up their research team.

VON WERRA： So we - our job is not to make money. Our job is mostly to...

KENNY MALONE： To spend money.

VON WERRA： ...To spend money and build things that are very useful.

KENNY MALONE： And what's been useful lately is DeepSeek or, you know, playing around with DeepSeek's new chatbot that partially freaked the markets out about the future of AI.

MARY CHILDS： Because there are really two reasons why the market freaked out - first, that it was made in a way that was cheaper and more efficient than how things like ChatGPT were made. And the other reason was that DeepSeek's model was allegedly really good. So the big question hovering over this entire week has been, is all of that real and true, or were markets overreacting?

KENNY MALONE： So let's take these one by one. Is DeepSeek actually as good as the fancy American AIs? Well, Leandro says, we have ways to test this. There are these, like, standardized tests, benchmarks, for AI models. They used to be, like, pretty simple math problems or whatever, but as the models have been trained more and more and have gotten better and better...

VON WERRA： We've upped the exams a little bit, so now we're closer to, like, Ph.D.-level exams. And we can measure quite well right? how many of the questions does the model get right?

KENNY MALONE： Is DeepSeek passing Ph.D.-level coding, Ph.D.-level math?

VON WERRA：Yeah. So those models are getting, like, really good at solving certain kinds of questions. For example, these models can solve some of, for example, Math Olympiad questions.

MARY CHILDS： And here, I will just interject to note that we do have on staff one person who has a math degree. And it's Kenny.

KENNY MALONE： True, not an Olympiad, but I was excited. Quickly downloaded some math Olympiad questions. Pulled them up on my screen for Leandro.

MARY CHILDS： This is such a big day for you, I feel like.

KENNY MALONE： Oh, yeah.

MARY CHILDS： This never gets to happen.

KENNY MALONE： The first all right, hold on show that for each n, we can find an n-digit number with all its digits odd which is divisible by five to the nth power.

VON WERRA： Yeah.

KENNY MALONE： DeepSeek can do that?

VON WERRA： Sometimes.

KENNY MALONE： I mean, I can only sometimes do that, so yeah. All right. Fair enough.

VON WERRA： Exactly. I also - I'm, like, a physicist by training, and it takes exercise to be good at those questions.

KENNY MALONE： Yeah. OK. So that's what we're talking about here, huh?

VON WERRA： Yeah. Capability-wise, we don't see any benchmarks that show that they have, like, some gaps in the knowledge.

KENNY MALONE： Yeah, no apparent gaps between how DeepSeek's model performs and how the other models perform.

MARY CHILDS： Leandro has also checked to make sure DeepSeek is getting its exam answers legitimately.

VON WERRA： So we test these models on kind of exams. If those exams are already in the training data, naturally, the models are much better.

KENNY MALONE： OK, so this is the classic, are they teaching to the test? - like that's the...

VON WERRA： Exactly.

KENNY MALONE： Yeah.

VON WERRA： Yeah. And we haven't seen any indication of that either.

MARY CHILDS： From what he's seeing, DeepSeek does seem to be in the same tier as the fancy American AI models.

KENNY MALONE： So OK, appears to be good. That answers everybody's first question about DeepSeek, whether it was playing at the same level as other big AI models. But the second question is, do we really think that this thing is more efficient in some way?

MARY CHILDS： And to test that, Leandro and his team are, in fact, attempting to build this themselves, basically from scratch, to replicate it.

KENNY MALONE： When you sit down to, like, replicate DeepSeek, I don't even know like, what do do you sit down and you open up a computer and you're like, all right, crack your fingers, open up a Microsoft Word document - you're like, DeepSeek v2, let's go.

VON WERRA： Yeah. I mean, that's pretty much what we did, so...

KENNY MALONE： Now, the reason this is even possible is because unlike a lot of the recent American AI models, DeepSeek has been pretty open about their methods. They actually put out a big report that was kind of a set of instructions for how the model was built.

VON WERRA： So we're not, like, reverse engineering in the dark. We're actually more, like, following their recipe and translating their paper to code. And I think we're making good progress. So I think in a few weeks, at the latest, we're going to have a pipeline that works that people can use, and we're going to see if we get the same numbers.

MARY CHILDS： Yeah, those numbers. Again, DeepSeek's latest version was reportedly much cheaper to train and much cheaper to run than the big American models.

KENNY MALONE： Are the claims that have been made about DeepSeek the cheapness, the fact that it can run on less powerful processors do all of these things seem to be checking out?

VON WERRA： Yeah. So I think that's something that we want to investigate a bit. So far, it seems like napkin calculation it's probably the right order of magnitude.

KENNY MALONE： Yeah, in the ballpark, which is notable because there had kind of been some murmurs of skepticism around the specific numbers DeepSeek was putting out.

MARY CHILDS： But Leandro is pretty convinced, so far, it really is way cheaper than the existing American models for basically the same thing.

VON WERRA： And I think one thing that people underappreciate is an open model is kind of a leveling - levels the whole field because everybody has access to the same level of knowledge, so everybody can immediately build on top of that.

KENNY MALONE： I've heard people talk about this moment as a shift towards AI models as a commodity. And that is a completely different vision than what markets seemed to be betting on before this week. Like, seemingly overnight, we went from an imagined future where a handful of gigantic American companies controlled the most powerful AI models to a future where it seems very powerful AI models can be built and used by maybe anyone anywhere, someday.

MARY CHILDS： That is a lot to process in one week, in just a few days. The markets for their parts have moved ever so slightly back towards where they were before Monday. Still shocked, but, you know, Nvidia is still one of the most valuable companies in the world.

KENNY MALONE： Yeah. Investors are, of course, still betting on a version of the AI revolution, which, of course, will be excitedly televised.

30'10" (SOUNDBITE OF MUSIC)

MALONE： If you want to nerd out more on what an AI future might look like, you can subscribe to our newsletter. Newsletter author Greg Rosalsky is working on a piece about why the AI community is suddenly obsessed with a 160-year-old paradox, Jevons paradox. He's got the history of that idea and the latest AI ideas. Subscribe at npr.org/planetmoneynewsletter.

CHILDS： This episode was produced by Willa Rubin with an assist from James Sneed. It was edited by Keith Romer and engineered by Neil Tevault. Research help from Sierra Juarez.

MALONE： Special thanks this week to Haim Israel from Bank of America. I'm Kenny Malone.

CHILDS： And I'm Mary Childs. This is NPR. Thanks for listening.