DeepSeek's AI Innovation Sparks $2 Trillion Market Correction and Reevaluation of AI Industry

When It Comes to Investing, Is A.I. Worth the Hype?

Jeff Sommer writes Strategies, a weekly column on markets, finance and the economy. A.I. chatbots are fun, sometimes even useful and, until recently, endowed with the uncanny ability to mesmerize investors and fuel the U.S. stock market. But the excellent performance of a new, relatively cheap artificial intelligence engine from a Chinese start-up, DeepSeek, has perturbed the market and complicated the A.I. story. Investors are re-evaluating prominent companies swept up in A.I. fever, including Nvidia, Meta, Alphabet, Microsoft, Amazon, Tesla and the private start-up OpenAI. The notion that full-blown superhuman intelligence is imminent has spurred the-sky-is-the-limit valuations, as well as concerns about the political and social risks posed by advanced intelligence. One immediate question: Is the main approach to developing A.I. in the United States -- pouring billions of dollars into chips and infrastructure -- worth the expenditure for all companies if similar results can be achieved far more cheaply? DeepSeek's lower-cost innovations add urgency to bigger, longstanding financial questions: How much are artificial intelligence companies really worth, and what will the broader economic value of A.I. ultimately be? Daren Acemoglu, a winner of the 2024 Nobel in economic science, gave me some answers. "There is a lot of hype in the industry," he told me in a telephone conversation. Yes, he said, A.I. companies have made some "impressive achievements," but he added that many financial and economic calculations were being based on mere "projections into the future that are sometimes exaggerated." Professor Acemoglu, an M.I.T. economist with an interest in the impact of technical innovations on global economics, is skeptical about the more fervent A.I. claims. He ranks A.I. as a significant advance, perhaps with a macroeconomic effect akin to the telephone, which was no small thing. But don't get carried away, he said, at least not yet. He doubts that full, advanced artificial general intelligence "that can do anything a human can do, but more," will be achieved. Therefore, over the next decade, he estimated, increased productivity from the diffusion of impressive, but limited, A.I. engines will increase the size of the U.S. economy by only about 1 percent, or roughly 0.1 percent a year. That doesn't seem like enough to count as a technological revolution in economic terms, I said. "Well, it's not trivial," Professor Acemoglu said, "but it's one or two orders of magnitude less" than A.I. bulls "would like you to hear." Of course, he added, if one or more companies achieve true, complete, artificial general intelligence within the next several years, then his estimates will turn out to be far too low. Relentlessly Upbeat It's earnings season on Wall Street, and over the last two weeks, some of the U.S. companies that are developing and investing heavily in A.I. have offered entirely positive -- and, frankly, self-serving -- estimates of the A.I. future. I read transcripts and listened to several conversations of top executives and analysts. The most extravagant narrative undoubtedly came from Elon Musk, Tesla's chief executive, who took time off from his government work to talk about his company's earnings. While conceding that Tesla will have a tough year meeting production and profit targets, he was ecstatic about its A.I. prospects. Already, Mr. Musk claimed, "there is no company in the world that is as good at real-world A.I. as Tesla." Once its cars are approved for "full self-driving" on the roads -- he promised that he wasn't "crying wolf" and that it would really happen this year, though he has been saying the same thing for many years -- the company's fleet will increase in value "10X," he said, thanks to A.I. Moreover, he said, Tesla will produce millions of A.I.