Ep6. AI Demand / Supply - Models, Agents, the $2T Compute Build Out, Need for More Nuclear & More

00:00:00.000 | If World War II never happened, the atomic bomb never happened, and someone just showed up in 2024

00:00:05.440 | and said, "I figured out this thing, it's nuclear fission," people would be like, "Oh my god!"

00:00:11.600 | Like, probably be more excited than you are about AI, right? Like, because it would solve all of

00:00:17.360 | our problems, and...

00:00:18.720 | [Intro]

00:00:31.840 | Hey man, what's going on? Nice hat.

00:00:34.640 | How you doing?

00:00:35.600 | Hook 'em horns.

00:00:36.800 | Definitely.

00:00:37.360 | I mean, is this a little promotion for the state of Texas we got going on?

00:00:41.200 | Yeah, yeah, I'm proud of the state of Texas, and a little bit, you know, I'm sad that the...

00:00:47.360 | especially watching the women's tournament quite a bit, and Texas just came this close

00:00:52.000 | to the Final Four, but the games last night with Iowa and LSU were just amazing. So anyway,

00:00:58.320 | it's been a fun time. So yeah, reminiscing perhaps a little bit.

00:01:01.520 | March is a good month. Well, speaking of reminiscing, I've been reminiscing, you know,

00:01:06.960 | I'm on spring break with my son, and we see these markets around AI getting... there's a lot of

00:01:12.400 | commentary about how frothy they're getting. And it reminded me of that question, you know,

00:01:17.840 | it's different this time. You know, people say the four most dangerous words in the investing

00:01:22.960 | universe, but yet every big breakthrough we've gone through in tech, it actually has been

00:01:28.720 | different, right? And so as analysts, as anthropologists, as forecasters, right, we have

00:01:35.360 | to try to sort this out, both in the short run and the long run, right? And it's hard to do.

00:01:40.320 | I mean, you do this bottoms up, you do this tops down, you study it. Lots of people in 1998 knew

00:01:46.560 | the internet was going to be massive, right? I mean, Henry Blodgett calling Amazon 400,

00:01:53.840 | right? People thought it was blasphemous, but it didn't stop us from having a boom

00:01:59.120 | and a bust along the way, right? And today Amazon's at 3000 bucks, almost 10X what Henry,

00:02:05.680 | you know, got shouted down for saying in 1998. And I think at the time he said it was a 10-year

00:02:10.400 | call. But the fact of the matter is trying to marry up the short and the long-term, I think,

00:02:15.680 | is really, really tough. Well, and I think it gets even tougher if enthusiasm builds, because

00:02:23.840 | that impacts the entry price on the marginal investment that one might make. And I think

00:02:30.000 | this particular moment in time is very, very difficult for investors that are looking at the

00:02:38.640 | marginal investment, because the prices infer some amount of optimism already.

00:02:46.080 | For sure. And so that's really what I think we're going to dig into a little bit today. Tap on this

00:02:52.480 | and go a little bit deeper. Maybe starting with this idea is there is increasing evidence that

00:02:59.360 | demand for training and inferences is maybe deeper and wider than we thought. You know,

00:03:04.160 | we're reading a lot of headlines about the world building ever bigger supercomputers,

00:03:07.920 | and then a lot of conversation about what those bottlenecks become. But why don't we just start

00:03:12.160 | with this question about why do we need bigger, right? And I guess first principles, you know,

00:03:19.120 | generative AI produces these tokens. These tokens are a proxy for human intelligence.

00:03:24.560 | And there's a lot of conversation about there's no limit to how much incremental human intelligence

00:03:30.080 | we want to buy, right? You and I've talked a fair bit about co-pilots for engineering. We talked

00:03:36.720 | about that with Dara, co-pilots for call centers. And in a pretty short period of time, these have

00:03:42.400 | become really ubiquitous development projects for almost every enterprise in just a few short years

00:03:49.520 | wanting these co-pilots. And now we're seeing a lot of conversation about autonomous agents. I

00:03:54.560 | mean, I think Benchmark is investors in Lang chain. Harrison Chase gave a great talk last week about

00:04:00.560 | what's next for AI agents that I encourage folks to watch around planning, UI, and memory. And,

00:04:07.760 | you know, we hosted an AI dinner last week, and there was an interesting conversation that came

00:04:12.480 | up just about a search use case within the enterprise. And in this particular company,

00:04:18.320 | the CTO took their million, well, 100,000 lines of code, their entire code base, about a million

00:04:24.560 | tokens, dropped it all into the prompt, and then just asked if it spotted any common bugs. And he

00:04:31.680 | said it found six, like non-trivial bugs in the code base. So the point is, we've gone from co-pilot

00:04:39.360 | to some of these maybe a little bit more autonomous systems, magic and cognition, all of this in less

00:04:45.440 | than two years. And let me ask you the question, if we compare this to 1998, like looking at just

00:04:52.800 | demand, you know, do you think this demand is as real as it appears? And what do you think are the

00:05:00.880 | most interesting use cases you're seeing from a demand perspective today? I think it's super

00:05:06.320 | difficult to unpack, not opting out of the question, but, you know, you and I went deep

00:05:13.200 | on the topic of full self-driving at Tesla and Model 12. And based on what we were told and what

00:05:21.520 | we learned, that's an application use case of AI that's pretty amazing, you know, coded differently

00:05:31.360 | than it had been for the past decade, and has results that people are claiming they can just

00:05:39.280 | feel is demonstrably better. And we've talked about the use case like coding. We've also talked

00:05:47.520 | about the fact that when you move away from coding, the efficacy drops a bit in terms of the

00:05:54.000 | productivity gain you might get. And I've watched, you know, certain companies say, oh, we're going

00:06:00.320 | to use AI, or I think they mean LLMs, but we're going to use LLMs to improve workflows in every

00:06:06.560 | aspect of our business. And when I've circled back with those people and say, did it change

00:06:12.080 | every aspect of your business? They come off of it a little bit. So, and one thing that I think

00:06:20.640 | is really different between the full self-driving data point, and it's not built on LLMs, it's a

00:06:28.800 | traditional use case of AI, which allows it to, I think, have much more finite input/output.

00:06:35.040 | And that example use case is fairly finite input/output, rather than this kind of open-ended

00:06:41.200 | conversation thing. So, I'm insanely curious, as I have always been. And so, in the two weeks

00:06:50.080 | since we've talked, you know, if someone has a use case that they've uncovered, and they're

00:06:55.680 | willing to share it with me, and I'll throw this open to the audience, like anyone, feel free to

00:07:01.040 | reach out. I'm infinitely curious. But one thing that I've seen that is causing me to pause a bit

00:07:08.720 | is a lot of the incremental work that's being done on AI projects are stitching together

00:07:19.440 | external databases with the LLM. And the LLM gets relegated to being an interpreter of the human,

00:07:27.840 | or to being an interpreter of the data back to the human, but is not involved necessarily

00:07:35.600 | in the data store, and is partially involved but not wholly involved in the decision-making.

