Ep20. AI Scaling Laws, DOGE, FSD 13, Trump Markets | BG2 w/ Bill Gurley & Brad Gerstner

And this was Friedman opining in a very quick, rapid-fire interview of the 14 government agencies that existed at that moment in time, which he would abolish and which he would keep. And you know, we'll roll it and put a clip in here. Keep them or abolish them? Department of Agriculture: abolish, gone. Department of Commerce: abolish, gone. Department of Defense: keep, keep it. Department of Education: abolish, gone. Energy: abolish. How, except as energy ties in with the military? Well, then we shove it under defense. The little bit that handles the nuclear, right, that ought to go under plutonium and so forth goes under defense, but we abolish the rest of it. But you know, they asked him: Department of Agriculture: abolish, Commerce: abolish, Education: abolish.

Well, and I would add a few other things. Ilya was also quoted as questioning whether the scaling laws were continued. And for whatever reason, Mark and Ben over at Andreessen Horowitz also made the same statement. So in a very short window, we got this point of view from quite a few number of people. I would add that in Dario's five-hour interview with Lex, which I only grabbed pieces of from the transcript, he was very positive on scaling laws in that interview, which just recently dropped. So there may be a disagreement between some people, but the breadth of the feedback suggests something's up. I think something's worth paying attention to. And I would add that there were people who raised this question early on, and I think it's specific to LLMs. You know, I don't think you should say AI scaling laws. I think there was always a question about whether LLMs would run out of steam, and that was tied to three different things, which you and I talked about on July 11th, and I'd even raised this question a year ago. One is, you know, will the parameter count run out? Historically, mathematical algorithms that do some type of fit, and this is a very sophisticated form of that, but when you take the variables up to a certain level, they stop adding value. You just get too close to the fit. There's a question about how big the context window will be, and that came up on the Nvidia call. I think we can talk about that. And then data was brought up, and a lot of people, I know this was a big point Melanie Mitchell had raised, like, are we just going to run out of data? And you know, there was pushback that there's synthetic data, but there's been papers published that show synthetic data creates a lot of chaos. So I do think there were people that were thinking this might happen. And there was also a belief, and I share this belief, you don't have to, so we can disagree on this, that there was an argument being made by OpenAI and Anthropic that, oh, you're just going to see the next number drop, and the next number is going to be way better than the last number, and that's going to happen routinely. And they both said, you know, we spent whatever a billion on this model, and we're going to spend 10, and then we're going to spend 100. And that implies, I would say, at least linear scaling or maybe above, and they all said that. So if the comment you read earlier is true, and that these models are not getting the benefit, there are implications. It doesn't mean AI is in trouble or AI is done. I mean, there's tons of positive AI news out there, but it may mean we're shifting directions. And it's worth, I think, talking about: if this is true, what are the implications?

The first thing that comes to my mind, and once again I'm not asking you to agree with this, it's just like what's in my head, is that this is not the expectation that Sam and Dario had 12 months ago. Because they were promoting a thesis that it was just going to keep going up and up, and they were willing to put 10x over and over again on the size of the cluster they were going to train on. And if it's a shift, and it may very well be a shift that leads to insane innovation and continued growth, blah blah blah, I still think it's a surprise shift. Like, I don't think it was what was expected, and so I think that's worth acknowledging if that's what happened. Two other things that I think are, let's just call them questions that are raised. One: the Nvidia differentiation, as we've talked about, is greatest at the largest cluster size. So if pre-training, once again I only equate this to LLMs, you know, I don't think of the FSD problem, I think that is more of a core AI problem and that thing may scale way beyond where it is today. But for LLMs at least, if we've hit a cap, then does that impact Nvidia demand at all for the bigger systems? And then there's a differentiation question, which is: if the world shifts to inference sooner or faster, and if the path of differentiation for the language model companies becomes more inference-based than pre-training-based, what does that mean for the hardware systems that people are using? And then the third thing, and I'll go quick and you can respond, is: might a ceiling on that pre-training allow other people to catch up faster with OpenAI? And there was an article out today about a pseudo-Chinese model that was performing at or near some of the latest... DeepSeek, what you're referring to. So anyway, those would be the three things that I would say that you would want to analyze if you thought that was true.