-embedded Optimus robots in the not-too-distant future, creating a "path for Tesla being the most valuable company in the world by far." Mr. Musk elaborated: "There is a path where Tesla is worth more than the next top five companies combined." And, he added, "that is overwhelmingly due to autonomous vehicles and autonomous humanoid robots." Mr. Musk's assertions are not universally accepted. From Jan. 24, when DeepSeek's A.I. innovation began to roil the market, Tesla shares fell 7 percent through Thursday. Other A.I. companies fared worse. Shares of Nvidia, whose chips run much advanced A.I., dropped 12 percent. Aswath Damodaran, a New York University finance professor who has evaluated many tech companies, said DeepSeek's efficiency implied that fewer and less-advanced chips would be needed for many A.I. functions. As a result, he wrote recently, the market for Nvidia's high-end chips isn't likely to grow as rapidly as expected. So, he said, Nvidia shares will be worth less than anticipated, even after the recent price decline. In addition, shares of nuclear-powered electricity providers like the utilities Constellation and Vistra, which had soared in the expectation that A.I. data factories would need ever-increasing quantities of power, sank on reduced projections of the required electricity. Meta, Alphabet and Microsoft, which have invested billions in A.I. development, have had mixed performances since DeepSeek's arrival. Alphabet and Microsoft fell, while Meta rose modestly. The companies are complex and enormous, with different products and strategies, but the chiefs of all three said they would continue pouring vast sums into A.I. infrastructure in hopes of developing a competitive edge, while extending A.I. offerings throughout their consumer services. This past week, in fact, Alphabet said it would increase capital expenditures to $75 billion in 2025 from $52.5 billion last year -- a huge A.I.-driven jump that surprised Wall Street and that, along with a slump in Alphabet's cloud-computing sales, may account for the decline of its shares. Military Supremacy One company making heavy use of A.I. whose shares have surged this month is Palantir. It's not a consumer brand, but its technology is widely used, not just by corporations but by the U.S. military, police forces and U.S. Immigration and Customs Enforcement. These have been growth areas in both Democratic and Republican administrations. Shyam Sankar, the company's chief technology officer, told analysts that DeepSeek had made basic A.I. cheaper. But he added: "I think the real lesson, a more profound one, is that we are at war with China. We are in an A.I. arms race." Alexander Karp, the company's chief executive and one of its founders, spoke unabashedly of Palantir's role in ensuring U.S. military supremacy. "We believe we are making America more lethal, making our adversaries increasingly afraid of acting against the interest of America and especially Americans." Careful readers may recall that Mr. Karp and Mr. Musk appeared together in Strategies in June. They were the two highest-paid executives of U.S. public corporations in 2023. This was according to filings with the Securities and Exchange Commission that for the first time required universal disclosure of "compensation actually paid." Mr. Musk gained $1.4 billion in 2023 while Mr. Karp had a windfall of nearly $1.1 billion, the filings showed. Both of their fortunes continue to be tied to A.I. and, in idiosyncratic ways, to the U.S. government. A.I. companies come in many shapes and sizes and will need to be valued in tiers, Professor Damodaran said. Consumer-facing companies embedding A.I. chatbots in services available to millions will benefit from lower-cost, commoditized A.I, while cutting-edge A.I. with military, corporate and scientific payoffs may receive premium valuations. Infrastructure companies like Nvidia can benefit from these variations and more, but not all ventures will require huge expenditures on the most formidable chips. Given the complexity and uncertainty, it makes sense for long-term investors to diversify while A.I. fever cools down, Professor Acemoglu, the new Nobel laureate, said. "I have a balanced portfolio," he said. "So I've got tech stocks, health stocks, real estate stocks, everything. I don't go out of my way to pick, you know, Tesla or any other company. I'm an index fund kind of guy, and I don't do anything other than that."