00:07:42.800 | And on one hand, you could say, "Well, that's just how this is going to evolve." And maybe that is

00:07:47.760 | how this is going to evolve. On the other hand, for me, it questions whether we're still on the

00:07:54.400 | same evolutionary track we were with the LLMs. And look, there's all kind of open questions, like,

00:08:00.960 | "Will the LLMs that are being worked on at Anthropica OpenAce start to have a native data

00:08:06.880 | store inside of them?" Well, on the other side, we're seeing companies that have a ton of corporate

00:08:12.640 | data bolting their own LLM on top of theirs, and saying, "You don't need the traditional LLM

00:08:20.720 | vendors because we'll just allow you to query the data store you already have." So I think there's a

00:08:26.640 | lot of questions that caused me not to have the overwhelming confidence that it sounds like you're

00:08:34.000 | building, that it can go even higher from where we are today. And as you and I talk about all the

00:08:41.200 | time, like, I think about the fact that these entry prices are quite accelerated and bake in quite a

00:08:48.240 | bit of optimism. So one thing we've seen since you and I talked is a couple of, we'll just call them

00:08:54.960 | early falls from grace with inflection and stability AI. And when you raise money

00:09:03.920 | for a non-revenue company in the billions, it's pretty easy to lose confidence when you actually

00:09:15.920 | start having revenue, because all of a sudden, you can see the delta from what it's going to take

00:09:23.040 | to get to your next round. And in some way, I used to always tell people that valuation is discounted

00:09:32.720 | future expectations. They think about it as they won an award. What they should really think is,

00:09:39.600 | "Holy shit, we just set ourselves up." And that gets back to that point we were making before,

00:09:47.200 | which is Amazon 400 in 1998 was ahead of where the company was. The fact of the matter is it

00:09:55.680 | ended up being a bargain relative to the ultimate price that Amazon achieved. But we all know there

00:10:02.400 | were a lot of companies along the way that never achieved it. Yeah. And I mean, I've made this point

00:10:07.600 | over and over again, but I I'll state it again for the record, which is public companies structures

00:10:13.760 | are much more flexible for downward performance than private company structures. Right. And they

00:10:22.400 | just like, if you raise money, let's say your generic AI company, we call it company G, and

00:10:31.120 | company G raises money at $1.2 billion, and they have very little revenue. And 18 months from now,

00:10:38.080 | they need to be in the market again. And because they raise that much, they're burning, I don't

00:10:42.960 | know, what, $50 million a year? And the revenue's $2 million, $5 million, $10 million, $15 million.

00:10:50.000 | Like, all of a sudden, that looks very difficult to raise an incremental round.

00:10:55.120 | Well, I would I would say it's, you know, and again, I think inflection did some important work.

00:11:00.800 | I don't think it was a terrible outcome for the investors, for Mustafa, for Microsoft. But it does,

00:11:07.760 | I think, shine a light on one of the things you said. If you are trying to build your own LLM

00:11:12.640 | today, as inflection was, and you want to get your hands on 40,000 H100s, which I think they did,

00:11:19.520 | that's like, you know, digging a gold mine, right? The capital intensity of that undertaking

00:11:26.080 | is very different than starting a software company, right? And so the risk reward to

00:11:32.800 | both the founders and the risk reward to the investors changes a lot. I want to touch on

00:11:38.160 | something you said, you know, I was getting perhaps a little bit more bullish and excited.

00:11:43.040 | You know, I think I have a good company. I think I have a good company here.

00:11:49.280 | Remember, it was just not even a year ago that, you know, perhaps those of us who owned NVIDIA

00:11:56.880 | went from 125 to 250 or 300. And everybody yelled and said, you got to sell your NVIDIA.

00:12:02.320 | And I said, why would I do that? My numbers have gone up by more than the stock is appreciated.

00:12:07.680 | And the argument was that we pulled forward all the training demand models are going to get smaller.

00:12:12.720 | And that, you know, perhaps some of the production use cases are overhyped.

00:12:17.280 | And, you know, what we've seen from our bottoms up, both in terms of the training demand,

00:12:23.840 | okay, so this is training small models, medium sized models, frontier models,

00:12:28.880 | that is a lot bigger than we thought. And then the production use cases around inference,

00:12:35.760 | the co-pilots that we talked about, like, I can't think of another technology that became

00:12:40.000 | almost ubiquitous within enterprises, right? The need for it. Like if you're an enterprise with a

00:12:45.840 | call center, and you're not leveraging AI today, like what are you doing? If you're not leveraging

00:12:49.920 | co-pilot today, what are you doing? So in a matter of just a few years, this enterprise technology

00:12:55.200 | has become ubiquitous. And I want to explore two quotes that, you know, I came across recently,

00:13:02.640 | both from interviews that I think get to this point. And I want to try to break those down a

00:13:08.800 | little bit for where we go from here. But the first was from Jensen Huang. And he said, "You

00:13:16.080 | have to redefine the way you think about the traditional sources of demand for data centers

00:13:21.200 | and compute. The production at scale of intelligence matters to every single country

00:13:27.920 | and to every single industry." Now, the immediate reaction to this and to the quote from Sam Altman,

00:13:33.600 | I'm going to read here in a second, is of course, Jensen's going to say that he's talking his book.

00:13:38.720 | Right? But given that Jensen has more demand today than he can satisfy, right? There's no

00:13:46.320 | need to talk his book. And Jensen's been at this for 30 years. And I think he's been a pretty

00:13:51.280 | straight shooter for a long period of time. And so when you have somebody like him who's saying,

00:13:55.920 | "Listen, this is going to impact every sovereign in every industry." And then my own two eyes look

00:14:00.800 | at every single company, and every one of them wants a co-pilot for engineering. Every one of

00:14:05.600 | them wants a co-pilot for call centers. And the capabilities of these co-pilots is still on a very

00:14:11.760 | steep curve. It tells me that the use cases are unlike anything else we've seen since perhaps

00:14:18.880 | the database or the internet itself, where those became ubiquitous in every single enterprise.

00:14:24.800 | What do you think about Jensen's quote? Will this touch every industry?

00:14:30.320 | I think... Well, I already made a statement earlier that it is such an interesting question,

00:14:38.320 | too, because I really think you have to separate LLMs from AI. Now, Jensen's

00:14:42.880 | exposed to both, so it doesn't... One's a subset of the other.

00:14:48.320 | Explain that a little bit. Explain what you mean by that a little bit.

00:14:51.280 | Look, I'm always open to being wrong, and I've been wrong many times before. But I'll tell you,

00:14:58.080 | my current best thinking was just that LLMs were developed, starting with language translation,

00:15:06.560 | and then when the attention window part was added in, where context mattered, became very good at

00:15:12.800 | structured language. And it turns out coding is a subset of... And an even better subset of

00:15:21.200 | structured language, because it's more structured than language itself. And so, LLMs are very,

00:15:28.560 | very, very good at coding. Although I did see one tweet this week where someone said that the...

00:15:33.360 | They were saying the percentage of time spent on debugging has gone up massively because you don't

00:15:41.040 | even understand the code that was written, so to debug it takes longer, which would make sense.

00:15:46.800 | But anyway, as you... So, LLMs work really well against things that are language intensive. So,

00:15:54.320 | programming is a very particular use case of language that's even better for... And then

00:16:00.480 | customer support, you know, because the data content stores that have all the answers are

00:16:07.520 | very textual, and the user representing their problem is in language. And so, that one's a

00:16:15.200 | really good one too. And then, you know, people have talked about legal. There are others, but

00:16:21.600 | LLMs not going to do much for you on a manufacturing floor. The LLM part, AI might,

00:16:29.920 | but the LLM won't. And so... And clearly, NVIDIA is taking... There are going to be all these

00:16:36.400 | different models, vision models, omniverse, models of the real world, et cetera, that tackle the

00:16:42.160 | other problems that you're talking about. But that's also going to lead to massive TAM expansion.

