Ep16. Nuclear Update, AI Fast & Furious, State of VC | BG2 w/ Bill Gurley & Brad Gerstner

There's a picture you can look up that's kind of disgusting, so people may not want to, but there's this thing called a gavage tube, which is what they use to make foie gras. It's how they force feed the geese to get them just super fat. And that's the image I have in my mind, like, are we overfeeding these startups?

Hey, Bill, great to see you. Good to see you, Brad, man. That was an amazing pod at Diablo Canyon. You know, the inbound regarding just nuclear has been off the charts, from literally senior policy makers, senators, and house members on both sides of the aisle. It really feels like the dam is broke. We half-joked, Bill, that Microsoft, you know, when we were down there, we discovered that there were four unbuilt new nuclear reactors that are all already plotted on the site. And we half-joked that Nvidia and Microsoft and Oracle could come sponsor these reactors and that they could have a new type of public-private partnership with the government and build data centers right next to them. And it turns out it wasn't so far out there. I mean, Oracle has announced that they may do some things with small nuclear reactors. Amazon is buying this nuclear power, Talen data center facility. And now Microsoft this week announces with CEG that they're going to bring Three Mile Island out of retirement. It's incredible to see the beginnings of what may be a US nuclear Renaissance. Certainly the momentum has, maybe it was headed up and then it's kind of reached an apex and kind of fallen over. The FT article yesterday or today highlighting that 14 different banks have shown up at a climate conference with a confirmation of a willingness to invest is just huge. And I think there are two things that are big takeaways for me. One, we were talking about one of the limits on SMR and on any new innovation in the space was that utility companies are traditionally very conservative. And I like to think about it in the framework of crossing the chasm. You are basically selling only to laggards, and that's very difficult for especially a capital-intensive startup to be selling to laggards. And what may have transpired literally in the past month is the hyperscalers — and this may have started before them because Amazon did the deal with CEG a little while back — but if the hyperscalers become part of the customer set for the nuclear startups, that may be like 10x better than selling just to utilities alone. You may have brought innovators to the table on the purchasing side that may be more open-minded, may be more understanding, may be more willing to share risk, which could be very positive for the SMR market. So that'd be my one big takeaway. And the second one is just that a lot of times I think people look at big problems and think they're insurmountable. I remember actually in the past two years being at an off-site conference at a think tank where we were talking about climate change. About 80% of the way in, someone raised their hand and said, "Why aren't we talking about nuclear?" And all the scientists in the room said, "Oh, no, we're not going to put that back on the table. That's too far gone. That's passed." And it turned out that wasn't true. It turned out there was an opportunity to get a Renaissance in thinking about this. And it started, I think, with people like Stephen Pinker, who are widely regarded scientists saying, "No, this is our best path out." But then you know we talked about Patrick Collison and others, kind of jumped on the bandwagon. And then there were plenty of pro-nuclear advocates that were sticking their neck out. And then Elon gets in the game. And then this data center thing may have been just the impetus you needed to get people over the top. And we were lucky enough to kind of time our thing as this transition was happening. But it is possible to create kind of wholesale change in how people think about something, but it takes a lot of work by a lot of people. And everyone that kind of stuck their neck out early — Josh Wolfe was another one that was sticking his neck out on this topic. So I congratulate all of them, and it feels like the momentum's now behind us. And I literally feel bad for the citizens of Germany. One thing that is very apparent is that the easiest thing to do is start with don't decommission any of these things, but second, if any have been decommissioned recently, try and bring them back. And I hope there are some sane minds in Germany that are watching all this, because I think the world would benefit from them reversing that decision and running back at this. One of the things I learned as well, because you and I talked a lot about how do they underwrite — how would the hyperscalers underwrite building out those nuclear reactors? One of the things I learned after our pod was that these companies that are considering nuclear, they are spending billions of dollars a year on carbon offsets. And you know, there's a lot of criticism about these carbon offset markets, but I dug up some data. In 2020, Morgan Stanley estimates that about $2 billion was spent on the carbon offset market, and by 2030 they expect that to be $100 billion by these large hyperscalers that have to buy these massive carbon offsets. Now, if instead you're investing in nuclear clean energy, if the source of the energy that is powering your data centers is clean, then you actually get to buy fewer of the carbon offsets. So that may make it easier for them to pencil out the math. I think a huge part of this is just the public consensus. Nobody wants to invest in something that all your customers are against. And we know we've shown the data here that this is now popular again among consumers because they understand it's clean, it's carbon-free. The other data point that broke since we did that was the Three Mile Island restart. I think there were rumors of it before we did the podcast with CEG, with Constellation Energy Group. And there were quotes in these articles about a survey of the Pittsburgh or the Philadelphia residents, and they were supportive. And what that says to me — and look, knowing that the customer is really Amazon and not the citizenry, I was just shocked by that. So we all know that one of the reasons this happened was that there was an irrational public response to the negative risk of these solutions, and it is super unfortunate that that takes so long to kind of heal. But time is the best way to get past something like that, and it's been a long time. I think people have a lot more data. We're not there yet — we need to keep the pressure on. We said we'd like to see Gavin Newsom extend Diablo by another 15 years. That's on his desk right now. That facility has at least another 40 years left in it. So I think we all need to keep the pressure on, but the nice thing is there's good bipartisan support. People view nuclear not only as a matter of climate security, but now it's a matter of national security because it's the prerequisite to all of AI. So I'm excited for the momentum. But speaking of the exploding need for more baseload power to feed the AI beast, let's talk about our first topic: where things currently stand in AI. I think since we last talked, people have continued to climb this wall of worry, like whether AI is in a bubble or not. Of course last week, OpenAI was rumored, or now is well known, to be raising capital at a $150 billion valuation, or at least that's the Bloomberg headline. I can confirm that Altimeter is talking with the company, so of course there's some things I can share and things I can't share. But you had Kevin Scott, the CTO of Microsoft, say demand for AI infrastructure is materially outpacing our ability to supply it, even as we are building at a pace unseen. Jensen said at the Goldman Sachs conference that they will be in under-supply not only this year but for a while to come. So where do you come down on this, Bill? Let's just start with the demand for training and inference and data centers and power. Do you think we're headed for a glut?

