New Tariff Implications, DeepSeek and Open Source Analysis, and Meta’s Big Bets w/ Byrne Hobart

I'm doing great. I've just landed in Sydney where I've come to hand deliver the tariffs personally.

Yeah, exactly. Just kidding. But while we're on the topic, before we get into it, if you had to predict, do you think that some of the vibe shift will come to Europe in a similar way that some of what's happening in the US might hit England, Australia, or these countries?

I think that in some ways the European vibe shift is more extreme in terms of voter behavior than it is in the US. There are like, and some of it is just that if you have a system that is designed to have multiple parties, then you typically have room for a more extreme right and a more extreme left. But even within that, there's been more reshuffling. And a lot of what Trump does is rolling things back to a status quo that existed for a while. That is what conservatives and reactionaries want to do generally: take something that they believe broke recently and put it back. And there's a debate on how far back you want to go, and there's also within that debate the question of how literally do you take any of these rollbacks. Are you trying to make it like the 90s again where China is not in the WTO and immigration is a bit more restrictive, etc.? Are you trying to do that, or are you trying to ask yourself, okay, what would a centrist Clintonite from the 1990s, if you transported them forward in time, what would they be doing today? Because those give you different answers, and I think that does make some assessments of this kind of tricky. But a lot of Trump's views on social issues are just things that were pretty centrist in living memory, or he's drifted but more slowly than the average person. And then on economic issues, he is more of a throwback, particularly with respect to the tariffs. But I also can't tell how much of the tariff stuff is a Trump-specific vibe shift, how much of that is just general skepticism of globalization, skepticism of the economist establishment. I also think that within tariffs, there have been people who are clearly opponents of tariffs broadly, or not Trumpy people, but who have said that the tariff debate is a little bit more sophisticated than just talking about the first-order consequence of there is deadweight loss if you make it harder to import goods into a country and that loss is going to hit consumers. It's not just like everyone ignores the fact that there are tariffs and suddenly everyone who exports goods to the US just accepts a 25% lower margin if they're from Canada. That is just not a realistic description of human behavior. That said, within that idea that you could be more sophisticated on tariffs, there is a lot of nuance to exactly how you do that. It is not trivial to set up a tariff regime that is going to encourage industries where the US could compete but maybe can't. The financial markets don't have a will to fund companies that will scale up to the point that they can compete, but the government can implicitly fund that by making consumers worse off. There are cases where that's a high ROI decision to make, and a lot of the East Asian countries that had enormous growth, particularly manufacturing-driven growth, one of the ways they would do that is they did protect their home market. So Japan was a great place to be an exporter and a pretty bad place relative to GDP to be just a consumer when they were at their peak growth period. But that works if it's an investment and if you can build up these export-driven industries such that they provide higher returns later on, such that you turn those exports into importing things that people actually want. That is the end goal. And the US is in a weird situation because we sort of achieved that end goal where yes, people are really happy to take dollars in exchange for things that American consumers want to buy, but the US is also just not in a position to manufacture a lot of pretty straightforward intermediate goods. And some of that is also just when you have expensive labor markets, you can't cost-competitively manufacture some of this stuff. It is just going to go where the labor is cheaper. So there are a lot of problems that are worth talking about, but blanket 25% tariffs on countries that are extremely economically integrated with the US, that makes sense as a bluff. It does not make sense as a long-term strategy. And even as a bluff, it is such a sufficiently weird thing to do and has so many economic downsides. Like automotive, for example, if you have goods that are just constantly zipping back and forth between the borders where there's some component that is made in the US, it's put into something else in a factory in Canada that's sent back to the US, Ford puts it in a car, that car gets sold to an American. If you're putting a 25% tariff on every leg of that transaction, then a lot of complicated supply chains just can't be that complicated. And maybe it's worthwhile to have these supply chains all be equally complicated but all be within the US. I don't think that category of good, like intermediate goods made out of other intermediate goods