All-In podcast host Chamath Palihapitiya on the current state of AI

[0:00] I want to bring in our special guest this morning. [0:03] He is sitting on set watching all of this play out. [0:06] Hasn't been here in a bit, and we're thrilled to have him back. [0:09] Chamath Palliapati is here. [0:11] He's the founder and CEO of Social Capital, CEO of 8090, [0:15] and host, of course, of the All In podcast. [0:19] Hello, Mr. Sorkin. [0:20] You're wearing a suit, no less, for us. [0:23] My entire job is selling enterprise offerings. [0:26] That's what I do. [0:27] That's what you do. [0:28] So actually, before we get into anything, [0:30] and I want to get into all what's happening in the Valley, AI, everything. [0:34] We're going to go back in time, talk SPACs, because people have asked. [0:38] I know your favorite topics. [0:39] But just your take on the IBM situation, [0:43] to the extent that you think it has ramifications in any larger way, [0:47] or maybe it's just specific to IBM. [0:49] I think the thing that we know is there are a handful of companies [0:54] that are compounding revenue a billion dollars a day. [0:58] And at some point, the downstream ecosystem has to make money as well. [1:05] And then the ultimate buyer of these tokens also has to make money. [1:09] And so at some point, the question has to be asked, [1:11] if two or three companies are going to generate two, three, [1:14] four hundred billion dollars a year, and it's doubling and tripling every year, [1:18] where is the rest of this money going to be made? [1:21] And who is implementing it in such a way where these profits are to be had? [1:25] And I think you're starting to see a little bit of the wheels come off. [1:29] Like, just at a very basic level, if you think AI is like oil, [1:33] let's just use this analogy to make it simple. [1:36] A barrel of crude is what, 86 bucks right now? [1:38] Right now, unfortunately, it's gone up. [1:40] Yes. [1:40] No, no. [1:41] WTI is 80. [1:42] WTI is 81. [1:43] 81. [1:44] Okay. [1:44] Let's say a barrel of intelligence, which is a million tokens, okay? [1:49] You can buy that barrel for 26 bucks from Anthropik's really good model. [1:55] You can buy from OpenAI for 26 bucks. [1:58] Anthropik's latest model costs you 56 bucks. [2:01] Elon is selling you a barrel of intelligence for a buck. [2:05] Zuck is about to sell it to you for a buck 50. [2:08] Demis and Sundar are trying to sell it to you for a dollar. [2:10] The Chinese will sell it to you for 50 cents. [2:13] So this rationalization has to happen. [2:15] We have the same input that has this crazy cost. [2:18] If you've made a bet very early around one of these folks that are selling [2:23] extremely expensive barrels of intelligence and you try to pass through the cost, [2:27] you may run into some downstream difficulty. [2:29] And that has to play itself out. [2:30] And you think that's playing itself out in the IBM story? [2:33] Or you're saying that's just a broader... [2:35] That's just a broader comment. [2:36] I think that Arvind, he's done a great job. [2:38] I mean, look, it's undeniable the trajectory of the business, [2:40] and he's been able to pivot it and orient it around cloud. [2:44] Right. [2:44] And I think he deserves a lot of credit for that. [2:47] But the reality is everybody that's in this ecosystem, me included, [2:51] we're all struggling to figure out how do we price this stuff [2:54] so that the ultimate buyer is making more money, is growing faster. [3:00] And that is still a question mark. [3:01] Well, I think you're getting to a point that Alex Karp, [3:04] who was just here about a week and a half ago, [3:06] was making about the sort of token maxing conceit, [3:09] which is that some of these companies are... [3:12] This is to the Anthropic Open AI piece of it, [3:15] where you have CEOs, and I've now talked to them too, [3:17] who say the math is not mapping for us. [3:19] The CEOs and the CFOs, in my opinion, [3:22] probably have no idea how much token maxing is going on [3:26] inside of their organizations. [3:27] I suspect what will happen is one day you're going to have a miss, [3:31] and EPS will be off by a few pennies. [3:35] And the CEO will say to the CFO, what happened? [3:37] Where did all of this incremental OPEX come from? [3:41] And they'll trace it to the $50 barrel of intelligence [3:44] versus the $1 or $0.50 barrel of intelligence. [3:47] That hasn't happened yet. [3:49] This is interesting. [3:50] There's a software stock. [3:51] What is also interesting, so it's a Dow component, IBM. [3:54] So the Dow is down 500 points. [3:56] Let's do the math, 50 points times whatever the divisor is. [3:59] The NASDAQ was up more than 100. [4:01] It is still. [4:01] So if it is a question mark about technology in general, [4:06] you're not seeing it yet in the NASDAQ. [4:09] And that big 500-point drop in the Dow is because... [4:11] I think, in fairness, the NASDAQ has a lot of other stuff, [4:13] which is all the hardware, all the memory, all the chips. [4:17] That entire complex is still going to make a ton of money. [4:19] The question is, again, eventually, if you buy