Right, right. And I agree with you. I think you said something really important, which is when you're planning out three to five years and we're talking about tens of billions of dollars, one has to assume that sane people like Jensen and Oracle and Amazon, etc., that what they're seeing justifies the spend. And one of the things driving the debate over that demand is all these new models that have been popping out. We finally saw Strawberry, the O1 preview model — a whole new vector of scaling around inference-time reasoning out of OpenAI, which is exciting. There's a new Anthropic Opus model 3.5 that's expected to drop this week. Meta Connect is in a couple days — I'll be there — and Mark Zuckerberg is expected to announce several new models there, smaller models and larger models. So despite the improvements in model training efficiency, the aggregate training and the velocity seems to continue to grind a lot higher. We were lucky to have Noam Brown, who is the inventor and the leader of the O1 preview. We talked about Noam before — he was at Meta and did Pluribus, Libratus, won the game of six-handed poker. He spent a lot of time thinking about this inference-time reasoning, and we're really early in this journey, but it's already kind of a big wow. This idea that in addition to pre-training, we're now going to allow the model to reason and think as part of the response process to the prompt. I don't know if you looked at any of the tweets out of Noam Brown or others about the O1 preview, but I'm curious if you had any thoughts or reaction. I know you were blown away by the voice model. It's been slow in dropping, but I don't know if you played with the O1.

Yeah, and whether or not the engine can interpret which of those it is will be important. I would go back and highlight that even in the announcements from OpenAI on Strawberry, they admit, disclose, or qualify that there are many instances where the extra iteration led to worse results. So you really do need to be able to figure out the type of problem and whether or not you're going to get improvement from that effort. I don't think we know enough yet to know which problems fit in there or not. But I do think people are super excited about what's possible on that front. My biggest thing about it is, listen, we've been talking about bigger models, more parameters — that's basically been the exclusive vector of conversation for scaling intelligence. This is a totally new vector. And now you have the compounding benefit of two different ways to scale intelligence, which is super exciting. But just real quick, I had a second thing before you move forward.

I couldn't agree more. I think there's going to have to be massive price discrimination. You can't charge in the Philippines what you're going to be able to charge in the United States. You can't charge the tail end what you're going to be able to charge the head end. But one thing that is true is that OpenAI has something like 200 million weekly MAUs. That's a number widely reported. It's a huge, huge number with little to no advertising. And it seems to me this is really the benefit of going first, Bill. Billions of free ad impressions. It continues to grow. So I asked the team to take a look at the time to reach 100 million MAUs. You could see this chart — it took ChatGPT a fraction of the time that it took YouTube, Instagram, or Facebook to get to the same place. It took Facebook about five years. And that obviously has continued to go up. Weekly average users of 200 million is pretty extraordinary. And even by country, if you look at the spread of ChatGPT, it's clearly spreading on a global basis. And then finally, this tweet by Ved on my team shows that ChatGPT is beginning to run away with it. Gemini, Meta AI, Claude are really not even keeping up. So if you set aside valuation for a second here, Bill, why do you think the game on the field as to consumer has changed so much? Has ChatGPT now in a flywheel? Have they broken out of the pack? Do you think they're going to be the winner in consumer AI?