00:16:46.560 | We met with... I spent time with Fei-Fei Li this past week, the incredible researcher, author,

00:16:52.960 | thinker at Stanford who's working on her own business, and that's going to require model

00:16:58.800 | training. So many people are still... And I would argue in the... Not only do I think it's going

00:17:04.880 | higher, Bill, I think we're in the very, very early phases, and we're going to look back at

00:17:10.480 | this period and say those were pretty incapable models. I want to talk about this Lex Friedman

00:17:15.280 | podcast, which I thought was pretty terrific with Sam Altman. But Sam, I think, referenced that,

00:17:21.360 | you know, the step forward in CHAT GPT 5 versus 4 is at least as big as 4 versus 3. I've heard

00:17:27.680 | this from others who've actually had some exposure to that. I also hear similar things about Anthropic

00:17:34.000 | where you have these agent capabilities that are, you know, built right in and become intrinsic,

00:17:39.040 | inherent in the model itself. And so I think it's not just language. I think we're going to have,

00:17:44.960 | you know, these real world models. Obviously, we've seen the video models, etc. And these

00:17:48.720 | things are just massively consumptive of data and compute. And so in that regard, I think,

00:17:54.880 | if anything, we're still probably underestimating where we go from here. But I will stipulate to you

00:18:01.120 | as venture investors, you're going to have a lot of roadkill along the way,

00:18:07.840 | right? You know, there were 20 search engines that went to zero, there are going to be 20 LLMs

00:18:14.160 | that go to zero, or many more. And I'm not even sure, right? If you said to me, where does value

00:18:21.920 | capture happen at the end of the day? Where is durable value capture in the land of AI? I'm not

00:18:28.800 | sure that LLMs, like if that's all you have going for you, I'm not sure that's going to be the

00:18:34.160 | winner. At the end of the day, you get paid by solving problems for consumers or solving problems

00:18:40.400 | for enterprises in a way that's durable and differentiated such that other people can't

00:18:45.760 | compete. And you can achieve, you know, some some dominant market share type pricing, some monopoly

00:18:52.160 | like pricing. And when I look at the commoditization of models, it seems to me that

00:18:57.680 | apart from the people who are on the very frontier, if you're on the frontier, and you

00:19:02.080 | have something totally different, it seems to me that that's a place where that is defensible. But

00:19:07.040 | if you're not on the frontier, man, it seems that these are going to be really fast depreciating

00:19:12.720 | assets that are going to be commoditized by the llamas of the world, other open source models.

00:19:17.760 | So I think you can have these two simultaneous truths. One can be that we're at the early stages

00:19:24.320 | in terms of the demand cycle for AI. But number two, that if you're building an index of LLMs,

00:19:31.040 | it may not be a very successful investment approach. Yeah, I mean, you raised so many

00:19:36.400 | different questions. I'll try and unpack a couple and then and then and then I want to go back to

00:19:42.240 | the same thing for a minute. But the the I really look forward to seeing five and seeing if I feel

00:19:51.760 | that way about it. I think part of what people are doing is extrapolating the feeling they had

00:19:59.600 | the first time they used chat GPT or the first time they had to write a letter, you know, and

00:20:05.600 | it just had this magical appeal to it. And then the next step was multimodal. And I would argue

00:20:13.440 | that didn't really land as well as people thought. And the first version of I think what you're

00:20:20.080 | calling agents were these third party attachments you could click and sign up for like that didn't

00:20:25.680 | really land. And so I what what I don't have the ability to understand and where this is where I

00:20:34.880 | could get it way wrong is. It doesn't appear to me that if agents are connectors to the external

00:20:41.600 | world, you know, to me, that's a lot like an enterprise when people would build connectors

00:20:46.560 | to other apps. And I just don't see it on this kind of exponential scale that we got, you know,

00:20:54.240 | as the LM was passing the LSAT, you know, you had each of these hurdles that it was driving through.

00:21:01.600 | And, you know, I don't I don't know that it'll feel that way. But if it does, then you I mean,

00:21:08.080 | this will happen. So when does five ship? Everybody thinks summer end of the year.

00:21:12.960 | OK, perfect. So today's April 2nd. So within six months, if it's so magical, then I want you to

00:21:21.120 | call me out and say, you see, but but but but we're playing a lot with hyperbole. And that's

00:21:27.920 | where I wanted to ask, which is like like just saying, oh, it's way better or oh, it's you know,

00:21:34.480 | I I have I it causes my skeptical meter to go up. I want to there's two things Sam did this in this

00:21:42.400 | recent interview that I think put him in the promotion hall of fame. And one of them was he

00:21:51.680 | read he juxtapositioned AI versus the smartphone market and then on the smartphone market as being

00:22:01.280 | small. He said, well, that's limited to five billion purchases every two years. And we're

00:22:08.400 | going to go way beyond that, which was I just thought genius. Right. Like like bring up something

00:22:14.880 | really big and then say degrade that and say it's going to be way bigger. It's genius. But the second

00:22:20.720 | thing he did was he said he's afraid they're going to run out of energy. Now, this is genius also,

00:22:28.960 | because the minute you leave that conversation and say, holy, how much energy is used by data

00:22:37.120 | centers, you've succumbed to the game, like because you've now accepted that this thing's

00:22:44.400 | going to be energy limited and you're going off and trying to figure out whether that's true or

00:22:48.640 | not. I think both of these things were and I it may sound like I'm kidding. I'm not. I think

00:22:55.200 | they're jobs and level promotional techniques like so much. So considering that I am on the other

00:23:03.680 | side of that from you, let's actually look at the quote and then let's take it apart in its parts.

00:23:08.800 | Right. He starts off by saying, I think compute is going to be the currency of the future. I think

00:23:12.960 | it'll be the most precious commodity in the world. OK, I'll give you that. That's that's nice prose

00:23:17.760 | and a little hyperbolic, perhaps. He says, I think we should be investing heavily to make a lot more

00:23:22.240 | compute. And then he he addressed this question of why are people skeptical of demand? I would say

00:23:27.520 | you're skeptical of demand. We have some other friends who are skeptical of demand. And he says

00:23:32.160 | people think they compare it to the market. So they think about the market for chips, for mobile

00:23:37.200 | phones. And let's say there's something I never did, by the way. I never thought about it that

00:23:41.440 | way. That's my point. But you know that seven billion people, they upgrade every two years.

00:23:47.440 | So the market's three billion systems on a chip. And he says, here's the important point, Bill,

00:23:51.600 | because it gets down to the elasticity of pricing. He says, if you make 30 billion chips on a system,

00:23:59.600 | you're not going to sell 10 times as many phones because most people only have one phone. Now,

00:24:04.800 | then he goes on to say, but what if you produce more compute? This is why it's so important.

00:24:10.400 | But compute is different. Intelligence is going to be more like energy, where the only thing that

00:24:18.320 | makes sense to talk about at price X is how much of this compute the world will consume and at price

00:24:25.680 | Y, how much it will consume. Because if you think about energy, right, and I was trying to come up

00:24:31.920 | with an analogy on this, Bill, if the application, let's talk about an application that we use of

00:24:39.040 | energy every day that we like, like a hot water heater. I like to take a hot shower. But if the

00:24:44.240 | cost of that was $100,000, right, not many people would take hot showers. But he's saying that if

00:24:51.920 | you drive down the price of the cost of compute, then the reflexivity is people will consume a lot

00:24:58.720 | more of it. Now, this is also known as the Jevons paradox, right? As price goes down, we demand more

00:25:06.800 | of it. The aggregate amount of consumed, of the compute consumed actually goes up, right? And

00:25:13.840 | that's really, you know, just a fancy way of talking about the elasticity of demand. And so to

00:25:19.760 | me, I think that is the way when people, when the traditional semiconductor chip analysts look at

00:25:27.440 | this market or data center analysts look at this market, they do, Bill, compare it to things like

00:25:32.880 | the smartphone market, et cetera. And they say, okay, well, how does this compare? Is it bigger

00:25:37.920 | or is it smaller? And I think the key point he was making, whether we agree or disagree,

00:25:42.960 | he was saying, if you produce it in abundance, right, there is going to be dramatic workloads

00:25:50.080 | demanded of that abundance since compute. And he used the examples, you know, whether it's,

00:25:54.640 | you know, tracking my email, doing very pedestrian things, running co-pilots for engineering and

00:26:00.000 | customer care, running these autonomous agents, or whether it's solving in the future,

00:26:05.040 | much more difficult problems like finding cures to cancer using some of these search and planning

00:26:13.200 | techniques, inference planning techniques that we're looking to in the future. So I thought that

00:26:17.840 | was actually interesting. But, so let me stop there, get your reaction because the second half

00:26:23.440 | of it I also think is important. Don't you agree that if we drive down the cost of compute, if we

00:26:30.320 | build a lot more of this, today we're clearly constrained that we're going to consume a lot

00:26:35.280 | more of it in the future. I mean, it's not limited by the number of chips on a smartphone.