oil, [4:24] you need to put it into a new engine. [4:25] That engine has to make you go faster. [4:28] It has to burn cleaner. [4:29] It sounds like you're suggesting that the LLM complex at the high end, [4:33] which basically is anthropic and open AI in terms of the pricing of whatever that is, [4:37] and that's still in the private markets, is going to break. [4:39] Is that what you're trying to suggest? [4:40] I think that's what you're trying to get me to suggest. [4:42] No, I'm not. [4:42] Let me be very clear. [4:44] The hardware and memory complex is going to run for another couple of years [4:48] because there are enormous constraints in that ecosystem. [4:51] So there's tremendous scarcity. [4:53] Those things are going to continue to do well. [4:55] NASDAQ will probably overperform the S&P. [4:58] Okay. [4:59] If you look at the only other pocket where there's true growth, [5:02] it's the LLM complex, as you said. [5:04] I don't know what the long-term valuation is. [5:07] But if you see a bunch of public companies in the next few quarters miss because of an OPEX miss, [5:14] and they trace that back to runaway spending that they didn't know existed inside of their organization, [5:21] you're going to take a little bit of the air out of that specific category. [5:24] Okay. [5:25] So we just talked about four of the big model makers, maybe five of the model makers. [5:29] We talked about OpenAI, Anthropic. [5:31] You mentioned Gemini. [5:32] We talked about GROC. [5:33] GROC. [5:34] And we talked about Meta. [5:35] How do you see them all stacking up longer term? [5:40] I think that you are seeing a convergence. [5:43] It used to be the case that when a model dropped, it was so superior to everything else, [5:49] you're like, oh, my God, we went from kerosene to Jet A, right, to jet fuel. [5:53] So, of course, I'm going to go to jet fuel. [5:56] Instead, I think the analogy now is we've seen the nth version of the iPhone. [6:01] We've all kept upgrading to it. [6:03] We're still using it for roughly the same behaviors. [6:06] My elbow hurts. [6:08] What should I do? [6:09] And you're wondering to yourself, well, why am I spending all of this money? [6:13] And I think that's where we are today. [6:15] So if you're going to direct all of this behavior, why don't you spend 50 cents or a dollar? [6:22] That's a pretty rational question. [6:24] But then the question is, is the Meta model, for example, or Grok, as good as or close enough of an approximation for whatever you think you need to get done using anthropic? [6:38] For most use cases, the answer is a screaming yes. [6:41] For some very narrow use cases, no. [6:44] But in those narrow use cases, I would say spend the $50 barrel of oil. [6:49] Spend it. [6:50] It makes sense. [6:51] If you're a cybersecurity company, if you're a Palo Alto Networks, and you can book billions of dollars of more incremental revenue by locking down all the infrastructure of big companies, use Fable. [7:02] But the freak out over the weekend, even inside of Anthropic, you know, we talked about it yesterday morning. [7:08] Anthropic, as you know, was planning to end the sort of free use of Fable for some of the paid users, sort of end that extension so that you'd actually have to upgrade or pay more to get access to it. [7:22] They can't afford to. [7:22] And then all of a sudden they say to themselves, we can't afford to do it because if we do, people will cancel their subscription and go to OpenAI. [7:30] Exactly. And the OpenAI model is excellent. [7:33] Sol is excellent. [7:34] Right. [7:35] Mythos, Fable, it's excellent when you can use it. [7:39] So all of these folks are hitting the same constraint. [7:41] They're massively power constrained. [7:43] They're massively data center constrained. [7:47] And so they are usage constrained. [7:49] Meanwhile, the guys that have all of that capacity, Google, SpaceX, and Meta, are finally getting their footing. [7:56] And they're starting to consistently release model after model after model that's 80% to 95% is good. [8:03] It's a very complicated dynamic for the leading labs. [8:07] Do you see a shift, by the way? [8:09] You know, there was, you know, for a long time, OpenAI and SAM were sort of under a lot of fire. [8:14] There's big questions about trust and the like. [8:16] And by the way, maybe even more questions given this Apple situation on Friday. [8:20] But there seems to be some dings taking place on the anthropic side, too. [8:26] And this maybe gets to the Alex Carpenter at all. [8:28] But people who believe that either the spending part, the math doesn't make sense of what's happening. [8:36] Or you look at the Figma story, for example, of them going into different businesses. [8:41] These are really concerning. [8:42] I mean, look, let's be clear. [8:45] There are three leaders of lab companies who have been in the spotlight for 20 years. [8:51] Elon, you've seen every facet of his personality. [8:53] And you can judge over 20 years of behavior. [8:56] Sundar and Demis, 20 years. [8:58] Zuck, 20 years. [9:00] And I think those folks have shown that they basically run a relatively