Yeah, it's pretty extraordinary — the pace and velocity. And frankly, we see that on the team side as well. Just extraordinary team. The best people continue to go there. Everybody asked me a year ago, they said, "Oh, Gemini is coming, Gemini is going to go first." But the fact of the matter is, right through the chaos of the moment, Bill, everybody's responding to them. They launch Strawberry and O1 preview first. They launch advanced voice first. So you got to give them some credit for that. But when it comes to 200 million weekly MAUs that are reported out there, I would also make this argument. The team has made a chart on this, and I'm curious of your thoughts. It appears to me that OpenAI is seeing more and more of a network effect as well as scale advantages. You've talked a lot about network effects, but there's been debate as to whether or not they exist here. But here's the argument I would make on the network effect side: it seems like more users is leading to better data. The data coming from the interactions with those users is leading to better models and cheaper models, because you can do more of the work in post-training, which then leads to more users. Here's the chart we made on it. Do you buy the network effects argument? That flywheel? Because if that's in place, then it explains why we're seeing them break away from the pack when it comes to consumer AI.

Yeah. The data on the field — I'm just looking at the data on the field. I'm looking at the number of users of Meta AI, etc. It looks to me like among the new consumer entrants — and I like the guys a lot at Perplexity as you know, and I know lots of people like that product — but just from a usage perspective among the new entrants who have the capital, who have the surface area to compete, it looks to me like ChatGPT has clearly broken away. And this is going to be a game between them. Meta I think is probably in the second best position. Google probably can't be underestimated, obviously. Sacha has consumer Copilot with Mustafa there. But it's really interesting to see that game. The other vector here that's interesting is there's a lot of rumors out there about OpenAI's revenue, Bill. And we'll post some of these — four or five billion in revenue, that's what I've heard, growing at over 100% a year. I'm just intrigued by that. If that trajectory were to continue, that'd give you roughly $10 billion next year. And the rounds rumored to be at $150 billion, so that's about 15 times forward revenue. So I asked the team two things. I asked them to compare that to other companies, namely Google and Meta, both in terms of the pace to get to $5 billion in revenue and the valuations once they got there. This first chart just shows that OpenAI was able to get there roughly in two years from the launch of ChatGPT in November of 2022. It took Google about two or three more years than that to get there. It took Meta almost six or seven years to get there. So that's interesting — they got there a lot faster to $5 billion, we can agree on that. But then I asked what were the multiples at that point in time, because I remember when I bought the Google IPO, everybody said it's overpriced. When Microsoft invested in Meta, everybody said it's overpriced. But what's interesting is Google IPO'd in 2004 at about 10 times forward revenue. Microsoft invested in Meta in 2007 at about 50 times revenue. Meta IPO'd in 2012 at about 13 times revenue. And now again, if all these rumors are correct, you're talking 15 times revenue. So it's basically in an EV valuation zip code that's similar to those other companies, if you accept the trajectory. And I know you have some real thoughts about margin here, whether or not the quality of those revenues are the same. So I thought I'd just throw that out there and ask you.

Yeah, and I think it's such an important point to make. When you're investing in a company, you got to get the top line — who's going to be the winner — and then you need to be able to forecast that top line. But that's not ultimately what drives valuation. What drives valuation, as we've all often talked about here, is the future cash flows that those revenues can produce. And there's a real question on the table that you've articulated well, which is: is there going to be a tax here that Nvidia and the cost of inference and the cost of training imposes in perpetuity on these companies, such that it's going to always be a less profitable business than a Meta or Google on the consumer side, or an AWS and Google Cloud or Microsoft Azure on the enterprise side? I certainly think you're right that at the start of the super cycle — if you go back to the start of AWS, remember the debates then, Bill: 2009, 2010 — "Can it ever make money?" Because they had to get to massive scale, and the cost of delivering that scale had to come down. So you're betting on two things with OpenAI with respect to margins. The first thing you're betting on is that they can get to scale, because this is clearly a scale business. The second thing you're betting on is you have to believe that the cost of inference is going to come down meaningfully over time, and the cost of training will come down meaningfully over time. We already know the cost of inference has come down by over 90% over the last 18 months. And our friend Sunny over the weekend, when we were in our group chat, said he expects it to come down by another 90% over the next several years. But those are the types of things you're going to need to be true in order to have a margin structure that is consistent with those legendary businesses like Google and Meta.

Well, it's going to be interesting to watch it unfold. But there's certainly a related topic that I know you're amped up about, and is probably even more amped up given this rumored OpenAI deal: how the venture model is changing, and whether these structural changes are good or bad. Whether they're good or bad for LPs, whether they're good or bad for GPs and founders. So why don't you lead us in a discussion on that topic?

Yeah, no, I mean, listen, I think it's a great analysis and framing of the issue. You bring up a lot of great questions and the right concerns. Let's try to break them down, because I think they fall into roughly three buckets. Let's start first with the question of more dollars, bigger funds, more competition, and higher valuations. There's no doubt that VC has grown from $100 billion 10 years ago to $300 billion over the last 10 years. If you set aside the COVID period, where we all know because of zero interest rates public markets lost their minds and venture markets went to high levels, we're back to this $300 billion level, which was pre-COVID. So while we call all of this venture, Bill, one of the big differences I have here — and I've said many times — is that the venture market really hasn't grown that much. Much of the investment we're counting as venture, when you look at all these data sources, is into companies that are higher than $1 billion in valuation with huge revenues that would have been captured by public market investors before. So I think it's important as an industry that we start thinking about these things as different. I call them internally, and to our LPs, I call them quasi-public companies. Companies like Databricks, Stripe, OpenAI — I think it's silly to call them venture at this stage when they have $5 billion in revenues growing 100% a year. So yes, I would say the late-stage quasi-public market is much more competitive, just like the public market is more competitive because it leads to better price discovery, but it also means that there's less arbitrage and returns are more dependent upon long-term compounding than some misinformation in the market. As we saw in 2020 and 2021, the public markets corrected. In fact, a lot of the IPOs that happened during that period are still down over 50%. So the public markets corrected, just did it quicker. In the quasi-public markets, we've seen a lot of these companies shut down, sell, and are still down over 50% from the high. So I don't think there's a lot of difference there. When I look at the early-stage venture markets, Bill, I would agree with you that there's a lot of excitement around AI, but outside of AI, Series A follow-on rounds or Series B follow-on rounds are down dramatically. The number of first-time funds that are getting funded as second-time funds is down dramatically. So I see a lot of reversion to the mean happening rather than structural change. I think the big structural change that this data leads me to conclude is that because of the regulatory burdens of going public, because of the change in Silicon Valley around sentiment, because of the ability to get liquidity in secondary transactions, because of the liquidity of this late-stage quasi-public market where institutions like Altimeter or Coatue or Fidelity or Thrive are here to provide that liquidity — I agree with you that there's a lot more money there because those companies are choosing to stay private. But we should think about these and compare them to their public company competitors, not to what's happening in the Series A market.

I understand. And that's the third point I was going to make. The quasi-public isn't the same as public in terms of how the public markets might shape the motivation and execution of the team. We go back to the Meta example. They went public, stock went down. Wall Street says, "You're not ready for mobile." Zuck later said, "[__] that, that kicked me into gear." Those things don't exist when this isn't here. I agree with you, this is playing the game on the field. I'm not blaming anyone, I'm just highlighting this is where we've matriculated to. This is where we stand today. I think it's right. So let's move to the second big point, which is liquidity and IPOs. It is true the number of IPOs has been anemic and the exit amount in venture is now at about $100 billion a year, while down a ton from the $700 billion peak in zero interest rate policy in 2021, which was really a one-time COVID high. We're basically back at the same exit level for venture that we were pre-COVID. Lots of people talk about the zombie corns. We have a thousand companies that were unicorns. A lot of those will never get back there. I've said 80% of those companies will never get back there. Those companies need to get merged into other companies, get sold, get shut down, or do down-round IPOs like Instacart did, which is now off to the races under some great leadership. But I look at just our pipeline, Bill, to give you a counter example. We may have four IPOs in the pipeline in the next four to six months. The rumors out there around companies like Cerebras and CoreWeave and Databricks — those are all in our portfolios. And on top of that, we recently sold Tabular to Databricks at a price rumored to be $2 billion. So exits are increasing. Interest rates are coming down, which I think will lead to more of that. The world is healing. So I'm not so sure this is — I do think there are some things that are structural. The regulatory things are structural. The more dollars in quasi-public is structural. But there is still enough incentive. These companies want and have boards that will get them public. I just think they may come public a lot later, when they're at $10 or $50 billion valuations rather than at a $2 billion valuation. As far as the companies I'm involved with, I'm pushing for them to come public sooner, or at least when they're ready.

No, I think it's a super important conversation. By the way, you talk about if 80% of the zombie corns are never going to get out or are going to get out at reduced prices, I'm telling you flat out: those are being held on the large endowments as LPs at unrealistically high prices across the board.

Let's talk about this question that I think is the most important one, Bill: does excess capital lead to companies being overfed, which leads to poorer outcomes for innovation? Because the potential for that — at the end of the day, that would be the worst thing. I think you and I have a lot of shared belief that too much capital does ruin corporate culture, it does lead to higher burn rates, it does lead to lower financial returns, and it does lead to less innovation. I pounded the table on this topic: the time to get fit. I applaud Mark Zuckerberg. He stepped out in February of 2023 and wrote a letter called "The Year of Efficiency." He said, "We started doing this just thinking that it was about getting back to the office, but what we discovered was that smaller is better. Flatter is faster, and leaner is better." What he meant is the cycle time on innovation — delayering the organization, getting rid of layers of LPs and VPs — really getting the organization tight and fit was better for future growth, future profitability, and future innovation of the business. I tweeted over the weekend that it's great news that Zoom and Salesforce and Workday are starting to get sober about stock-based compensation. That is another component of too much capital, too many people, leading to excessive SBC, which you and I have talked a lot about. Just because you raise a lot of money doesn't necessarily mean you're unfit. Remember, OpenAI is not really a VC company at this stage. Google went public on $2 billion of trailing revenues. These guys are rumored to have $5 billion already — 3x that amount. Building AI is just a lot more expensive than building the things that came before. So I just think we need to have an apples-to-apples comparison. But if you want a champion to stand with you on this issue of companies raising too much, spending too much, they got to be really careful. I gave some advice to one of our fellow founders the other day. He's got money coming in over the bow at a multi-billion dollar valuation, and he's like, "We don't need it. We already have hundreds of millions on the balance sheet. Should we raise it?" I said, "Here are all the downsides of raising it. You can't go raise another $500 million and not have pressure from all your employees for employee secondary, for spending more money on more projects. The NPV on those other activities will be lower, and the incentive your employees have to stay with you once they sell $10 or $20 million worth of stock is going to be less." So finding that balance is a critical function of leadership of these companies.

Yeah, no, net-net, I agree with you. VC is harder, it's more competitive, but I think the structural changes we're seeing are more reversion to the mean than VC is forever bad. Listen, this has always been a hard investment category. If you're not backing a VC who has a right and a process that gets them into the top deals, then returns are not going to be great. It's a power law business, always been a power law business. I think these are super important issues, but why don't we — in the spirit of time —

No, I think it's another way of saying negative reflexivity. When dollars come in, it actually leads to poor behavior. We'll take a deeper dive on that sometime, because I also want to get your thoughts on how that might change. But why don't we finish, just in the spirit of time, with a quick tech check like we always do? Awesome. So the big event — and this macro world is more you than me — but the big event obviously was the Fed decision to lower by 50 basis points. What's the impact from your point of view?

Well, I'll just tell you where we are today. We talked about taking a bunch of units of risk off, which we did in June and July. As we were heading into these rate cuts, we're back at average levels of exposure today. That's because we saw a lot of really positive statements coming out of the earnings this summer, and the economic data continued to be constructive. We expected a 50 basis point rate cut, which I'd been sharing with you and folks on this pod for some time. We think that's a good setup heading into the fall. Now we have an election we got to work our way through, and a lot of other question marks, but you got to take those data points as you get them. Your team is listening to so many different earnings calls and whatnot. What's your take on the consumer and that side of the demand equation? Irrespective of inflation, is there a mini recession, or have we literally landed a soft landing, or do we not know yet?

Let's wrap it there. It's good to see you. Let's do it again next time. I know I'm going to be down in Austin, so let's get a poker game. All right, take care.

As a reminder to everybody, just our opinions, not investment advice.