I would add to what you said. I listened to the call. There was certainly no concern raised by Jensen about either his demand or his sense of what was going on at the architectural level. And Microsoft has been increasingly boisterous as well. If this thing is true, it certainly goes against what Kevin Scott was saying on a podcast four or five months ago. And so I doubt that, even if it were true, it would play out in the numbers as fast as we're talking. This stuff just came out like weeks ago. But I wouldn't ignore it. I think it's important to understand what the drivers are of incremental investment, incremental use, incremental performance, and competitiveness, and to see where that's going. And I think there are a lot of questions that we need to ask and understand about inference. Do models that are, as I understand it from the startups that we work with, you can train a model on a big Nvidia cluster and you can run it on a different system. And there are certainly performance tests that are in the public that show Groq and Cerebras and other TPUs, including the ones that Google has, outperforming on inference. Now on the call, Jensen said something that I hadn't heard him say before: he talked about context windows potentially growing really large, and that might then require a larger cluster, which might bring a point prior. On the previous call, Jensen had just said, "Well, the reason you'll use us for inference is because you've got old machines laying around." This was a step in a different direction, saying, "Well, if context windows are huge, we might be differentiated yet again here with larger systems." So I think those are the things that people need to ask and figure out.

I would just make one comment, kind of over again, on this scaling wall or ceiling starting to arise. There was talk prior to now, or if you assume that weren't true, that these companies were just going to train on bigger and bigger and bigger systems. They said, "Oh, someone will train a model on a hundred billion dollar system one day." If it's true that it's topped, that's not going to happen anymore. So that is a... it doesn't mean that Nvidia may still be sold out till 2030, but there was this notion that there was just going to be a constantly building larger model for pre-training LLMs, and that may be off the table if this is true. And let me suggest, as an investor at least in one of them, OpenAI, it may not be a bad thing that that's off the table. They don't have to build that system. I do think there was this question around returns and where the money was going to come from.

I mean, all those things, memory, voice, they can be utilized in an enterprise fashion for sure, and will be. So yeah, I think you're right about that. I think those innovations will matter for the exact same reason you said, which is utility will go up. I think Snowflake said on their call tonight, 3,200 of their customers, so they have about 10,000 enterprise customers, 3,200 of them are using some of their AI use cases, AI products now in daily use. So again, starting to see some of that indication. I noted that Microsoft, which is on a $66 billion run rate at Azure, has said they expect Q1 and Q2 Azure revenue to accelerate. Think about that, Bill. Azure's growing at 34% on a $66 billion base, and they say they expect it to accelerate. So it's really significant growth on big software businesses. And I think what we've seen in the season is that software certainly has not decelerated, if anything it's accelerated after a couple years of digestion around this. But if I had one takeaway having watched all of this, I would agree with you, and you've been rattling my cage about this for a year, and I think you've been proven at least partially right that we're seeing a deceleration in the rate of improvement movement on pre-training. But my conclusion is I don't think it necessarily matters to the consumption. On the Nvidia side, certainly it would be higher, could be even more I suppose if they were able to produce that many more, which I don't think they can. But I think that they're still in a very strong position. And I think these models, for example, OpenAI could double or triple the number of users of ChatGPT with frankly not making any improvements in the core model, because consumers today barely scratch the surface on core model capabilities. Product improvements like voice, memory, and actions, I think in and of themselves could drive tremendous traction and revenue increases. But I do think they'll still continue to see some of those model improvements.

There's a couple things I would respond to that I think are going to have to play out. First of all, as of now, Tesla hasn't had cars driving around without drivers picking up people, and Waymo's got years under their belt doing that. And there's going to be some learnings that come from that, there just are. So they're going to, I think I read that they're going to maybe start doing that in the next few months or something like that. But in the states that are allowing Waymo, you have to apply, you have to get approval, so there's a step that they need to take. And I think there's some natural learning that comes from that. I mean, you can hear how they talk about SpaceX, and there's a similar thing here. Like, until you get the cars out on the road, most people believe Waymo still has oversight from a data center somewhere where people are watching when engagements happen or disengagements happen, so they can take over and pay attention. And so is that same infrastructure going to be built at Tesla? I don't know. That might be an important thing as well. To the national standard, I unquestionably think it can be helpful. There are states in this fine country of ours that like to override regardless of a national standard, most notably California. So that would, but of course it would make more sense for this to be done nationally, and so that'd be great if they stepped in. There's a lot, I'd always been a big proponent of open source in this area. There's a lot of stuff that could be beneficial from a safety standpoint if everyone were doing it the exact same way. And I could even imagine a communication network. The status of a traffic light, for example, should, there's no need for that to be inferred. I mean, we're going to use how much compute to look at the light and determine whether it's red or green? That's a state machine, it's either red or green, and that can be known. And so you could, as you build these intelligent systems, start to have the state of different things, which could include road openings and road closures and construction work that's going on and traffic and all kind of other things that should be built into the system as well. So I do think a national kind of standard could be great. It could also be really bad, like if you get over-regulation and this massive stuff, and the incumbents might just use it to protect themselves. But my own belief is that you've got Waymo who's got the most cars running autonomously on the street right now, Tesla who's got all the advantages you just talked about, you've got Uber who's got this huge network of users. I'm less certain who the next player is after those three that's got a strong argument right now.

Well, let's talk about that, because I think that's an interesting question, a market structure question that I've been thinking a lot about lately, Bill. Certainly on the public fund side of our business, when I have a disruption of this magnitude in an industry that's a multi-trillion dollar industry that has 20, 30 different public players in the supply chain, OEMs, etc., I just start thinking: who are the winners and who are the losers here? It's feeling like 2008. You see the iPhone, you've got Nokia, you've got Motorola, you've got Blackberry, you've got all this stuff, and you're like, somebody's winning and somebody's losing. I remember my mother didn't understand the concept of shorting in 2009. I told her to buy Apple and to short Nokia, and that was my canonical example for her. It worked out okay, it actually went against her first, but then it worked out okay. Let's just talk about OEMs for a second. I think the consensus view in the world is that all these OEMs, so I'm talking the US OEMs, the Stellantises, the GMs, the Fords, the Europeans, the Volkswagens, the BMWs, the Mercedes, the Japanese, the Toyotas, etc., the Chinese, that FSD becomes a commodity and that they all end up with some version of it. Maybe there's one provider like Nvidia or Mobileye or somebody else who's providing this capability into these cars, and you know, think of that as maybe like this Android version that they're all going to, this box they're going to plug into their cars, and they're all going to go compete as though they were before with Tesla. That's one version of the world. The other version of the world is that this is a tectonic disruption. Tesla currently sells 2 million of the world's 80 million cars, but that in the future, this isn't going to be about buying a car, this is going to be about a mode of transportation. And that Tesla could end up restructuring the entire industry, end up being tens of millions of cars. You may only have two or three or four global players, which we have in a lot of other industries. So the question I have for you is: what is your pattern recognition? Do you think that we're going to have all the same OEMs, they're all going to have some system, it's going to be basically commoditized, or is this an extinction moment for some of these OEMs?

Such a hard question. I mean, there's so many different ways this could play out. One of the things that I question in my mind is who's going to own these vehicles. So on one hand, you have the Waymo situation where they own all the cars, but they also know what that means from a capex standpoint. And if you look at the subtleties of the kind of the statements they've made in conjunction with Uber, they're at least going to experiment with some models where maybe they don't have to pay for all cars, or licensing their technology to the other OEMs that we talked about. And people I've heard Dara talk about at Uber talk about maybe large debt companies get behind and own these things, like maybe the containers work on a rail yard. Like, who owns these things? They move around, but someone owns them, people lease them. So that's a big question. And the Tesla argument is that they'll just be owned by the people that own the Tesla, which is great from a Tesla doesn't have the capex burden, everybody else has because they've talked them into buying these cars. But what percentage of people are going to lease out their car every day? That's what would you guess? I have no idea what to put on that number. I mean, Airbnb does exist, so it's maybe not zero, but most of us still have our own house. So where does it fall? I think my own hunch on that is that it's going to be a higher percentage, particularly as the price points on the cars fall. Right now they're leasing cars for $200 a month, and as those subscription prices fall, I think you're going to have a whole category of people who don't mind at all earning incremental revenues on their car. I imagine they've done internal survey work on this. And I think if you look at the places where they're likely to launch the robo taxi, number one, it's going to be in marketplaces where they have sufficient safety, so where they have a density of data that tells them they can put it in robo taxi mode and they won't have any problem. I think that could be somewhere like maybe the Bay Area, somewhere like Austin. And then secondly, it's going to be a market where they have a lot of Teslas sitting on the sidelines. So they come with 25 or 50 Teslas, but in both of these places, they happen to have Tesla factories, there are a lot of people with Teslas, and there are a lot of people who will drop them into the pool. So those will be good test markets. But if I think of a market like New York City, New York City seems to me to be a problem because there's massive demand for cars in New York City, but there aren't a lot of Teslas sitting around latent that can be tapped into for supply. So I don't think it will work equally well in all markets, but I think it is a massive advantage. Remember, out of their, I think they may have 7 million cars on the road today, something like that. About 2.5 million of those I think are Hardware 4, and it's going to require Hardware 4 in order to run on all these cars. So we're really talking about a multi-year undertaking. But if you said to me, over/under on middle of next year where we have evidence of a single market or two where robo taxi is working, I think we will. I'm going to take the over on that, and that is a huge difference from where the world was 18 or 24 months ago.

Maybe we can have an expert on the China thing. I've certainly seen videos online that suggest the Waymo equivalents are already working in China, so that would be a data point. Freda went and rode in every single one of them, I think there were 12, and she took videos in every car and sent it back to the team. Some of them were totally terrifying, obviously anecdotal. Everybody from Huawei to Baidu to Didi to BYD, etc., has their version of it. And there's another question, I think, which is both Waymo and Tesla have hinted at potentially licensing. And so that'll be an interesting maybe that happens down the road between the historical incumbents and these two new companies. And whether that type of model plays out seems quite likely to me.

Well, I think that's a perfect segue to talking about DOGE and what's happening with this tectonic election we've just had. And I think those really big ambitious objectives, which frankly we haven't seen a lot of ambitious efforts to reform over the course of the last few administrations, maybe both Republican and Democrat. The one thing about the current administration, it strikes me, has a level of ambition that I don't know that I've seen in my lifetime, Bill. One is the Department of Government Efficiency. So this is the unofficial government department which is run by Vivek and Elon and is overseeing trying to reform government spending, government regulations, frankly reducing the size of the federal government, with the overall objective of getting control of our national deficit and our national debt, which we all agree is pretty egregious. And one of the videos I saw going around, I saw several people tweeting it, including Elon, was of Milton Friedman. And this was Friedman opining in a very quick, rapid-fire interview of the 14 government agencies that existed at that moment in time, which he would abolish and which he would keep. And you know, we'll roll it and put a clip in here. Keep them or abolish them? Department of Agriculture: abolish, gone. Department of Commerce: abolish, gone. Department of Defense: keep, keep it. Department of Education: abolish, gone. Energy: abolish. How, except as energy ties in with the military? Well, then we shove it under defense. The little bit that handles the nuclear, right, that ought to go under plutonium and so forth goes under defense, but we abolish the rest of it. But you know, they asked him: Department of Agriculture: abolish, Commerce: abolish, Education: abolish, etc. And I think in tweeting that, it kind of sets some boundary conditions of a level of ambition that I don't think many people, they heard the words, but I don't think many people said, "No, they're not really going to do that. They're not really serious about going after that." What do you expect out of the Department of Government Efficiency? And is the agenda going to be that ambitious, or do you think it will hit the wall of some congressional reality and maybe they're just setting the boundary way out there that they can negotiate back from?

Well, first of all, I think if you extrapolate the growth of government two or three decades into the future, we would get to a point where such a large percentage of the population work for the government that it would self-implode. And many very smart people have also highlighted that our debt is too large, our US country debt is too large, and that interest payments are now a huge part of the federal budget. And how can you possibly solve that if you don't shrink the size of government? Now, prior to the election, the papers said that both candidates were going to spend the same amount. So this new department is orthogonal to what most of the media thought Trump was going to do. I would add, I highly recommend everyone watch the two-hour interview between Lex Friedman and Milei from Argentina. I think a lot of this is being provoked by him. He was recently in town and hung out with the DOGE team. If you take him at his word for what has been accomplished in Argentina in such a short window of time, six to nine months, it's pretty spectacular. And maybe it's important for someone to fact-check that, but really remarkable what can happen if you improve efficiency. And things like you got rid of a ton of regulation around rent control and housing improved for everyone. And for me, that's like an "of course." Of course that's what would happen. So how do you tear this apart? We could simultaneously, if they're successful, see a reduction in government and an acceleration or proliferation of growth, because you find out government's getting in the way. As I have said on the past two or three podcasts, Governor Shapiro in Pennsylvania keeps getting accolades for moving regulation out of the way. It's so ridiculous. If that's what you get celebrated for, we should just rewrite regulation and get it off the books. And we know there are states that are more productive and are growing faster right now and are actually increasing their population because people can build things. They can build factories, they can build distribution centers, everything can happen, they can build solar farms a lot faster. So I do think it's super important. I would really encourage people to watch the Milei thing. The part that's most interesting, because I think the easy pushback is, "Oh, they'll never be able to get it to happen. Washington's too ensconced, and you just can't move it, or these people will protect themselves." The methodology that I'm going to assume was Elon's idea is basically to shine a flashlight on the idiocracy that exists. Transparency can be a hell of a disinfectant. And while you and I have started this, I took a flash look at Twitter and VC just announced that he's going to pause his own podcast to launch a new podcast with Elon called "DOGE Cast," where they're going to give regular updates on the cost cutting and what they're going to do. And they've asked people to send in their favorite example of government waste. They're going to talk about it. And I think when you highlight to the American population really idiotic spend, they're going to get behind this thing. And so that's a point here. It comes to this question of can they get it done? There have been plenty of presidents, Clinton, Reagan, etc., who've talked about getting government spending under control, talked about empowering the individual versus the state, and then they meet the leviathan of Washington. Maybe just a little background here on the executive power as it relates to this. There's this constitutional doctrine that we've been talking about in our thread, the doctrine of impoundment. And it's the ability of the executive branch to withhold or delay the spending of money appropriated by Congress. This was a question as to whether or not the executive branch actually had this authority. Congress passed an act in 1974, the Impoundment Control Act, which basically restricted the president's ability to withhold funding, saying Congress is the one that allocates funding. If we allocate funding, then the executive branch has to implement the law, has to spend the money. Now, notwithstanding that fact, there was a big in Trump one, the Trump administration withheld $391 million in military aid to Ukraine that Congress had appropriated, and some argued that this violated the law. Now why do I bring this up? I bring this up because on the one hand, there is going to be a constitutional authority question here. But over the weekend, I talked to several members of Congress, both in the Senate and in the House, on the budget committee. And I will tell you, there is strong agreement, and it wasn't just from Republicans. There were people who were trying to figure out how they can support and interact with the Department of Government Efficiency on both sides of the aisle. So I think that notwithstanding the, if they can get Congress to cooperate, then there is no constitutional issue. And then the final one, Bill, which is the one that I think is the most powerful here, at the end of the day, what we've seen time and time again, the power of Twitter, the power of X, going directly to the people with your message, saying, "Here's what government's standing in the way of, here's what we want to do." I think that's overwhelmingly powerful, shaming the government into making what I think a lot of people will agree are logical cuts. And so I'm bullish. And one of the things I keep hearing in the investment world is lots of people are trying to read between the lines. They're saying, "Well, what do you think the president really says? Or what do you think's really going to happen?" And I said, "Stop trying to guess. Just listen to the words." After the election, Trump is literally recording on X reiterating exactly what he's going to do. People are tweeting about what the Department of Government Efficiency is going to do. You're going to have these DOGE casts. And I think that's exactly what they're going to attempt to do. And I think this is going to be a once-in-a-generation resetting of the balance of power between the individual and the state. And frankly, I think it's well needed. I hope it's successful. Certainly there'll be some things they get wrong, just like anybody who takes on an ambitious project will get some things wrong. But I think you can undo those. One of the things that's already come up is that I think two of the largest departments in the country have failed audits for like three to five of the past years. If a single company on our public markets fails an audit or has to change auditors, as we recently saw with Super Micro, you get taken to the woodshed. How about that? Down 90%. Yes, it's considered a failing of the worst kind. And I've already seen interviews where someone's asking these department heads about these failed audits, and they're like saying it's okay. And that should not be okay. The dollars are massive. These people work for the citizens of the country. And if they have an audit requirement, they should be meeting it. I'm shocked to find out that that's true.

Agree. It'll be interesting to watch. And once again, I would encourage people to watch. I'd say watch just because I think Milei is such a fascinating human. And I'd watch it on YouTube just to see his facial expression, although it's dubbed because they actually did the interview in Spanish. But he's a very unique individual. He was in a rock band. He obviously has studied the history of economics like no other, a disciple of Milton Friedman, which comes out in this thing. And if what he said is true about what's been accomplished, it may be the fastest impact a president of a country has ever had on the trajectory of that country. So anyway, I'm going to agree with you. I'm in the optimistic camp. It's about time. There is zero doubt in my mind that making the government more efficient in our country would be very positive for everyone that lives here.

Well, and I saw some data that suggested auto delinquencies are on the rise. So any concern about the consumer for sure. I think that listen, we know that they've spent through all their stimulus money, that they're not getting the pay raises and all the overtime that they were a couple years ago. And we know that when you look at their credit, they're really spending a lot on credit now. And we see that same data on auto delinquencies. And so if you just go through the things that Trump says he's going to do, imposing tariffs is bad for retailers like Target. So that's one side of the equation. Then on the other side, you have somebody like Snowflake or Palo Alto report tonight, and they beat earnings, but expectations are low. Snowflake is up 20% in the after hours. So just think about that spread trade: Target down 20, Snowflake up 20. One has super high expectations, super high valuation. One has super low expectations. And in the case of a company like Snowflake, there's massive room for efficiency. They're getting the efficiencies from AI. They don't need to hire at the rate that they were hiring before, and they still just grew 29% in the quarter. So that's far from the massive deceleration that many people were projecting. And so I do think that the beneficiaries in technology are both beneficiaries of this top-line growth but also bottom-line margin expansion. I was looking back in preparation for this pod with Satya on Monday. Since he became CEO of Microsoft, the net income margin has expanded by a thousand basis points. If you look at the story of Microsoft over the last 10 years, revenue has doubled, but earnings have almost quadrupled because of the expansion of the earnings margin in that business. And that's what I think we're still going to see in technology, and it's going to be accelerated with AI. So I head into next year not all in, but I'm definitely constructive on the backdrop. I just think it's going to be a stock picker's market. It's going to be a bunch of stuff that you want to avoid, but you can find your 5, 10, 20 tech companies that are enjoying the tailwinds of AI and also enjoying the possibility of this margin expansion. I think they can continue to do reasonably well. So that's how we're positioned.

Yeah, another way of saying it: I don't expect a 25% up year for a broad-based index. And if you look at it, I don't know where it is as of today, but there is one point in the year, Bill, that if you took the top 10 tech companies out of the S&P, the S&P was not up 20%, it was actually down 5%. And I think that's the type of dispersion that I think we're going to see again next year. There are going to be a lot of winners, either as the result of changing government regulations, changing government policies, lower taxes, or as the result of this tech super cycle. But there are going to be a whole lot of losers. We just talked about the ones in FSD. It may very well be that this is terrible for a lot of OEMs, but really great for Tesla. We have yet to see. I'll tell you as an investor, and we could leave it on this: the rate of change has never been higher in technology, and the rate of change of this election is massive. The magnitude of the policy change is massive. You combine those two things, and let's just say we've been working overtime. Makes sense to me.

Well, one of the things we may want to talk about in the future is whether the Trump administration is going to lean into AI regulation as well, which could have implications for everyone. Anyway, thanks so much, man. Good to catch up. Sorry the Seattle thing didn't happen, but we'll get it next time. We'll get her done on Monday. G2.