straightforward, well-understood business. [9:09] Right? [9:10] There aren't any rug-pulling situations. [9:12] Maybe Facebook had some in the past. [9:14] I think they've cleaned that up. [9:16] Okay? [9:16] But Google certainly never has. [9:17] And Elon never has. [9:18] You may dislike them for other things. [9:20] And then all of a sudden you have these two guys who are incredible entrepreneurs, clearly. [9:25] But the amount of collectible data that you can look at, behavioral instances where they've had to make hard decisions, [9:31] are still relatively limited because they're only a few years into being this public and being this scrutinized. [9:37] So if you look at the first few years of Facebook, messy. [9:41] I don't know about Google. [9:42] Maybe the first few years of Google. [9:43] We're going to talk about you and your 20 years. [9:45] You're going to get to talk about it. [9:46] I mean, you know, like I was caricatured as the guy at Facebook that made everybody cry. [9:51] But then you look back, it's a $2 trillion company. [9:53] So the tears are crocodile tears. [9:55] Can I just say, IBM is saying that, you know, we just missed closing a couple of big deals. [10:01] That's, they always say things like that. [10:03] But this is interesting. [10:05] The clients that they were trying to close were distracted by industry-wide cybersecurity concerns. [10:11] And the stuff that they went on. [10:14] If you have somebody... [10:15] Do you believe that? [10:16] If you have somebody... [10:17] Now they're going to say the weather was bad. [10:18] No, I have a real issue with this. [10:20] And I've said this a couple of times, so I'll just say it again. [10:24] You can look at a graph of the financing needs of these companies. [10:30] Right. [10:31] And you can pinpoint when they oscillate between two different messages. [10:36] Message one, we've created a super god, which then the VCs are like lemmings. [10:41] They run and they're like, well, I need to be on the right side of the super god. [10:44] And then they flip to the other two, which is, this is a complete weapon. [10:47] Let's shut the world down. [10:49] And why don't you regulate everything? [10:51] They just go back and forth. [10:52] We've had this game. [10:54] Now the problem is, at the scale of a trillion dollars of market cap, [10:57] that game has these ripple effects that touch everybody in the industry. [11:00] This is why, by the way, I think what Alex Karp did, he deserves a medal. [11:05] This is an incredible human being who had the courage to come here. [11:09] Everybody watched that clip. [11:11] All these S&P 500 CEOs called me after that clip saying, help me interpret this. [11:16] And I'll tell you, he was on the right side of history. [11:18] Because what he's showing you is that, hold on a second, [11:21] like you need to have a much more predictable way of running your business. [11:25] Because if you are, you know, bending to the vicissitudes of private company funding cycles, [11:30] or private companies desire to need to live up to trillion dollar valuations, [11:37] you're taking a level of risk that you probably didn't know you were taking. [11:41] That's a fair critique, I think. [11:43] We've got to take a pause for a second. [11:46] Just one last question related to it, because it's on the Alex Karp topic. [11:50] Effectively, what he's saying is that Palantir needs, [11:52] you are going to need some kind of layer, though, in the middle. [11:55] I believe that, too. [11:57] Partially to protect you, I mean, he was arguing in some ways from the LLMs, ultimately. [12:04] Here's the thing. [12:04] The LLMs do a very good job of a very superficial form of privacy called zero data retention, ZDR. [12:11] That's the term that they use. [12:12] And they say, enable this, and everything's hunky-dory and everything's OK. [12:16] But when you look at the fine print, let's just say Joe puts in the secret formula for Kentucky Fried Chicken [12:22] into the model and say, help me optimize it. [12:24] That part is ZDR'd. [12:27] But then if you look at a little thing, like if he just clicks the Like button, [12:31] is there any guarantee that that information is not stored somehow, that he liked this kind of a thing? [12:36] And the honest answer, technically, is no, because we don't know how to do that. [12:40] The other interesting thing that I would point out to you is that when Alex came on, [12:44] I think he said something extremely legitimate. [12:46] I think it applies to companies. [12:48] It also applies to countries. [12:49] So at some point, we should talk about this. [12:51] And at no point did Anthropix say, you know what? [12:54] He's got it totally wrong. [12:56] Here's the exact technical details. [13:00] And so there's nothing to see here. [13:01] And I thought to myself, why wouldn't you, the minute after that clip started to go viral, [13:07] be completely specific and repudiate the claim? [13:12] And I think it's funny that we're here 10 days later and nary a word has been uttered, [13:16] other than Alex's version, which the reason is because we all know is the true version. [13:21] Thank you.