Jeffrey Emanuel and the lessons we should all learn from the $2 trillion DeepSeek AI market correction

Over the course of a Friday, Brooklyn-based blogger Jeffrey Emanuel penned The Short Case for NVIDIA Stock, which proceeded to viral and helped wipe $2 trillion off global markets. In an interview, he tells me that it was one of those stocks he had not invested in nor shorted, and the process of writing and organizing his thoughts helped him to understand why it was so overvalued. He says: When I published it, I couldn't take a short position because I finished it Saturday at 3 AM. I had no position in NVIDIA. Theoretically, I could have done a pre-market trade, but I hate doing that because you're just chasing the thing down. Besides, it wouldn't have been meaningful for me to do that. For clarity, Emanuel has a dim view of technologists' opinions on stock prices, meme stock investing, and most crypto, but he is bullish about Bitcoin. He is also bullish on the long-term financial prospects for AI, but fears it will erode opportunities for many. His initial motivation for posting his essay was to build traction for his YouTube Transcript Optimizer service that creates transcripts, summaries and quizzes. (This service was used to transcribe the interview for this story, and it's at least as good, if not better than Otter,) That service has seen a minor bump, but he has made far more money consulting with hedge funds at $1000/hour since the essay went viral. His post is a long read that starts with the bull case for NVIDIA and walks through each component of the AI hardware moat that helped the company rapidly grow into a $3.5 trillion company. The essential argument is that NVIDIA's high value is based on its near monopoly protected by moats, including its leading chips, efficient drivers, CUDA programming framework, existing skills base, and interconnect infrastructure. Since these moats allow them to earn 90% returns per chip, the industry is eagerly exploring ways to chip away at this moat. Cloud providers stand to see better value even when their combined chips/software/interconnects are half or even twenty percent as efficient. He notes: I didn't quite realize that every single hyperscaler customer to a man had their own custom silicon like in various stages of completion and readiness. Perhaps most concerningly, the relative value of chips deteriorates rapidly as the energy per compute costs improve over time. Companies that over-invest in AI hardware today will thus face higher operating costs than those waiting for future chip generations. The cost-efficiencies introduced in DeepSeek are likely to rapidly be adopted by other AI developers. This could further reduce short-term demand for AI compute, creating a costly glut in data center capacity and an air pocket for NVIDIA growth. He also explores how the balance of the different kinds of chips used for training versus inference may quickly shift. More than just being cheaper, the success of DeepSeek also highlights a shift in future demand for inference versus training hardware required for new reasoning models. The NVIDIA $.6 trillion correction had a knock-on effect that broadly eroded $1 trillion off US tech stocks with residual $2 trlilion total effects on smaller worldwide markets. Shortly after the drop NVIDIA rose as people suggested a Jevon's Paradox effect could accelerate AI adoption. However, the stock is still down thirteen percent from its recent high. Emanuel's divergence from the previous market wisdom is a case of the Dunning-Kruger effect, where technologists opine on markets without a proper understanding of valuation principles. He says: I think a lot of people are very smart, high IQ, they understand semiconductors, they understand AI, but then they try to start talking about valuing a stock, and I'm like 'Oh this is a completely different game.' If you don't know what you're talking about, it's very easy to make a complete fool of yourself. These people often just embarrass themselves because they talk about this like it's straightforward, but there's nothing simple about valuing a company. It's not just about what it's going to do next year -- it doesn't matter that NVIDIA has contracts for the next two years that cover a lot of product. Markets are smarter than that. These big hedge funds look through that -- they want to know what's going to happen after that. It's all about expectations, and when expectations have to be revised downwards, that has an impact -- even if the future could still be bright, but it's not quite as incrementally bright as the most bullish people thought. Then, all of a sudden, the stock can be down 20%, and it's wiped out hundreds of billions of dollars of equity value. It's amazing to me - the Dunning-Kruger that you see among technologists trying to opine on the stock market is laughable. People who couldn't even do a very simple valuation of a very uncontroversial company and have it be compelling are so confident that they know how to value a company that, up until a few years ago, made a few billion dollars a year and now it's being valued, even after this decline, at close to $3 trillion. Maybe the future is exactly as bright as