that are put into final manufactured goods, I don't know that that is the strategically important industry that the US really ought to be protecting. So the tariff stuff is a separate part of the vibe shift, but I think there is just a general, it is very old news that there are a lot of people who have defined an establishment that they are very skeptical of, and that used to be a very left-coded thing. We don't really talk about the man anymore, but it used to be a very left-coded thing that you sort of viewed the military and the Democrats and Republicans and big business and tech as all part of this big conglomeration of special interests that did not have your interest at heart. You can find old, I think there used to be this, so on IBM punch cards they used to have this warning: do not fold, spindle, or mutilate. Which of course, if you are folding the punch cards in order to stuff them in something, it's harder for the machine to read the middle, jam, etc. And people did have to learn that. But that term, do not fold, spindle, or mutilate, people started using that as just a general political slogan, which was basically I don't want my entire identity to be represented as a number that's run through a computer, especially if that computer is deciding who gets drafted and who doesn't. So now that whole vibe of there is an establishment, it is out to get you, that still exists. It is now right-coded instead of left-coded, but I think it was directionally true in both cases that you can't have an establishment, it can get kind of stagnant, it can end up pursuing its own interests at the expense of everybody else's, and that it's useful to shake things up. So that is the vibe that is shifting. I think it's very hard to impute that kind of thinking to just everyone who is participating in this, and there are just lots of different flavors of populism. That's part of the nature of populism, and part of this vibe shift is that people are looking more at their country's particular interests rather than at universal goals, however laudable those goals might be. And when you're doing that, you have a different set of interests if you are in France or in Hungary versus in the United States. So you should expect populism to just have very different flavors in different parts of the world.

I think the big downside to that is that people have been, for very good reasons, skeptical of the EU's cohesiveness and sustainability pretty much from the beginning. It did grow, but it felt like every time it grew and every time it got more powerful, there was just a lot of local opposition and a lot of people felt like this can't go on forever, and it did just kind of keep going. And also, if you have a country that is aging or if you have a region that's aging and its economy is growing more slowly, it's going to be more conservative in the more literal sense of trying not to shake things up, because there are just a lot of ways to get more downside and they don't see as many ways to get that much upside. You obviously have these kind of hopeful glimmers of European startups that are able to compete globally and are able to produce high returns and provide excellent jobs for people with technical skills, but there's a lot of the continent where it just feels like a lot of things can be routed around. And then there's a question of does it happen sooner or later, but there are fewer cases where it's just an absolute lynchpin in a broad sense. And then there are a lot of narrow industries where, yeah, there are European midsize European manufacturers who are absolutely indispensable to some supply chain, so they'll be around for a while. But I think when you have a case like that where you sell this one little component with a lot of pricing power, you sell one particular chemical that is used in some really high-value process, you again probably don't want to shake things up too much. You probably want to keep roughly the status quo because anything that changes that might mean that your product is going to be manufactured somewhere else by somebody else, and there goes a little bit of your trade surplus.

So at least in the case of Mexico, it sounded like the tariffs were really not about Mexican companies competing with American companies to create jobs and that Trump wanted those jobs in the US and not in Mexico. It was more about border-related things, about climate, immigration, and drugs. For Canada, it's harder to make a case that immigration from Canada is a huge problem or that drug smuggling from Canada is a huge problem. It's not that there's zero of it, there's just a whole lot less. So I'm actually less sure what concessions Canada can actually make to the US that feel like they're worth whatever the hypothetical revenue of a 25% tariff would be. But I'm sure both sides can think of something. Maybe there's some line on the border that could squiggle its way over and maybe that's where things end up. But I don't know, it seems like the Trump administration's territorial ambitions are much more literal with respect to Greenland and the Panama Canal than Canada, where the 51st state thing just seems like more teasing than anything.