00:26:39.680 | It's just not that interesting, Brad. Like if I drove down the cost of housing, I would produce

00:26:44.720 | more of it and people would use more of it. If I drove down the cost of gasoline, people would use

00:26:49.440 | more of it. If I drove down the cost of a robot, there'd be more automation. Like, of course,

00:26:54.720 | that's true. If an airplane cost one-tenth the price that it does today, we would fly more.

00:27:00.560 | That's just not that interesting. It's not provocative. It's not provocative.

00:27:06.560 | Okay. So let me tell you why I think it is provocative. We did drop the cost of flying

00:27:10.640 | on an airplane by one-tenth since 1971. And guess what? We got 10X as much travel. Okay. So all he's

00:27:17.520 | saying is that there, you know, today we're limited in the amount of these things that we can do. We

00:27:24.560 | can't build a multi-trillion parameter model today if you don't have the compute to do it.

00:27:30.160 | And so I do think that you're going to get to intelligence faster. You're going to get to more

00:27:35.280 | use cases faster, just like you got to more plane tickets and more hot water heaters, but we got to

00:27:39.920 | actually build out the infrastructure. The truth is, whether we believe it or not, it's very clear

00:27:46.720 | to me, and maybe we can move to this topic, it's very clear to me that enough people in the world

00:27:52.400 | believe, because I'm reading headline after headline about massive supercomputers that are

00:27:59.120 | being built, right, in almost all of the GCC countries. So Omnivan, Kuwait, MGX is backing a

00:28:06.480 | lot of efforts in the Emirates. The Saudis are doing the same. The French, a lot of talk about

00:28:12.720 | what's going on with the Singaporeans. The United States is out talking about their supercomputers.

00:28:17.680 | And just this week, we saw some headlines, I don't know whether or not they're true,

00:28:22.960 | about this project that Microsoft and OpenAI execs are drawing up on something, I think

00:28:28.640 | they're calling, what is it, Stargate? And again, I don't know the specifics whether or not that's

00:28:34.160 | true in '26 or '27, that they're going to spend $100 billion on a single supercomputer, but it

00:28:39.760 | does seem to me that there are enough breadcrumbs in the world today that people are making forward

00:28:45.520 | bets, that the need for more training and the need for more inference, there's sufficient evidence

00:28:52.160 | to cause them to put real money up against this. And so I think we're, you know, remember, we

00:28:56.800 | talked about a few weeks ago, that chart where Jensen said the five-year replacement cycle for

00:29:02.320 | data centers is not going to be $1 trillion, it's going to be $2 trillion. And then we'll reshow it

00:29:07.200 | here in the pod, and we broke down what that's likely to look like on an annualized buildup.

00:29:12.640 | In many ways, this is just a lot more commentary about the same, Bill, which is there are people

00:29:20.960 | placing orders, hyperscalers placing orders, enterprises placing orders, and now sovereigns

00:29:26.320 | placing orders that get you to a much bigger number than the traditional data center market.

00:29:31.600 | Well, then Sam's problem is solved.

00:29:33.760 | Well, I think it will be solved. There will be more capabilities brought online to train

00:29:41.360 | ever bigger models. But I think that he's suggesting, and this is, you know, I think

00:29:48.160 | as forecasters, the hard thing is, are these things going to be sufficiently demanding of

00:29:55.280 | compute and power that on current trajectories, we're going to fall short. And I think there are

00:30:01.040 | a lot of people sounding alarm bells on this, including, I saw a headline from the Biden

00:30:07.440 | administration, says the Biden administration wants to accelerate its conversations with big

00:30:12.480 | tech companies on how to generate more electricity, including with nuclear power, to meet their

00:30:17.520 | massive demand for AI computing. So on this one, I think we have a forecast. You know,

00:30:24.320 | if you look at the general forecast for power demand, it's been about GDP growth, right?

00:30:30.720 | Couple percent a year. And the fact of the matter is with renewables, we got sufficiently more

00:30:35.920 | efficient every year that you really didn't have to bring on a lot of new power generation to meet

00:30:41.680 | the demand that was going on in the world. But now if you look at some of these forecasts,

00:30:47.040 | and here's a forecast, I think coming out of semiconductor analysis, it's similar to a lot

00:30:52.960 | of the other ones. And it has data centers as a percentage of US power generation going from

00:31:00.000 | something like four percent today to something like, you know, 18, 19 percent in 2030. Right.

00:31:08.400 | So these are parabolic moves in terms of consumption that aren't satisfied with the

00:31:13.680 | current improvements that we have in terms of both power generation and our grid. You've been

00:31:20.080 | out front, I should say, on calling for the need to have regulatory reform as it comes to comes

00:31:26.720 | down to nuclear fission. A lot of people are talking about nuclear fission with respect to

00:31:32.080 | solving this problem. Are you a let's say that you believe the demand is going to be great,

00:31:37.040 | you know, and we just talked that there's some debate back and forth. But talk a little bit

00:31:41.440 | about what would need to happen from a nuclear fission perspective in order to meet some of

00:31:46.240 | this power demand. Yeah, I think there are better people like to talk about this than me. But

00:31:52.320 | when I look across the number of people whose intellect I respect and this came to my thinking,

00:32:01.680 | you know, maybe five years ago, and that's everyone from Steve Pinker to John Collison to

00:32:08.000 | Ilan to Toby at Shopify, like they all believe that nuclear is the very best way

00:32:15.920 | to get us past the climate change problem. And one of them, I can't remember which one,

00:32:21.600 | tweeted like if if if you're vocal about climate change and anti-nuclear, I don't have time for you.

00:32:28.320 | And and I I'm in I've been in that mindset. What what has happened and another person that's been

00:32:35.520 | remarkably outspoken on this is Josh Wolfe at Lux, who I spent time with at Santa Fe.

00:32:40.560 | And he even went in front of Congress and was talking about how much China's investing in

00:32:46.480 | fission. I think we have some charts we can show here, but their plan is to build 100 or 200 new

00:32:52.320 | fission plants while we're, you know, and if you go back three or four years ago, we were only

00:32:58.320 | decommissioning plants. We weren't. And of course, people talk about the horrific thing that happened

00:33:04.160 | in Germany where they were a leader in the market and turned them off and now are aggressively

00:33:10.240 | using coal and buying energy externally. And, you know, I think the best quote I heard,

00:33:17.840 | and maybe it was from Josh, it may have been from Pinker. He'll love that. I confuse those two

00:33:22.640 | was that if World War Two never happened, the atomic bomb never happened. And someone just

00:33:29.680 | showed up in twenty twenty four and said, I figured out this thing. It's nuclear fission.

00:33:34.640 | People be like, oh, my God, like probably be more excited than you are about. I write like

00:33:41.760 | because it would solve all of our problems and we let there become this negative association.

00:33:47.200 | Now, in the past three or four years, that's been flipping and there's been a lot. I think

00:33:51.680 | it's because of thought leaders like the ones I mentioned being outspoken that people have come

00:33:58.000 | to realize that we are we are it's it is the most energy efficient, the most energy dense,

00:34:04.640 | the cleanest thing we have available. It's way more durable because it can be put in so many

00:34:10.960 | different places that don't have geographic limitations, don't have sunshine limitations,

00:34:15.520 | don't have wind limitations, don't have transport limitations that some of the renewables have.

00:34:21.120 | And so it's been great. I would just say exceptional even to see this flipping of expectation about

00:34:29.760 | what's possible with nuclear and, you know, the the plant that didn't get shut down in California

00:34:36.320 | and then this Biden thing this week, which is all new, right? It's all and it's new in a different

00:34:41.600 | direction. So winds have been going this way and they're going this way. So it's fantastic. Now,

00:34:47.680 | despite it being fantastic, China is way ahead of us. And we'll we'll put a link in the notes of

00:34:56.240 | Josh Wolfe telling this to Congress, but they're way ahead of us. They're executing faster. And

00:35:02.240 | the biggest problem, as I understand it, once again, from talking to experts, not from my

00:35:06.960 | my own direct knowledge, but our cost of deploying new nuclear fission infrastructure

00:35:14.960 | is limited by our own regulatory framework, not by the technology. Or if you look at the

00:35:22.080 | cost differential between us and say China, that is the problem. And I don't know that we know

00:35:30.480 | how to do regulatory reform. And if I if I could snap my fingers and say, what would I want? I'd

00:35:38.240 | love for there to be a a zero based like like you do zero costing in an organization, a zero based

00:35:46.560 | regulatory rewrite of nuclear fission and one done without oil executives in the room. Now,

00:35:54.160 | I don't know how to make that happen. Well, you know, it's you bring up so many good points.

00:36:00.560 | One of the points is, you know, we've got this chart from our world and data

00:36:06.000 | that nuclear is both the safest and the cleanest source of energy. Right. And so if you just start

00:36:12.800 | there, you have to ask the question, like, how did we end up here? Like, we're all of the age

00:36:19.120 | that we remember Chernobyl and all the fear that got propagated in the world. One of the things

00:36:24.720 | that worries me about A.I. itself. Right. Here's a technology much like supersonic aviation,

00:36:32.080 | where scare tactics literally shut down all of the innovation in massive industries that would

00:36:39.760 | have inured to our national strategic advantage, the environmental advantage, right, geopolitical

00:36:47.440 | advantage. And we literally shut it down. And worse yet, there wasn't even a robust debate

00:36:53.440 | about it. It's just like, you know, the anti nuclear lobby declared victory and it was game

00:36:59.840 | over. And it's only what's interesting to me if you say, well, why, Bill, has have the winds begun

00:37:07.680 | to change on nuclear? I think the reason the Biden administration came out this week. Right.

00:37:13.360 | If you think that A.I. is on the front lines of the new Cold War about national economic security,

00:37:21.600 | national defense and offensive security, and if you think the limiting factor to pushing forward

00:37:28.960 | in A.I. is power, then all of a sudden. Right. You're sitting here worrying about China from

00:37:35.120 | the perspective of the video chips. You think you're ahead of China, but all of a sudden you

00:37:40.160 | say if the bottleneck is power, at least on nuclear, we're way behind, way behind on two

00:37:46.640 | dimensions. They have over 300 plants currently either being built or in development. We have zero,

00:37:53.280 | zero plants being built. I think 13 plants that are proposed according to this data that we have

00:37:59.840 | here and the cycle time to put up a plant. Right. If you if you do it as fast as humanly possible,

00:38:07.280 | something like seven to 10 years. Right. And that's that's assuming that somebody says go

00:38:13.280 | tomorrow. On top of that, if you look at some of, you know, just enriched uranium,

00:38:19.440 | I think a lot of the plants in Canada have been mothballed. We've got to spin those plants back up

00:38:24.320 | in Saskatchewan, whereas I think, you know, we have another chart on this. Eighty percent of

00:38:29.680 | the world's uranium is basically coming from Kazakhstan and is being shipped to the border

00:38:35.200 | between Russia and China. So like the U.S. has got to get serious about this. But if I'm right,

00:38:41.600 | that the lead time is 10 years, then all the stuff that we just talked about, the demand

00:38:48.000 | that we've got to satisfy over the next 10 years is not going to come from nuclear.

00:38:52.240 | It's got to be an all of the above strategy. But, you know, here's an FT article that just

00:38:57.520 | just came out. It said data centers, voracious power needs are set to rocket as cloud storage

00:39:03.120 | facilities, crypto mining and AI all strain the grids. Microsoft alone's opening a new data center

00:39:10.320 | globally every three days. These power hungry operations will together consume more than 480

00:39:16.800 | terawatts hours of electricity or almost a tenth of U.S. demand. So when you start to think about

00:39:25.120 | that, wind and solar are important. Costs are plummeting. They're clean, but they have baseload

00:39:32.000 | issues and they certainly can't scale at the rate I think that we need this to scale. So it seems to

00:39:37.600 | me that the power source that we're going to need to scale up is not gas. And part of the reason I

00:39:44.800 | think you're seeing a lot of this go down in the Middle East is because there are abundant sources

00:39:50.320 | of natural gas. The U.S. has abundant sources of natural gas. You can spin up a natural gas facility

00:39:56.320 | much faster, right, two to three years, very well-known technology, less regulatory headwind.

00:40:02.160 | But it seems to me that there's a lot of energy in Washington around national AI strategy. We need

00:40:08.480 | to have an equal commitment. Part of the same conversation needs to be a national energy policy

00:40:13.600 | that supports the buildout we're going to have to see. Well, and look, you could leave the AI part

00:40:19.680 | out. I think there are enough people who have such a strong view of climate risk that you could do it

00:40:26.640 | just for that reason. And let's say both, you know, OK, that's kind of a slam dunk. When I talked to

00:40:35.840 | our friend, Phil, who would be thrilled to hear your comments on natural gas, you know, he told

00:40:42.560 | me that the base load in the U.S. has been flat for many decades now. And so obviously the demand

00:40:51.040 | for energy has gone up over those decades, but it's been offset by efficiency gains. And some

00:40:57.520 | of that will play out here. Like the more you make an issue out of this and the more people like us

00:41:03.120 | talk about it is part of the market adjusting. And so you'll see some of that get corrected.

00:41:09.200 | One thing that set off a little bit of fear in the markets was Amazon went and did this deal

00:41:16.080 | where they bought some land right next to a nuclear facility. And once again, according to Phil,

00:41:23.360 | about 30, 25, 30 percent of the production in the U.S. is from independent energy producers who

00:41:29.600 | basically sell to whoever they decide they want to sell to. So it's a plant owned by a corporation

00:41:37.040 | that sells. And if you start to see data centers pick off the independents, that could put pressure

00:41:46.960 | on the grid in a way that might bring the regulators in to voice their opinion. So I

00:41:54.160 | think that'd be something to watch out for. Back to the nuclear thing, like I just, once again,

00:42:00.960 | if I could do anything, I would just encourage those in Washington, if they do get on this

00:42:07.600 | bandwagon, to commission, to build, to create a commission of people who have the best interests

00:42:14.960 | of the country in mind to think about reevaluating nuclear regulatory structure from the ground up

00:42:22.240 | and starting over from scratch. Because I suspect that the 10-year thing that you're talking about,

00:42:28.480 | in addition to the cost thing, are heavily impacted by regulatory. In fact, if you'll allow

00:42:34.560 | me, sorry to spin off of this, when I was thinking about this, I had a flash in my brain about this

00:42:43.760 | bridge that went down on I-95 in Pennsylvania. And I'll put a link in here. And for those of

00:42:49.280 | you that don't remember, I mean, now everyone's focused on the Delaware Bridge, but this other

00:42:52.960 | bridge collapsed. And within 12 days, it was replaced. And there's articles that say the

00:42:59.360 | governor's now a presidential candidate, like he's definitely going to get reelected because of how

00:43:04.080 | wonderful this is. And when you peel underneath that and read the articles about what happened

00:43:10.160 | is he went with no bid contracts. He was able to get people to work 24/7. It basically broke the

00:43:17.600 | rules of the regulations that were put in place that slow everything down. And if we've gotten

00:43:24.400 | to a place where we celebrate how quickly things can happen, and you've probably seen these things

00:43:31.120 | on Twitter about how quick the Golden Gate Bridge was built and how much it cost and all these

00:43:36.480 | infrastructure projects and how today they're like exponents of that. And I also thought about

00:43:44.720 | when they cleaned up San Francisco for Xi in like a week. The problem isn't that we don't know how

00:43:52.400 | to solve these problems. The problem is that we're our own worst enemy. We put in place the things

00:43:58.640 | that limit what we're capable of. Another article that I want to link to shows that renewable

00:44:05.040 | projects are way more successful in Texas than California, which is the ultimate irony, right?

00:44:11.760 | - But what does that tell you? What does that tell you, Bill?

00:44:13.920 | - It tells you that there's less regulatory mud in Texas than in California. And despite

00:44:23.280 | the intent of both the people of California and the people that I guess are representing them,

00:44:30.720 | the intent doesn't matter if the outcome's not possible. In fact, in some ways, you should

00:44:41.520 | hold representatives only accountable for output. Intent is kind of silly if it never achieves the

00:44:48.720 | output. And so anyway, there's a long way of saying that.

00:44:52.720 | - I think it's a super important topic, Bill, because when we look at great advances in

00:44:58.720 | innovation that have been stymied by regulatory excess, right? I talked about supersonic aircraft,

00:45:06.880 | right? Think about the productivity gains that would occur in the world if we were able to get

00:45:10.640 | from point A to point B faster. But literally, it was the environmental lobby that stymied

00:45:17.040 | supersonic aircraft. You think about calcium CT scans, which I've been promoting from the top of

00:45:23.040 | the mountain. Calcium CT scans cost $100 and save lives, can bend the healthcare curve around the

00:45:28.880 | 3 million sudden cardiac events we have on an annual basis in this country. But the reason

00:45:32.880 | it doesn't happen is they lost the lobbying game in Washington. You think about nuclear fission.

00:45:37.920 | I mean, think about the amount of dollars in energy the government is using to subsidize

00:45:42.400 | nuclear fusion, right? Nuclear fusion, which we all would agree would be a fantastic thing if and

00:45:51.120 | when it happens, but it's long dated, right? We have nuclear fission that has all these benefits

00:45:56.960 | that exist today. But again, for whatever reason, mass hysteria and otherwise, government regulation

00:46:02.320 | got in the way of that industry. And this is what we kicked off this pod talking about.

00:46:07.440 | And consumer perception, which you talked about. The hysteria around Chernobyl. Actually, I think

00:46:18.080 | you have data in here, but the number of deaths that happen per year from traditional fossil fuels

00:46:24.080 | is exponentially higher than the number of deaths we've ever had from nuclear energy. Yet, there's

00:46:31.360 | this weightiness to it. It's like fear of flying. The irony is they're not even fearful anymore. I

00:46:42.320 | mean, the data suggests that well over a majority of people understand that nuclear fission is safe.

00:46:47.440 | But what happens is it takes a long time for the government regulatory process to revert back

00:46:53.840 | to that pre-paranoia. So it takes some external force to cause us to change. Well, I think the

00:47:02.240 | external force that's causing it today is when we start forecasting ahead power demand needs

00:47:08.960 | to power all these things that we're talking about over the course of the next five to 10 years,

00:47:12.960 | we're going to have to take in all of the above strategy. And I'm glad Phil's going to be excited.

00:47:17.760 | I'm not sure the price of nat gas is going to go up that much, but I do think that natural gas

00:47:23.120 | is super plentiful. The mechanics of it in terms of building production facilities are super easy,

00:47:30.400 | well done. And I think it can come online in big chunks, like two, three gigawatt chunks,

00:47:37.760 | which we're going to need to power some of these big data centers.

00:47:40.640 | Yeah. And some of the graphs that you've included here show that the cost of producing

00:47:48.320 | new nuclear facilities in France has been going down and in the U.S. it's only been going up. And

00:47:54.800 | that should be remarkably upsetting to everyone. Just that core reality. And I'll pose a question

00:48:05.280 | to you. Do you think they'll go up or down in China? They're going to go down. Okay. You know

00:48:10.640 | that innately, right? Yes. Okay. And even worse, Bill, I mean, I love to visit France as much as

00:48:21.040 | the next person, but if it's not the height of embarrassment, if it doesn't prove the point

00:48:27.200 | better than any other point we could possibly make, the French are driving more innovation

00:48:33.280 | and more efficiency out of nuclear fission than the United States. And we pride ourselves on

00:48:37.840 | innovation, right? That should be a shock and a national embarrassment. It should. And at risk of

00:48:43.520 | being redundant, in case anyone who is listening has any authority whatsoever, I would just highly,

00:48:50.880 | highly recommend that there be a consideration of a zero-based re-regulation of the market,

00:48:57.840 | because it's clearly the problem. I don't think there's anyone in the energy market that doesn't

00:49:05.440 | think the bureaucracy and regulation is the problem in the American nuclear fission market.

00:49:10.960 | Well, I think that before we move to the market check, the one thing that seems to me is that

00:49:19.120 | we have increasing entanglement. You know, you gave this talk that one of America's greatest

00:49:24.000 | strengths was, you know, the fact that Silicon Valley is 3000 miles away from Washington DC,

00:49:30.000 | right? That's been a source of our innovation. And you and I both know 25 years ago,

00:49:34.400 | Washington wasn't all that interested in what was happening in Silicon Valley.

00:49:37.840 | Well, today, these are all matters of sovereign importance, right? AI is our national security.

00:49:45.840 | It is what's driving productivity gains, you know, in our economy. Technology is what's driving

00:49:51.440 | productivity gains in the economy. And now I look at power being integral to all of this,

00:49:56.880 | you know, whether, you know, the front lines of our battle with China, the hot war and the

00:50:01.520 | cold war, if you will, are now being fought on whether or not we're going to give them chips,

00:50:05.760 | whether or not we're going to let certain engineers work on certain projects, etc.

00:50:11.040 | And so if Washington wants to get serious, it needs to be an integrated,

00:50:15.200 | right national policy, you can't fix AI without also taking on our future energy needs.

00:50:21.200 | How about we talk a little bit just about a little tech check. Speaking of energy needs,

00:50:26.480 | I think I know somebody on this podcast, and it's not me, who was faster to invest in some of these

00:50:32.720 | companies that would benefit from energy policy. But we have a chart here for CEG. They're one of

00:50:38.240 | the leading nuclear companies in the United States. And that looks like a pretty good

00:50:43.600 | chart over the course of the past couple years, in particular this year. But yeah,

00:50:50.320 | I bought this stock. And by the way, I bought a lot of stocks that went down. So I think

00:50:55.520 | we got to be careful about cherry picking. But I bought this stock solely based on the bet that

00:51:00.640 | perceptions would shift on nuclear. This is a company that was carved out of a larger producer.

00:51:07.040 | They took the nuclear assets and made it a singular company. What I didn't know was that

00:51:13.840 | the AI thing would happen. And there are clearly people talking about this company. I saw a Morgan

00:51:19.920 | Stanley chart that listed everything about AI. And I saw their ticker. I was like, holy crap,

00:51:25.040 | like I didn't know it was an AI play. But once again, I'm happy that the nuclear thing has

00:51:32.880 | shifted. But certainly Wall Street is focused on this and this energy issue that you've brought up.

00:51:40.000 | As a reminder to everybody, just our opinions, not investment advice.

00:51:44.160 | But why don't we talk a little bit just about the Fed met since we last were on, Bill.

00:51:49.760 | And they update, the Fed every quarter updates their own estimates for things like GDP.

00:51:55.680 | December of last year, they expected our GDP to grow 1.4% this year. And now they're

00:52:00.880 | expecting it to grow 2.1%. So expecting our economy to be more robust. In December,

00:52:07.200 | they expected core PC. So this is their favorite inflation metric to watch was going to be 2.4%.

00:52:13.600 | Now they expect it's going to be 2.6%. So coming in a little bit hotter. And now the market is

00:52:19.360 | wrestling with how many rate cuts, if any, we're going to get. I'm still in the camp that we're

00:52:23.440 | going to have rate cuts ahead of the election. But the market is now pricing in, I think,

00:52:27.680 | only three rate cuts between now and the end of the year. And we see the 10-year just today in

00:52:32.640 | the markets is up to 4-4. I think we had bottomed at 3.5. So that's adding more restrictiveness

00:52:39.840 | to the economy. The markets, particularly high growth stocks, are starting to sell off again.

00:52:45.440 | Stocks that are levered are starting to sell off again. And so when we started the year,

00:52:51.680 | I think one of the big debates, we covered the AI debate, whether or not we had pulled

00:52:56.400 | forward all the demand for training, whether or not these stocks had gotten ahead of themselves.

00:53:00.320 | But there was a second question. Are we going to be able to avoid a recession?

00:53:05.360 | How many rate cuts we'd get? When would we get those rate cuts at the start of the year? People

00:53:09.520 | thought we'd get rate cuts in March or April. And so I think, as I read the tea leaves, I see

00:53:17.600 | pockets of weakness in the economy, particularly if you look at housing, autos. Tesla really missed

00:53:24.240 | their number today in terms of new cars shipped in the quarter. Some of that has to do with China,

00:53:29.520 | but it also has to do with demand here. So I think that that's worthy of watching.

00:53:34.880 | And then I've got three charts that we spend time looking at. One is a Morgan Stanley chart.

00:53:40.640 | So the first chart is the hedge fund net exposure. So this is how many of their dollars

00:53:49.280 | they have invested on the long side of the market in a particular sector. So this is in technology.

00:53:56.880 | So you can see that in October of '21 or the end of '21, the MAG 7 or TMT X, the MAG 7 had become

00:54:05.920 | 31% of its book. So what is that? That's technology if you exclude the largest names. And now that's

00:54:11.200 | down to something like 19%. And so hedge funds are starting to pull back a little bit on that

00:54:18.720 | part of the risk curve to your point as to whether things are getting a little too bubbly, Bill.

00:54:23.840 | The next one is people have seen this a lot. This is our software index. So where is software

00:54:29.440 | trading relative to interest rates? You can see that in the blue line is it's trading at about 6.1

00:54:36.960 | times. The 10-year average is 6.9 times. Rates have obviously moved up a lot off the bottom.

00:54:43.200 | But this is a long-winded way of saying that software is not really participating in this

00:54:48.320 | run-up. So software came down hard coming out of COVID, and it kind of remains there. It's

00:54:53.920 | bouncing around a little bit. Everybody's trying to figure out whether or not AI is going to

00:54:58.960 | catalyze a re-acceleration in software. Within software, kind of the data stack companies,

00:55:05.600 | so the Databricks and the Snowflakes and maybe even the ServiceNow and Salesforce,

00:55:10.080 | they've caught a little bit of a bid. But a lot of software is trading well below its pre-COVID

00:55:17.600 | multiple. And then the final chart, because I knew you would ask me, I just asked the team to

00:55:22.720 | pull together, get rid of Tesla out of the index, show me the MAG-6 index. So I just want to see the

00:55:28.560 | multiples for that relative to interest rates. And there again, you can see the 10-year average

00:55:34.160 | has been about 22 times. And so that is trading up. All of these companies have re-rated a bit.

00:55:40.480 | And the interesting thing here is if you look, you had like a reverse correlation

00:55:47.920 | between these two things, which is what you would expect up until about March of '22,

00:55:54.880 | where rates kept going up, but the multiple expanded on this group of companies.

00:56:01.760 | And so some people would say, hey, that's a warning sign, right? Rates have gone up,

00:56:06.640 | but these MAG-6 has kind of a 10-year high multiple. Shouldn't that be a warning sign?

00:56:12.960 | Shouldn't these companies be trading down a little bit? But I would argue to the reason

00:56:18.000 | that they're trading at those multiples is that people's confidence or their forecast for future

00:56:23.600 | growth is a lot higher than it's been at over the course of the last several years.

00:56:28.720 | Bill, you pointed out a tweet to me that was, I think, where the generative AI company

00:56:35.600 | multiples were trading. And it says OpenAI at 18 times, Anthropic at 20 times revenue,

00:56:42.240 | Glean at 56 times revenue, et cetera. Perplexity has a bunch of names on here. It has Huggingface

00:56:49.280 | at 150 times. It's just a question whether or not this is the prelude to a bubble.

00:56:55.520 | So I think this kind of brings us a little bit full circle. We started the pod asking the question,

00:57:00.800 | you know, you can believe that the long run in AI is going to be massive, but you can kind of have

00:57:05.920 | booms and busts, winners and losers along the way. When you look at these multiples for these

00:57:11.680 | companies, what do you sniff out? Well, I mean, the thing that made me forward this to you was

00:57:17.440 | that it lists them all and then it says, is OpenAI a bargain? So rather than the takeaway being,

00:57:27.600 | oh my God, these things are really highly priced, there's this relative valuation game,

00:57:33.280 | which is how bubbles are built. And because you adjust up, so you re-rate to a new norm, right?

00:57:42.640 | You know, I would just say, like, literally in, you know, since in the past, maybe not since we

00:57:48.000 | talked, but let's say in the past four weeks, a couple of things have happened. So we've had

00:57:53.520 | what I called these fast failures. You might, maybe the inflection team will get their money

00:58:00.560 | back, but that's not what venture is about, right? Like, so the 10X is off the table. That seems to

00:58:07.280 | be true at stability AI as well. You had a really interesting data point in the past week where

00:58:12.560 | I think people are abandoning the notion of the $20 premium personal AI tool. Perplexity came out

00:58:22.080 | and said that they were, you know, considering advertising and they, there were statements in

00:58:26.640 | their previous releases that were very negative on advertising. And then OpenAI said there was

00:58:32.560 | going to be a non-login version of ChatGPT, which would just inherently be cannibalistic

00:58:39.360 | to their $20, even though I think most of us believe that's over a billion run rate already,

00:58:44.400 | right? But if they're abandoning that, that's an interesting data point. Like the leading players

00:58:50.160 | are saying they don't think that's durable. Or it might be that, you know, the land grab is on,

00:58:56.240 | I want to sign up as many consumers to use my product as anybody else. And I'm willing to

00:59:00.560 | forego some of those early revenues, but I would agree with you. There's marginal cost here.

00:59:05.840 | I think the run rate here that's been rumored is that OpenAI consumer revenues are about a billion

00:59:11.280 | and a half. So if you look at that, you've got just under maybe a million customers using,

00:59:18.080 | you know, paying for ChatGPT. I do think I've said this many, many times, listen,

00:59:23.200 | I don't think a $20 a month fee for my consumer AI is going to be defensible. And the reason I

00:59:31.520 | don't think that's defensible is because they're going to be way too many people in the pool.

00:59:35.600 | Apple's going to have one, Google's going to have one, Meta's going to have one,

00:59:40.560 | ChatGPT is going to have one, go through the list, perplexity, et cetera. It only takes

00:59:44.880 | one person to give away a frontier quality experience, you know, at a much lower price.

00:59:50.320 | I'm going to say two things that are going to sound like they're at odds with one another.

00:59:53.840 | I agree with the current state of LLMs. Exactly what you're saying is true. I think if one of

00:59:59.840 | them can seamlessly integrate memory in a way where I can become reliant on it, I would gladly

01:00:07.680 | pay 20 and maybe more. A lot more. And so, you know, I think it's TBD. I think we got to see

01:00:14.400 | whether or not someone can, you know, everyone seems aware of the memory issue that we started

01:00:18.560 | talking about a while back, but that doesn't mean it's solved. And the architecture doesn't

01:00:23.280 | have an elegant solution. So it's the number one thing I'm interested in and looking out for.

01:00:28.800 | The last data point I'd bring up on this front, I did a tweet where I was curious why, you know,

01:00:36.240 | there's a 60X reduction from top model at a LLM company to their next model. 60 is just a lot.

01:00:44.160 | Like, you don't see this. Just so we understand what you're talking about. Like, the difference

01:00:50.240 | between, I don't know, maybe a CHAT-TPT4 and CHAT-TPT3 is a huge differential in price.

01:00:55.840 | 60X. And I would use this, you know, relative to your, like, that's not true in energy. That's not

01:01:03.040 | true in cars. I said to you, if new car production were limited, what would happen to used car

01:01:08.240 | prices? They would go up. So there's something weird here. Now, I got some good responses. One

01:01:13.600 | person said that there are throughput limitations on the high-end model. And so they're priced

01:01:19.360 | artificially high on purpose, which I could see that making sense. But the second thing people

01:01:25.360 | said was that the runtime models are just highly competitive. And many people think that some of

01:01:31.440 | the models are already priced under cost, which goes back to the credit investment theme that you

01:01:38.000 | and I have talked about. But if things are already that kind of hyper-price competitive, it's just a

01:01:45.040 | data point. Like, it's a data point worth paying attention to. So when I look at all those things

01:01:50.960 | and the entry prices that a marginal late-stage investor would be asked to pay, which are these

01:01:56.560 | multiples here, I'd be, like, nervous. I'd be nervous. I'll leave it at that.

01:02:02.160 | No, I mean, listen, it brings me back to the conversation about investing in Alta Vista,

01:02:08.000 | Lycos, AOL, et cetera, versus Google. I mean, the reality is we were in the fog of war,

01:02:13.840 | and we didn't know it. Every venture firm was paying up to get a search logo in '98 and '99.

01:02:20.000 | There were early revenue curves that looked promising for all of those companies. But the

01:02:24.240 | fact of the matter is we hadn't even gotten to the starting line. Like, we didn't figure-- like,

01:02:29.280 | we hadn't even determined who the market leader was going to be. And ultimately, Google, by 2004,

01:02:34.320 | 2005, it was very clear that they were going to capture a dominant share of that market,

01:02:38.400 | that the market was going to grow longer, be bigger. And so '98, '99% of all the profits ever

01:02:44.400 | created in internet search went to Google and went to them five or six years after, right?

01:02:51.760 | Search really emerged as a category. And so you at least have to leave open the possibility that

01:02:58.000 | that's the moment that we're in, that you look at these names on the list. And obviously, OpenAI is

01:03:03.760 | an early leader, is in pole position to be the non-Google Google, right? But the fact is we're

01:03:09.840 | super early, and it's not even clear that they're going to have sustainable economics in order to

01:03:16.240 | do that. But if you said to me, I look-- we passed on a lot of these companies simply because we

01:03:22.960 | couldn't get comfortable with the entry multiples, given how opaque it still is at this moment.

01:03:29.040 | But here's what I also believe. I believe the winners, the ultimate winners in AI have the

01:03:34.800 | potential to be way bigger than the winners of prior generations. And so you can afford to miss

01:03:40.720 | those early rounds, perhaps, in some of those companies. Not if you're benchmarked, perhaps,

01:03:45.520 | and your stock and trade is Series A. But if you're Altimeter, and we can-- like, guess what?

01:03:51.680 | Investing-- we had Bob Milad on last week. Investing in price line at a billion dollars,

01:03:56.240 | that was 120x in the public markets. Google also produced venture-like returns in the public

01:04:02.880 | markets. Amazon produced venture-like returns in the public markets. So I think the most important

01:04:08.800 | thing is-- But anyway, you were able to buy all of those at very reasonable relative prices,

01:04:14.080 | all three of the examples you used. Yeah, but I would also tell you-- I mean,

01:04:17.760 | at the beginning of the 100x in the public market. Yeah, but my mentor, Paul Reeder,

01:04:22.160 | used to walk into my office, and Google would-- query volume would go down a little bit,

01:04:26.320 | and he would be like, why are you so confident? This is 2005, 2006. And so it never appears clear

01:04:33.440 | at that moment in time, right? You had to believe in the size, the length, and the width of the

01:04:39.360 | market. And so all I'm saying here is that you've got to tread with caution, right?

01:04:44.560 | By the way, NVIDIA's already-- 2 trillion.

01:04:48.720 | Won. 2.2 trillion. Yeah.

01:04:51.680 | Like, that already happened. That's liquid today. It clearly has been the AI winner.

01:04:59.920 | 18 months ago, 18 months ago, NVIDIA was at $125 a share, and Meta was at $90 a share, okay?

01:05:08.720 | Meta's at $500, NVIDIA's at a huge multiple, $900 of where it was then. Markets shift very quickly,

01:05:18.320 | right? And then they forecast ahead levels of durability that may or may not exist.

01:05:23.760 | And so when you're at these moments, the beginning of a phase shift,

01:05:27.280 | we talked about-- you and I on this pod-- ERR, experimental run rate revenues,

01:05:34.320 | versus something that you believe is monopolistic, annually recurring.

01:05:38.720 | Those are two radically different things. One, you have no ability to forecast the future,

01:05:44.000 | and therefore it should have a one or a two X multiple because you don't know if it's going

01:05:48.320 | to occur again. The other one, if it's monopoly-like recurring, then you can give it a 10

01:05:54.320 | or a 20 X because you think it's going to last for a very long period of time.

01:05:58.320 | Anybody who tells me that these revenues deserve monopoly recurring like multiples today,

01:06:04.800 | that's a real struggle once you get to the late stage.

01:06:08.240 | I think it's a good lead-in for maybe a future podcast, which is I have this big question about

01:06:14.960 | whether the leading LLMs are going to incorporate their own data store, which if developers become

01:06:22.320 | reliant on, would greatly increase switching costs. Because today, the switching costs based

01:06:28.800 | on every developer I've talked to are-- we mentioned this before-- but about as low as

01:06:33.600 | you could possibly imagine. And if you're just using big context windows and not doing fine

01:06:39.360 | tuning, then you're just not-- you can bounce from one model to the next.

01:06:43.600 | And this gets back to that point I was making about are you just using the LLM for

01:06:49.360 | language translation versus are you relying on it for something greater?

01:06:53.600 | And so anyway, it's something I'm watching out for. Could someone build an LLM with an

01:07:00.880 | inherent data store that developers start to become reliant on because then they're in a

01:07:05.760 | different place than the ones are today? And answering the question in that tweet,

01:07:10.320 | is OpenAI a bargain? I will tell you, the only way I see you getting to switching costs

01:07:14.880 | is you have to develop something that is fundamentally differentiated,

01:07:18.960 | right, that gives you monopoly-like capability, whether because you've got memory and nobody

01:07:23.600 | else could do it, whether it's because it-- Maybe it'll be the agent you're talking about.

01:07:28.400 | I'm skeptical, but people that have seen them are excited, including you.

01:07:32.480 | But if I had to make a bet, you bet on the smartest group of people that's come together,

01:07:36.160 | that's delivering at the fastest cycle time. I mean, an OpenAI would certainly be in that camp.

01:07:41.280 | Anthropic would be in that camp. They're going to compete against these big hyperscalers.

01:07:45.680 | I do think the winner here is going to be massive, but you're making an excellent point

01:07:52.800 | that that could be true, and 90% of these other things could go to zero.

01:07:57.040 | So for venture investors who are paying up for things with that level of variability,

01:08:02.080 | that's a very different game than betting on a software startup with low capex,

01:08:07.200 | high predictability, $10 or $20 million in early indicative revenue in 2016 or 2017.

01:08:14.000 | Yeah. Hey, Bill, it's been fun as always.

01:08:16.960 | Take care. See you soon.

01:08:19.040 | Bye.

Ep6. AI Demand / Supply - Models, Agents, the $2T Compute Build Out, Need for More Nuclear & More

Chapters