they think. I think it in fact will be, but that still doesn't mean that we're not going to hit this near-term air pocket where the supply-demand imbalance, which has been so skewed in favor of companies like NVIDIA in particular, is going to temporarily skew in the other direction. People are going to say 'Oh, we bit off more than we can chew in the near term, so let's cut back on new commitments of capital. Then all of a sudden it creates this air pocket where expectations in the near-term have to be adjusted lower." Was the $2 trillion market correction too extreme or does Emanuel believe this is a temporary pause in AI-driven market growth? He explains: Yes in aggregate, but no in terms of NVIDIA in particular. If you read the piece, I'm very careful about saying these comments are very specific to NVIDIA because it's not enough for the whole industry to continue growing very quickly. They also have to maintain their exceptionally high market share and their truly unprecedented margins, which is basically only possible if you essentially have a monopoly. He does not believe that it was an overreaction in a single day but rather more of a catch-up. Many highly leveraged companies build other sorts of data infrastructure that suddenly earn more than imagined thanks to higher margins. For example, companies selling air conditioning equipment and other ancillary parts have high fixed operating costs, and when revenues go up, profits can rise. But these can also go down just as quickly if there is an air pocket in demand. If cloud providers slow down AI infrastructure plans to take advantage of DeepSeek innovations and delay and defer future chips, they will also be buying less of this ancillary equipment. Meme stocks Emmanual has sometimes been characterized as part of the meme stock community. But Emmanual argues this is the opposite of what he does. Referring to the Game Stop phenomenon a few years ago he explains: I would very much separate what the meme stock people do from what I do. What I do is make an argument where a totally sober, realistic, careful analytic person who's really being mindful of their investors' capital and their fiduciary obligations and not trying to do anything crazy might find it compelling enough to perhaps sell their position or enter into a short position as a hedge based on the fundamental content of the analysis. When I started writing it, I wrote it because a friend of mine who works at another fund called me to ask me about it. This was on Friday, January 24th, when I was writing most of it. I started around 3 p.m. Eastern, and I finished it around 2:30 in the morning on Saturday. I just wrote it essentially in one sitting. He also believes there is a big difference between saying that a stock is not a good thing to buy versus you should bet against it. Most stock falls into a no-man's land where he does not have a strong opinion either way. He would not have considered buying NVIDIA until he started writing the article because the valuation was too frothy. But it was also too scary to short because the AI story had such momentum. His opinion shifted as he dug deeper: It was only after I was at the last couple of sections that I thought to myself, man, this is actually way worse than I thought for NVIDIA. When I found out -- I knew about Groq and Cerebras, but I didn't quite realize that every single hyperscaler customer to a man had their own custom silicon in various stages of completion and readiness. That's when I was like, whoa, do people really know this? That every one of their major customers is trying desperately hard and spending billions to replace their reliance on this company? So then I went back and changed the title. I think that's part of why it spoke to people -- I'm not like most of the people who will tell you NVIDIA is a short, who are skeptics of the AI narrative. Because then if you're not a skeptic of the narrative, you will immediately discard that person's opinion without another care. He has even harsher words for the meme-coin crypto community exemplified by Dogecoin and the recent Trump coins: I disapprove of that stuff. I think it's very corrosive to social cohesion. I think it sends such a horrible message to young people when they hear about some guy who just happened to buy the Trump coin at 18 cents and made 50 million dollars. It makes a mockery of the idea that you should wake up every day and work hard and if you do, you'll make some money and have the American dream. I think it's very bad. But Bitcoin and other types of crypto infrastructure are different animals: I'm a huge believer in Bitcoin -- I have most of my money in Bitcoin. But Bitcoin actually solves real-world problems. It does something very useful, interesting, and special, whereas a meme coin is truly just this new version of a Ponzi scheme where there's no utility. It's just a pure 'greater fool' thing, and the last people who get in are just the dumbest ones and end up losing all their money. I think it's horrible. So, I disapprove of all that stuff, but I don't disapprove of crypto in general -- it just has to do something useful. It actually has to serve a purpose other than facilitating speculative