So my understanding of Belt and Road is that it was a tagline to describe a lot of things that China was already doing. And the way that I would look at it is that if you have a country that wants to exercise economic influence on other parts of the world, it is a lot easier to do that if you allow the free flow of currency and of people, etc. The US has a lot of influence on the rest of the world because it's relatively easy to move dollars around, and the dollar is a good default currency to use for transactions between two countries who don't really want each other's currency. So it's also because the US financial markets are so developed, borrowing in dollars or raising capital in dollars is generally more straightforward than other currencies. And there is actually just a valuation premium for equities that are listed in the US that you can't explain with any statistical test that people try to do. There's a good piece from For Research on this recently. You control for everything you can control for, you know, control for size, control for growth, control for margins, and you still get a valuation premium purely for being US-listed. So all that makes the dollar just a really powerful tool for the US government, because once everyone's using dollars, you can threaten to cut people off from the dollar system. It's a really big deal. Whereas, you know, if I were cut off from the ruble system, there's just no point in my life at which I would notice or care. There's just nothing I want to do that requires rubles. But there are a lot of things that various powerful people in Russia want to do that are just much more convenient if you can handle and transfer dollars. So that's something that just naturally emerges if you have a large economy and you trade a lot with the world and you have multinationals. And China has a lot of those things, but it doesn't have a good way for Chinese currency to circulate outside of China and get back in. That process tends to be pretty controlled. So what China tends to do is take some of their trade surplus, use it to build up infrastructure, use that infrastructure to solidify economic relationships with other countries, and then kind of monetize it through importing more raw materials. And so they get some of that soft power and they also get some additional economic growth, at least if they make good investments. And they do that without losing that much control over their currency. But that's something that China had been doing for a long time. They had a huge infrastructure buildup when they needed infrastructure, and then when they responded to the financial crisis by just radically increasing state spending, pretty close to single-handedly pulling the global economy out of a recession just by using so many different kinds of raw materials and building so much stuff. I think towards the end of that process, they realized there just aren't a lot of railroads that are not yet built and ports that don't yet exist and highways that are like incremental highways needed. Like China's pretty much set in terms of that kind of physical infrastructure. But there are other parts of the world that still need some, and if China's economy is growing and China does have a significant military and influence on global affairs, etc., they can cut some fairly lopsided deals. But right now, I think they have to be just more focused on the domestic economy and in particular on the overhang from a real estate bust, which had to happen at some point, but does mean that there are a lot of assets underwater, loans that really maybe could get paid but probably should not get paid, probably should be restructured. And if they do get restructured, if they were marked to market, a lot of financial institutions would have very large holes in their balance sheet. So there are solutions to that, and countries have been able to recapitalize their banking system or just grow out of it. China did actually grow out of a similar problem in the late 90s where their banking sector was actually insolvent, or at least the big state banks were insolvent, but there wasn't going to be a run and the banks were structurally profitable if they weren't required to make politically motivated loans to state-owned enterprises that were not actually going to pay them back in a timely fashion. So there was room for China to just kind of roll with it, and there's less room now because growth is just less baked in and the creditworthiness of the marginal Chinese borrower is a lot worse if the country is just a whole lot more levered in general. So they do have fewer options there, and even fewer options if they don't want to concede that they made policy mistakes in the last round. And you can do a thing where the central government says, well, everyone one level below us completely messed up their economic planning. And the thing is, if you do that in one province, it's probably pretty believable to say, you know, we set up these incentives to produce economic growth, we knew that they could be gamed, we just couldn't imagine that one of our own would do that. The problem is if it happened everywhere, if everywhere people were following the incentives too literally and creating fake growth and doing a ton of off-balance sheet borrowing, at some point you have to say that actually the incentive structure was encouraging everybody to do this. If you don't have exceptions, if you don't have exceptional cases where people follow the rules, then obviously the implicit rule was break some of the explicit rules. So it is somewhat embarrassing for them. Maybe in retrospect there was some time where they could look at some high-growth provinces and say not only is this fake growth, but we're going to seriously punish the people responsible and let that be a lesson to everyone else. And they occasionally did things like that. There was at least one province that just had fraudulently inflated its GDP by some huge amount years ago. So they did have an accounting fraud scandal in the Chinese growth story, but I think that it's at the point where enough people have been following the actual incentive that there's just a lot of bad debt in the system and it can't get worked out without conceding that there were some bad incentives that were put in place by the CCP.

Yeah, it was so that was just a really funny, really fun time. Basically, what often happens when there is bad news that affects some theme in the market is that you'll actually see the pretty high-quality, high-growth, large-cap companies go down because they're all tied to that theme, and you'll also see a bunch of lower-quality bets on the same thing go up because people will want to do this offsetting bet where, you know, if we think that people are going to spend more time online and they're going to look at more online ads, but also that the small-scale ad tech companies just don't have the scale to monetize as well as the big platforms do, then you might be long Meta and Google and short a bunch of other midsize companies. And if you are running a totally factor-neutral book, you probably have other large-cap shorts, other small-cap longs, such that you've balanced your exposure to the size factor. But that is generally roughly how it works. And so that's what I was expecting. I was expecting that at a lot of funds, the risk person would say, hey, you are paid to pick individual stocks. I noticed that most of your performance comes from things tied to AI, so it looks like we're actually paying you a lot of money to say AI is good and express that in 15 different ways in your portfolio, and we don't want to pay you very much money to do that. That's an easy one. But it turned out that actually people were getting paid to do that. And what I did see, which was really interesting, was I just went through the list of holdings for a large momentum-focused ETF and I looked at the performance of everything that wasn't an AI bet, and all of them were up. It was not up by a ton, but like Philip Morris was up and Progressive Insurance was up and there were pharma companies that were up. There was just a bunch of kind of lowish-risk stuff that had done well over the last year was way up. And so I realized that a lot of people were long AI, and since that means since AI is working, that means they were long momentum, so they were shorting other momentum stuff to offset it. And then when they exited all those positions, the non-AI momentum stuff went up. So I thought that was one interesting piece. And I'm still a little bit confused as to who was paying close enough attention, who had a sufficiently sophisticated view that they were able to update on the DeepSeek thing, but they also owned some of the more fringe kind of garbage AI bets. And I won't name a bunch of them because people have talked about them in detail elsewhere, and also I'm short some of them and it's always a dubious honor to be publicly identified as someone short of a specific stock with a large retail following, you tend to get death threats. So I'm going to avoid that. But a lot of those stocks went down too. So that either means like a lot of retail investors are actually pretty keyed into day-to-day developments at AI, or that a lot of institutions were speculating on some fairly dodgy stuff that happened to be AI-flavored. So yeah, I think that what's possible is that, and I'm sure some funds do that, I'm sure some funds come up with these qualitative ad hoc risk factors and tell their portfolio managers don't bet on this risk factor and we won't pay you just to make bets on this factor. But I think there's always room for refining how you think about what is idiosyncratic performance for one stock and then what is just betting on a theme in a bunch of different ways. One of the other interesting highlights in terms of stock price performance on that day was that there were actually companies that did worse than Nvidia that were utilities or companies that build data centers, things like that. And one thing that occurred to me was I don't know that there's been a case where a US-listed utility stock dropped 20% in a day because of lower expected demand. I don't know that that's actually happened since the 1920s and 1930s. Like utilities, the demand picture has been very, very clear for a very long time. But when that demand picture changes, when you have a historically very low volatility company that has very, very predictable earnings power and you know roughly what the growth rate will be, any uptick in that growth rate has just an enormous magnitude of impact on the actual valuation. So I think that was a case where if you look in risk-adjusted terms, that maybe buying power companies was actually a smarter bet than buying Nvidia. But that also meant that the marginal buyer, the price setter of that company, was no longer someone who was a traditional utility analyst who is thinking about the things that utilities people think about. It was someone who was really gung-ho about AI, looking for an interesting bet that was going to capture some AI upside. And when those people went away and utilities got valued like regular utilities again, that was a pretty big, very quick haircut to their valuations.

So I think that's one piece, but there's also the software moat piece, and that DeepSeek was able to not just use Nvidia's own software for programming GPUs, they wanted to layer deeper and found some clever optimizations that Nvidia's people apparently had not found or had not bothered to find for this particular use case, for this set of chips. The set of chips that they used was pretty much, from what I understand, basically there was an early round of sanctions and Nvidia started shipping a chip that was just designed to be as powerful as it could be while still sanctions-compliant. And that's probably just not a market that Nvidia saw as a really great long-term market. It was more like this is an opportunity to move some inventory and the rules will probably change again, especially if this product takes off. So exactly the kind of case where you wouldn't expect them to really focus on optimizing for that particular customer, that particular use case. So I think that is like, it was still a surprise that they were able to improve things as much as they were, at least surprise to me. But I think that is part of what people were reacting to, is that Nvidia, there can be like, Nvidia has such a software moat that there could be a time where they are slightly behind on the hardware side, but nobody is going to take the hit from the sunk cost of all the CUDA-specific stuff they've done, so they just buy the next round of Nvidia chips and then hopefully Nvidia catches up again, gives them a little bit of margin for error, makes a slightly more durable, less hit-driven business. And if that goes away, then it is closer to a commodity business. Now it's a really good commodity to sell if that's the case, because demand is booming and demand has over the last couple years persistently overshot anyone's willingness to supply enough chips. So still a high-margin business, but maybe one that deserves a lower terminal multiple, however you figure out Nvidia's terminal metrics. So I think that was a big part of it. And I also think that what people are underrating, like there's so, Jevons Paradox instantly became a meme, it instantly became just like the midwit explanation for why you should not update your views at all in response to this news. But I think if you flush it out a little bit, there is actually a case for much more consumption as measured in tokens, even if total user minutes don't scale that much, and even if revenue takes a while to scale, like even if revenue doesn't scale proportionately, because a lot of the more interesting AI tools that are being launched by the major labs, they are agents, they are agent-like, they engage in multi-step processes. And if you're trying to explore a problem space and you start out with here's the premise and then you come up with 10 different promising directions you could go and you follow each tree for a little bit until you find that okay this one is probably a dead end, we'll try something else, etc., like you can just generate a lot of tokens to get to your answer. You can pretty much consume just arbitrary tokens. Like if that is how you scale the performance of that kind of model, then you can just keep scaling. And if the cost of a token drops by 90%, then the same o3 model that was answering graduate-level math problems for $3,000 a problem, now it's $300 a problem. And if that happens again, it's $30 a problem. And at some point, just being able to have an on-demand grad student in any domain where the answers can be checked against something true but the answers aren't trivial, which is basically everything that is either math or uses lots of math, if you have that, there's just a much larger set of possibilities that you can look at. Like I would naturally turn towards investment research, and I would say that if you're looking at some company that is more on the deep tech side of things and they've been developing a product for a while, or let's say they have a drug and it's going through trials, being able to for $300, you know, that would be 90% off what o3's cost was when Meta talked about its performance in late December, being able to do that for $300, like for $300 I would be happy to just go through a list of biotech companies and dump in a bunch of data on what drugs they're testing and where they are in the testing process and just ask a really smart model, is it biologically feasible that this could actually work? And my guess is there will be cases where it's just not feasible, like it is just clearly not going to work, it is not a mechanism that actually does what the drug is intended to do, and that's a pretty straightforward short. And then if you have an industry that, like most industries, has roughly similar risk-adjusted returns to the rest of the market, and then you can segment out some set of the companies that go to zero, then you can also just buy everything else and now you have a nice factor-neutral or at least industry-factor-neutral portfolio where you're short all the biotech companies that will eventually blow up and long ones with lower but non-zero blowup risk, and that's a good place to be. You could imagine doing this kind of due diligence for a lot of other technical questions. So, you know, maybe Boom Supersonic, like just asking the model to start reasoning about things like how low does the cost of supersonic planes get under different scenarios and how many do they need to make before they start earning a profit on them, etc. These are questions where you could spend a lot of time, you could hire someone, you could hire an aerospace engineer as a consultant and have them work on this for a couple weeks and you'd probably get a better answer. My guess is that for a lot of these questions there is actually just enough domain knowledge that you get a better answer, but if you can get an answer that is like 85% accurate and doesn't take very long and the cost is comparatively trivial, then there's just a lot more stuff you can do and you can sort of outsource certain parts that require a very deterministic kind of rigor and then focus on the things that are just more vibes-based.

Yeah, and with Meta, like I was confused on Monday morning when Meta opened down, because Meta is a consumer, they are a buyer of tokens in the sense that they are monetizing generative AI by showing people more content, by prompting them to comment more, by pre-writing the comments for them basically, and by being able to generate a very long tail of ads. So if it gets cheap enough, every ad just gets generated on the fly based on what is going to convince you personally right at this moment to make a purchase, and that opens up a lot of interesting possibilities and it means that there are even bigger scale advantages to being someone like Meta. Like if they have more data on your behaviors and your preferences, then they have a better return on making those ads, so they make more of them, which means they can afford to scale up their hardware more so than other people. I made a few people mad on Twitter recently by pointing out that as a percentage of sales, based on Meta's current guidance, they will spend twice as much on capital expenditures as US Steel, so relative to sales. And I think that's a good thing. I think that they actually can get an incremental return on a lot of the hardware that they're buying and that they may be able to calculate that return more precisely than a lot of other participants in like a lot of other GPU buyers. But it is kind of funny to me that this classically asset-light business that literally starts as a group of teenagers in their dorm, they already own the computers, so the only expense is you need a server and you need a lot of Mountain Dew, everything else is just stuff you were already spending money on. So now they're a very capital-intensive company, but also a company that can probably project its returns on capital quite nicely. And if that's the case, then they should be really, really excited that there is a cheaper model. They said that they want to make the AI models as cheap as possible and they want these open models to be ubiquitous. But also my suspicion is that one of the reasons this is bad for Meta is that if you have a group that is building models that are open, that anyone can use, etc., that is really cool, like you're able to build something people use, you're also able to talk about what you're doing in a way that someone working at a lot of the more closed labs just can't do. So it is uniquely attractive to some of the people that Meta would really want to hire. And in some cases, if they have those and then they have some more prosaic task that happens to involve a lot of AI knowledge, so something like when they spun up Reels, targeting Reels was much harder than targeting other kinds of content because they are not just looking at first and second-degree connections, they're actually looking at the entire space of Reels content, and Reels are videos, so you have to actually figure out what they're about, you have to actually analyze the videos and do that at scale, etc. So that was part of why Meta made this large initial investment in GPUs, and then they were able to repurpose that. But if they had AI researchers and they have them on hand, they can, I suspect, tell some of them that as much as we like the pure research you're doing, as cool as we think it is, like for the next six months we have this really important business goal and we think you can help us accomplish it, and then they could do that. So Meta wants to have some reason for really good AI researchers to be on Meta's payroll and be available to occasionally do stuff that has more of a revenue or cost impact. And I think that's a harder case to make if there is another lab that's outperforming them and is using fewer resources. Although I guess the other important thing to point out is that we don't, I don't, as far as I know, we don't really have comparable numbers from other labs that compare directly to the specific cost that DeepSeek is talking about, because what they talked about was the cost of doing the final training run for the model, not the cost of all the experiments, not the cost of the data gathering, etc. And those costs are a non-trivial component of the overall cost. I don't know what the relative breakdown is, and I'm sure that breakdown evolves over time. But yeah, the specific headline, like 5.5 million number, is not directly comparable to a lot of the other numbers that's being directly compared to.