mania. Emanuel is cautious about clarifying his take on NVIDIA's stock price and its leadership and ethics. I ask about the trend towards enshittification in the tech industry, where vendors erode the value of their services while raising costs. He says: I wouldn't say any of that stuff applies to NVIDIA. I think they're a class act. I think Jensen's an extremely good guy, and I don't think they do anything sneaky. I don't think they've done enshitification of anything. I think they've been great -- they've helped support this research when nobody believed in it. Jensen tells a story about when he hand-delivered the first batch of GPUs to OpenAI back in 2015, back when this was an insane pipe dream. So he's sort of been nursing this from the cradle. And they open source a lot of stuff. This isn't a reflection of them at all. It's more that 'you're never as good as they say, and you're never as bad as they say.' This thing got ahead of its skis, and it puts him in a very difficult position because even though he's a man of high integrity and good intentions toward people, you simply cannot go out there and say, 'Our company's stock is way overvalued.' After all, you'll crash the stock and everyone will be furious with you -- your employees will be furious, your investors. And so you can't say it. I suspect that in his heart of hearts, especially recently, he realizes that maybe things have gotten a little too frothy in terms of near-term expectations. And that's because it became this treadmill -- as you keep exceeding expectations, not only do expectations go up to reflect that the results are higher, but now they start factoring that in and start to handicap it." Emmanual believes that AI will continue to get unbelievably better over the next three-to-five years. It is already smart enough to do most white-collar knowledge work, and it is just a matter of integrating it into existing business processes. This will look like replacing offshore call center workers with AI for basic tasks in the short term. Also, many paper-shuffling jobs will become more vulnerable as AI improves. The other thing is that humanoid robots that could replace many unskilled workers are also starting to get closer. And, much like DeepSeek, he thinks Chinese companies may be further ahead than expected. In particular, he notes that most American robot companies have shared canned and highly edited videos of their robot capabilities. In contrast, China's Unitree is showing more realistic videos of their robots that can do unbelievably agile things and have cut costs dramatically. He also has high hopes for a new generation of Solopreneurs who can discover ways to spin up new companies and test new ideas, using AI to fill in many of the software development, design, marketing, and accounting processes. But in this new world, if you're a 'solo-preneur' -- and it definitely requires a lot of ambition and a jack-of-all-trades mentality so it's not for everyone -- but you could theoretically not hire a single human. You could just do it all yourself. Even if you don't have any artistic talent, the AI will do the design work for you, and it'll look fine. It'll do the coding for you, and if you know what you're doing, you code at like 10-20x the productivity level of normal people not using AI. You can conceivably make a great product if you have a good idea in one or two months. Well, one-to-two months isn't long, so even if it doesn't work out, it's not the worst thing in the world. Which means you don't even necessarily need to raise money from an investor, you can just risk it yourself. But then, at the other end of the spectrum, there will be people who don't understand the technology or how to leverage it. These people may get completely squeezed and pushed out. This could create a surge of inequality. He explains: I'm very worried about social cohesion from this. It's hard to say what we're going to do about it. Universal Basic Income gets floated around, but UBI has significant problems. Even if we implemented it in a fiscally sustainable way, I think you'd see skyrocketing deaths of despair, like overdoses, because people find meaning in their work. There's dignity in work. Nobody wants to feel like they're just on the dole playing video games. It's hard to maintain self-respect and feel good about yourself in that situation - just sitting back while robots do all the work. Some people might be fine with that, but for many, it would be really tough to stomach. Like Emanuel, I am somewhat bullish on the prospects of AI. NVIDIA is also investing significantly more in physical AI and embodied AI innovations than competitors, which may further drive further gains. It's also possible that Jevon's Paradox effects may further accelerate short-term demand as people find more ways to create profitable AI-powered services. But by the end of the call, I realized he had brought up a lot of important points relevant to market mechanisms and over-supply I had not previously considered. His take on the potential for social unrest is also sobering. He also suggests an additional ray of hope when I asked about the strange green stringed instrument arrayed with others on the wall behind him: