The AI industry is shifting from a large-hub, large-spoke model to a large-hub, medium-hub, distributed-spoke model where enterprises will train and run their own proprietary models on their own hardware.

(00:00:00) All right, everybody, welcome back to the number one podcast in the world. It's your favorite podcast. It's your podcast's favorite podcast. It's the All-In Podcast, episode 279 with me. (00:00:12) Friedberg, Sachs, Chamath, the squad, the squad. We're here. It's the summer and we're ready to rock'n'roll. We got a power docket. We got a rocket docket. Palantir and Nvidia have announced a sovereign AI partnership. Where have we heard that term before? Palantir is going to use Nvidia's Nemotron, Nemotron, like the Pixar film, open models to build a custom frontier quality model to serve the US government. Palantir is calling this new platform (00:00:42) sovereign AI operating system. The US government agencies will own the hardware. The data and the model weights. Palantir also shared a viral tweet manifesto laying out the concept. Data retention is your treasure. Transfer it at your own peril. Transferring that data hands over access to your pre-existing winning plays and yields the means of production for new ones. CEO Alex Karp went on CNBC to announce the partnership in a classic Karp (00:01:11) Robin Williams' style monologue. Here's a clip from his 20-minute interview where he basically went after the frontier models like Anthropic. Play the clip. Our clients are just, to say they're unhappy with the Frontier Labs is to say I'm welcome at the Berkeley faculty. (00:01:31) It's like, there's just a level of discomfort and loss of trust. Sam and Dario, there's nothing more fun than debating Dario in private. So I'm not throwing shade at them, but something has gone completely wrong. And the basic view among enterprises in this country is, I'm gonna chillax and waste my time with tokens. I'm gonna get no value and they're gonna get my IP. When the Department of War goes to you and says, (00:01:59) I need this application. Do they get to control the weights to do it? Or do you get to control the weights? Are we really going to outsource the battlefield of this country to the consensus view in Silicon Valley? That is effing insane. All right. And so some folks refer to this as a televised nervous breakdown. We here at All In call that Alex Karp on a Tuesday. And we talked about this, Chamath, Sachs, Friedberg. We talked about this a whole bunch. (00:02:29) Back in February, I coined the term intelligent sovereignty here. Here's your victory fap. Do I want to give all of the secrets in our organization, every piece of intellectual property? (00:02:42) to Sam Waltman, who's got to make a billion dollars a year to keep up with his spend, right? He's going to build every application. I've been talking about AI sovereignty here for a bit, just in terms of how much more cost effective it is and how you're not training other people's AIs with your knowledge and your insights. This is why it's super important that open source, open source agents and local hardware be able to run these models and that consumers and companies learn how to roll their own language models. The intelligent sovereignty is different than privacy. (00:03:12) Privacy is, oh, you can't see my photos. You can't peek into my notes app and what I wrote there in my journal. Intelligence sovereignty is you can't tell me what to think. You can't use your AI to analyze my photos, to analyze my emails, to analyze my messages, and tell me how to interpret the world. That's actually going to be the next key piece. All right. And Sachs, you are in your long post era. Another long post this week. (00:03:38) from you on this very topic. And obviously, as AI czar there for the first half of the Trump administration, you've been very involved and very close to this. I would love to hear inside. I'd love to hear your take on this and your long posts, over 300 words. You can follow x.com slash David Sachs, but also like the palace intrigue here and what this means in terms of the relationship with the government, which we're going to get into in our second story. (00:04:08) look, J. Cal, I got to give you some credit there. The first part of your take was spot on. After that, it was kind of diminishing returns. I'm not sure why we had to listen to the next 30 seconds of it. But anyway, it started off really strong. But let's go back to this supposed crash out by Carp on CNBC. It was nothing of the sort. It was all these legacy media types making that claim. And that's the first clue that he's actually saying something insightful and maybe kind of brilliant. And I think the thing that he said that I hadn't really thought about in quite those terms, (00:04:38) firms is he started talking about AI safety in the enterprise and what that really looks like. And what he said is that what technical customers want is control over their compute, their models, their data stack, and their alpha, meaning their proprietary knowledge. They want to know they own the means of production, he said, and it's not being transferred to someone else. (00:05:00) And what he's referring to there is that these enterprises are at risk of transferring their knowledge, their know-how, their trade secrets, their customer data to these model providers who might eventually decide to compete with them, like you said, J. Cal. And you can see that enterprises are waking up to this threat and they're not happy about it. And I think Karp is exactly right about that. Now, I think this is a really interesting take on AI safety because (00:05:30) What safety means for an enterprise is, again, that they get to control their own data, their model weights, their compute. So A Frontier Lab can't hoover up their proprietary knowledge, their alpha, and turn it into their next product. And if you don't think that can happen, just look at what happened to Figma. So according to the information, Anthropic, quote unquote, blindsided its then business partner with the launch of Claude Design. So this was a new vertical (00:06:00) app that Anthropic launched to compete in the design category. And Figma's founder said that Anthropic had not been completely honest with them. Anthropic's chief product officer had actually even served on Figma's board and didn't resign until three days before the launch of Claude Design. So obviously, Figma again felt blindsided by this. And you can see the resulting impact on their stock price. Figma's stock has fallen something like 50% this year, while Anthropic's valuation has surged. (00:06:30) This is not an isolated example. Anthropic has also launched Claude Science, Claude Security, Claude Legal, Claude Financial, and of course, Claude Code. And every single one of these vertical apps expanded into categories that was previously served by companies building on top of Anthropic's own models. And really, if you want to go back to when (00:06:52) Anthropic's revenue explosion began, it was with the launch of Claude Code. And how did they know to launch that product? Because they saw that Cursor was doing extremely well. Cursor was one of their biggest customers. They created the coding assistant first. They created that category. And Anthropic said, oh, like why don't we vertically integrate? So in other words, they're watching where the value is being created on top of their models, then they're moving in directly. (00:07:19) And this is a formula that I think is very Microsoft-like. You could say it's very Google-like. They want to dominate the model layer. You could call that the operating system and then use that position, that monopolistic position to capture the most lucrative verticals. So if you want to think about like the Microsoft example, they had the Windows monopoly and then systematically they went and dominated every (00:07:42) lucrative category of business software. It started with spreadsheets and word processing, and then eventually it went to the browser, so forth and so on. If you want to look at Google, they basically had a monopoly or dominant position in search. And if you go back to the early days of Google, the search results kicked you off site. And in fact, they really pride themselves on how quickly they could send you off site. But gradually over time, they used that traffic to tell themselves where to build properties. And today, fewer than half of search (00:08:12) searches kick you off site. You stay on Google properties. And I think something similar is happening with Anthropic here. The pattern is clear. They are going to use their dominant position in the model to then grab more and more territory in any interesting and lucrative vertical. So again, back to Alex Karp's point, if you're an enterprise customer or a developer, why in the world would you ever want to share any proprietary data with them? You are mortgaging your future. You are sealing your fate. You are going to lead to disaster for your company. Just one last point. (00:08:42) and I'll turn it over to Jake Cal, is that Dario, at the same time that they pursue this business strategy, has been arguing that open source models are dangerous and need to be restricted. Well, dangerous to whom? Not to enterprises that want to retain control over their data is dangerous to his business model because this model requires that customers don't have a lot of choice at the model layer. And what Karp is pointing out here is that if you want to have true AI safety, (00:09:11) As an enterprise, you have to retain the ability to choose at the model layer who gets to see and use your alpha. Yeah, this is well said. I think you picked that carcass to the bone a bit, but Chamath, you're actually doing specific examples of this at 80-90. You've been testing some of the open source models. I saw you share that on. (00:09:34) Twitter X. So maybe you could give us a little feedback on what you've learned as the CEO of 8090 and your first-hand experience now with using open source for the first time in the last couple of weeks for this specific use case in enterprises. When you start a company, I mean, you guys all know this, you're not starting it for the moment that exists today. You're almost sort of trying to forecast (00:09:57) if such and such a set of things happen, then here is the scene that gets created because it takes time to build something and it takes time to get enough reps to know what you're doing and you go to market. This for me was the moment that I thought would arrive, which is the point where everybody wakes up and realizes, wait, hold on a second. Two things are true. The first is that (00:10:20) My business is complicated. I want AI to be able to accelerate it, but I want to be able to protect myself in doing so. And then the second is I want the flexibility where there's an independent third-party control plane that I use to get all these benefits so that I don't leak and cede my advantages away. And I think Alex is an incredible (00:10:41) smart, brilliant guy, and he completely nailed it. And I think he called out on its face the huge risk of this. So let me just give you this narrative in three tweets. The first one is I read this really interesting study from BCG, and what they looked at was the return on capital employed, or ROCE, of various businesses. And this is what's incredible. The cost of capital has now, with long-term rates, (00:11:12) moved back to what its long run averages, which is around 8 to 11%. What that means is like that is the actual cost that you would borrow money at effectively. The problem is that half of large US companies now cannot deliver returns that exceed that. is a really big problem. And then second, there's a further problem, which is that persistently low returns. So in the, you know, 1, 2, 3, 4, 5%, (00:11:38) is about one in seven companies all around the world. Okay, so why is this important to note? It means that being in business is complicated. It's hard. Not everything works all the time. There's a bunch of underperforming businesses. There's a bunch of underperforming segments. So in that lens, when you think about what Sachs said, which is you have this company that comes to you and says, I have a magic box. And all you have to do is tell me everything you're doing, and this magic box will make everything better. (00:12:08) But then all of a sudden, from the shadows, the magic box says, you know what? I've decided to compete with you. That is a huge risk. And now that you've seen enough examples of it, I think you have to figure out a different way to do it. So then you go to the next tweet that I saw, which I thought was interesting. And this is a woman who's an ex-meta PM. And what she essentially says is like, hey, hold on a second. There is this assumption that you can't use an open source model because it's all Chinese. (00:12:36) And what she says was, even if it's 100X cheaper, and their response is, no, because we care about safety and security. And her perspective is, don't you understand that you can actually host open source models with your own GPUs in US data centers that doesn't share any data back to anybody? And instead, what you're accidentally or purposefully doing is giving away all your data to a couple of frontier labs rather than owning it privately yourself. And 100X, it turns out, is a really big number to pay. (00:13:06) to do all of that by accident. And that's what Alex Karp is saying. He's like, why would anybody do this when there were alternatives? And so the third post that I'll talk about is something that we did. So we took our software factory, which is an agnostic third-party control plane, and we just wanted to see. And we ran it on a very typical enterprise task. (00:13:30) which is you have an old piece of code, you want to migrate it, and you want to maintain it in a new framework so that it's easier and more flexible. Pretty straightforward task. And so we ran it, and we ran an experiment where we did Claude by itself, then we did us plus Claude. And then we ran it on the best frontier open source model, and then us plus that model. And the data is crazy. So when you use our harness, (00:13:59) with Claude, it was simultaneously 1.4x cheaper and 1.5x faster than just using M-Propic Opus 48 alone. But if you wrap the open source model with our software factory, it was 16.4x cheaper. Now it was three times slower, but you know, you're talking about a couple of extra hours to save 16.4x. (00:14:23) On that one, the slowness, Chamath, is that slowness because of the hardware being served up by Claude, or is it this all? This is all using Open Router. No, this is all using Open Router on a very traditional hardware stack. So look, I think the reality is, could that be optimized even further? Absolutely. But my point is, if you take Sax's points, and then if you take Alex Karp's point and just this actual data, there is a very legitimate question, which is, if you are a reasonable company, (00:14:50) Why are you not finding an independent way to access this intelligence in a way that doesn't leak your edge away? To do so at this point now is kind of becoming derelict. (00:15:01) and irresponsible. Back then, you could be experimenting because you didn't know any better. But now when you know all of these data points, to continue to make the same decision, I think is really insanely dumb. All right, Dave Friedberg, you are also a CEO of the surging Ohalo, and you have a lot of proprietary data. Let me ask you point blank, do you trust your data to the frontier model companies, or are you doing what? (00:15:28) There seems to be consensus here, protecting your data, protecting the crown jewels, so to speak, and using open source, are you experimenting with it? And just yes or no, do you trust the frontier models with Ohalo's data? So I'll tell you, there's been an effort by Anthropic to go around and sign up life sciences companies to contributing to a new life sciences focused model. (00:15:53) That effort has been, they're approaching these large companies with large proprietary data sets and saying, hey, if you share your data, we will give you early access, some sort of proprietary value, sign this NDA, and you can participate with us. And I think nearly everyone I've spoken with has woken up to the fact that they are basically trying to commoditize everyone's business. Because fundamentally, if all of the 10s of billions of dollars you as a life sciences company (00:16:19) have invested in experiments and product development, and you've generated all of this proprietary data along the way, that data is a true asset of your organization. It's an asset that you've spent billions of dollars developing. And by handing it over to a model company to then combine with other people's data, you are effectively commoditizing the asset that you have, the one kind of core differentiation that you have. And so everyone is largely saying no. (00:16:47) The way I see this evolving is very much in line with what Alex Karp suggested on CNBC. If you go back a couple of years, I think we all assumed there was going to be this large hub, large spoke model for AI model development and deployment, meaning there would be these very large clusters, (00:17:06) these large clusters would be ultimately capital advantaged. So those who had the most capital, which is why everyone's raised 10s and hundreds of billions of dollars, would be able to train models. And then there would be these large spokes, these large clusters for deploying those models with inference. So everyone's using the neo clouds and the hyperscalers and whatnot to run models. And then maybe they've got their own proprietary data layer that sits in front of that. But I think what everyone's realizing is they're better off (00:17:33) developing their own weights and their own models using either an open source basis or there might be some intermediary business model that evolves, meaning there will end up being several large hubs that do all of the core foundational model development, then smaller hubs, meaning like clusters for training, that enterprises will use to train and develop their own proprietary advantaged models using their own data. (00:18:02) And then there will be these much more distributed spokes because I think everyone's also realizing the value of on-prem. By putting a set of servers and building a cluster in your own data center or even in your own enterprise IT closet, you can run a lot of the workflows that you're using AI for your enterprise locally. And so I think the model is shifting where we're going from large hubs, large spokes, to large hubs, (00:18:28) medium hubs, and then a distributed spoke model, where there will still be neo-clouds and hyperscalers that are being used for inference, but people will also have their own inference instances that they're going to run for their own enterprise setting. And everyone, I think, is walking this path, and they're going to walk this path over the next couple of months, because they're realizing quite quickly that in order to compete in a world of commoditizing knowledge and commoditizing (00:18:53) capabilities, you have to leverage the core differentiating assets that you have, which means you have to build your own models, and you will likely end up having to run your own inference with your own proprietary models. (00:19:03) And to just give a little bit of texture to how that's going to look, if you follow the Microsoft example, which we've talked about here before, Lotus 1, 2, 3, WordPerfect were their partners. They were replaced with Excel. They were replaced, obviously, with Microsoft Word. You don't even know those other two, WordPerfect, WordStar, Lotus 1, 2, 3, VisiCalc. That's exactly what they have to do. And they have no choice but to do that now because they have a trillion dollar market cap. They must win the application layer. And (00:19:33) Sam Altman went to Y Combinator and he said, we'll give you $2,000,000 worth of free tokens. And I came out and I said, listen, there's nothing personal against Sam. You know, Sam's a very aggressive deal maker and he wants to get access to those startups because he knows having run Y Combinator, that if he can get their innovations, those founders latest thoughts about what's around the corner, he can incorporate them into the platform. There is no free pizza. There's no free beer. When somebody (00:20:00) like Sam Waltman comes to you and says, here's some free tokens. (00:20:03) Your alarm should go up. (00:20:05) Zuckerberg did the same thing. (00:20:06) He said, hey, I'm going to give people a bunch of access. (00:20:08) I'm going to give them money. (00:20:09) Come to the Facebook platform. (00:20:11) Nobody who went to bed with Microsoft in the 80s, Facebook in the 2000s, or Sam Altman now in the 2020s, did not wake up with their throat slit. (00:20:20) This is a message to founders. (00:20:22) If you partner with any of these people, they will slit your throat and take your business wholesale. (00:20:27) There is nothing to discuss here. (00:20:28) Don't trust them. (00:20:30) Use your (00:20:30) I don't think, I think it's less salmon OpenAI, to be honest. (00:20:36) I think that the diversity of OpenAI in terms of its revenue streams and specifically its consumer business may actually be its savior. (00:20:44) In a relative value basis right now, OpenAI equity, I think is more reasonably priced than entropic equity. (00:20:52) And the reason is not because of the quality of the models or the teams, because they're both excellent teams. (00:20:58) But the reason is that (00:20:59) OpenAI can fall back on a really healthy consumer business. (00:21:03) Sure. (00:21:04) The difficulty that Anthropic is going to face is that I think what Sachs said is true, that they have lost this fundamental trust about being able to stay within their sandbox. (00:21:15) And if you consistently demonstrate this tendency to learn and then to try to disrupt your host organism, (00:21:23) eventually you get sort of pigeonholed and you get cornered and people find ways to work around it. (00:21:30) And so, look, I sent that text that I had about the, or that post that I had about our testing of our harness on these Chinese models to somebody well-known in the industry. (00:21:42) And he says, look, if you also add some (00:21:45) post-training with all of the telemetry that you're going to get from the harness itself. (00:21:49) He's like, I suspect you'll find that it gets as good as Mythos. (00:21:54) And I thought, well, if that's true, then why don't I just take GLM, control it entirely soup to nuts on my own hardware inside of the United States with only US citizens that can touch it. (00:22:06) just seems like the brain dead obvious thing to do. (00:22:09) And it's much, much cheaper. (00:22:11) Yeah, and 100% correct. (00:22:13) I will say, (00:22:16) If you're going to do this, you'll eventually wind up rolling your own LLM. (00:22:20) I mentioned a company, Abacus, goabacus.co, that we seeded in our accelerator. (00:22:25) What they're doing now for HIPAA client people is they're actually giving you this go on box. (00:22:31) They're literally saying, we're going to make your own model for you. (00:22:34) So once you start this, you start with, you know, Claude and use their wrapper and everything for OpenAIs. (00:22:40) Then you move on to the next step. (00:22:41) The next step is I'm going to use an open source model, use my own (00:22:44) try to find a harness, et cetera. (00:22:46) Where you will eventually wind up is you're going to fork these models. (00:22:49) You're going to build your own. (00:22:50) That is the end state on-prem, on your own hardware. (00:22:53) Don't trust anybody because there's too much at stake. (00:22:57) You cannot risk your entire business. (00:22:58) You might as well be part of the vanguard and start investing here. (00:23:02) You're going to slow down to speed up is what's going to happen. (00:23:05) Whether you're using 8090, Abacus, or any of these other solutions or rolling your own, you must have AI sovereignty, you must have intelligent sovereignty, or you're just giving your business over to your company. (00:23:14) competitors. (00:23:15) The only company, the only company at scale that's ever respected the developer community in this way is Apple. (00:23:22) Apple took a very strategic approach to wanting to build an app store business, to wanting to support developers, and they explicitly told people, if you make something super obvious that you can build in a week or two, (00:23:32) It eventually might wind up in our basic collection of apps, the stock app in your iPhone, the notepad app in your iPhone. (00:23:40) Those apps are incredibly basic, but if you looked at, you know, Robinhood's or Google Finance, and then you looked at something like, say, Evernote back in the day, which was an advanced note-taker, it took 10 years for Notepad to add the features that Evernote had 7, 8, 9, 10 years ago. (00:23:56) They specifically slowed their apps down to make them simple for users and not F with (00:24:03) their ecosystem because they want to take the 30% tax. (00:24:06) That's your choice here. (00:24:07) There is no 30% tax here when it comes to Anthropic. (00:24:10) There is no 30% tax equivalent with OpenAI. (00:24:13) The thing with Apple is that Apple was renting distribution. (00:24:17) This is not renting distribution. (00:24:18) This is where you're renting intelligence and judgment. (00:24:22) And so the problem that a company has is you can't rent the same. (00:24:27) And this is why your point is actually right. (00:24:29) I think it's for a different reason. (00:24:31) You can't rent intelligence from the same place that rents it to your competitor. (00:24:35) Correct. (00:24:36) You just can't. (00:24:37) You can't. (00:24:37) It's just stupid. (00:24:38) It will overflow. (00:24:39) It will go over the wall into your competitor's lap. (00:24:42) It becomes the lowest common denominator problem where you and your competitors now look exactly the same. (00:24:47) Why would you do that? (00:24:48) And again, if you go back to that thing that BCG identified, so many companies are already teetering on a very difficult position where they cannot generate returns on their invested equity. (00:25:00) And so why would you then go and pay all this money so that you end up with the same answer as your competitor? (00:25:06) You can't do it. (00:25:07) It's just not a choice. (00:25:08) And Sachs, part of what's going on here is the deflationary nature of technology. (00:25:13) Every single one of these tools is getting cheaper. (00:25:15) Tokens are getting cheaper sacks. (00:25:17) And if you look at Nvidia's role in all of this, they have something called Nemotron. (00:25:23) You can try it if you use perplexity. (00:25:26) You can just pick the drop-down menu. (00:25:27) It has deep thinking. (00:25:28) You will not be able to tell the difference between Jensen Wong's open source LLM sacks and Claude for 95% of your searches, I guarantee you. (00:25:39) Now, why? (00:25:42) has Nvidia and Jensen downplayed their open source model until this moment? (00:25:46) Why would he do that? (00:25:47) Why would he never bring it up in the all interview? (00:25:49) Never bring it up because his top customers were very concerned, from what I understand, about the fact that they had made so much project progress on their open source model. (00:26:02) But suddenly, after OpenAI announced their jalapeno chips, (00:26:07) after Anthropic started making chips, after AMD did successful projects with both of these companies, after Elon said he's going to do his own fab, Nvidia's taking the gloves off, David. (00:26:19) They are going to own the whole stack. (00:26:21) They are going to be talking about open source a whole bunch. (00:26:23) So maybe you could talk a little bit about Nvidia and their role in this as the open source, at scale, full stack provider. (00:26:31) You get the hardware from them and you're going to get a model that's competitive with open AIs for free. (00:26:38) and all you have to do is use one of their hosting companies, Cora, we have whoever's buying Nvidia, Colossus, et cetera. (00:26:43) What are your thoughts on Nvidia suddenly being willing to talk about their open source projects today? (00:26:49) Well, look, I think part of it is they need a time to make the offering compelling. (00:26:54) And, you know, they're up against some pretty great AI labs. (00:26:57) And I think some of it is just, hey, it takes time to train up these models. (00:27:01) Now, one question is, why is it that Nvidia and Palantir are partnering? (00:27:08) Like what makes them natural partners? (00:27:10) And I want to just explain that. (00:27:12) If you want to think about the AI stack for a minute, at the most basic level, there's three layers to the stack. (00:27:18) It's basically the chips, it's the models, and then it's the applications. (00:27:22) right now we have in the middle layer, at the model layer of the stack, you've got two dominant companies. (00:27:28) You've got Anthropic and OpenAI. (00:27:30) We know that Anthropic's around 60-something billion of ARR. (00:27:34) OpenAI is at 40-something billion of ARR. (00:27:36) As far as we know, no one else is really generating meaningful revenue at the model layer. (00:27:41) So we already have, let's say, an emerging duopoly situation. (00:27:46) We have Anthropic pushing for a regulatory capture agenda that would probably enshrine that duopoly situation at a regulatory level because they're pushing for a safety agenda where Dario explicitly says that these other models are not safe, you shouldn't have access to them. (00:28:03) So you've kind of got that situation. (00:28:05) You've got the market producing duopoly, you've got the government now potentially leaning (00:28:09) Not to bust up the duopoly, but maybe to enforce it. (00:28:12) So that's sort of the emerging situation at the model layer. (00:28:15) And so if you're an application at the top of the stack, like Palantir, or you're a chip company at the bottom of the stack, that's the last thing you want. (00:28:23) You want a competitive model layer. (00:28:24) Why? (00:28:25) Because if you're an application, you don't want to be beholden to 1 model provider, right? (00:28:29) You want to have a choice. (00:28:30) And if you're an enterprise, you want to have a choice because you don't want to have to give up all of your proprietary knowledge. (00:28:35) And if you're a chip company, you don't want a monopsony buyer situation. (00:28:39) situation where there's only one or two companies who can buy your chips, and by the way, they're producing their own, you want to have as diverse and healthy an ecosystem as possible where there's lots of potential buyers for your chips. (00:28:51) And if enterprises are rolling their own using open models, that's kind of an ideal situation because now there's like a long tail of buyers. (00:28:58) So I think really the whole ecosystem in a way, the chip companies, (00:29:04) developers, applications, enterprises, everybody has an incentive for a competitive layer of the stack at the model layer. (00:29:14) And really the only companies who don't are Anthropic and OpenAI, because obviously they want to dominate, they want to be a duopoly. (00:29:22) And my view is, look, if you earn a monopoly or duopoly in our system, we don't ban monopolies in the United States. (00:29:29) we ban anti-competitive tactics. (00:29:31) But if you lawfully achieve monopoly through amazing performance, we don't nationalize you or make you illegal, and I think that's fine. (00:29:40) However, the government, in my view, should do nothing to make monopoly or duopoly more likely. (00:29:46) They should make it harder for these companies to engage in monopoly tactics, and they should do everything they can to keep the model layer competitive because (00:29:55) Competition is what brings out the best and it's good for the ecosystem and it ensures our civil liberties and consumer choice. (00:30:02) And it's going to be amazing for pricing, Sachs. (00:30:04) If you think about what this competition is about to do, and you've been very vocal about this in your time in Washington, D.C., we want to have a level playing field. (00:30:12) We want to see massive competition. (00:30:14) I'll give you your flowers and your lei. (00:30:16) You did a great job of now setting the table for this in 2026 and 2027 going forward. (00:30:22) by letting people compete. (00:30:23) The cost of tokens, Freeburg, is going to go down 90% a year for the next three years. (00:30:28) You're going to be able to buy 1000 times as many tokens that are more intelligent because you're going to have free options. (00:30:35) The price will be free or close to free for many of these, and that could be incredibly disruptive. (00:30:41) Yes, Dave? (00:30:42) Yeah, and I think people are going to, again, deploy their own hardware against it. (00:30:47) I think there's going to be a buying frenzy in the enterprise, not just with the neo clouds and the hyperscalers. (00:30:52) I think the enterprise is going to be a buyer. (00:30:55) And when that happens, you do the simple math, and you don't want to have a dependency on server availability and cloud downtime, and you don't want to put these models on some third-party cloud, and you want to do stuff that's very cheap. (00:31:08) Like a lot of people are running, I mean, you guys do this, you're running day-to-day workflows for your enterprise, and you realize, hey, I could run these workflows on an open-source model on a machine in my office, and we don't need to be sending this stuff back to some (00:31:20) You're right. (00:31:21) The industry spent so many years convincing everybody to flip to the cloud, and the realization may be that all this idea of shared infrastructure may not be the best idea in a world of intelligence. (00:31:35) And this is what I mean by like a distributed spoke model, because I do think it's not going to be all or none. (00:31:40) I think you're going to end up being like 70, 20, 10 in how you're going to allocate your resources for model inference and running models. (00:31:47) You're going to probably be 70% in some big cloud. (00:31:50) Maybe you'll do 20% local, 10%, you'll try other clouds. (00:31:53) You know, you'll kind of mix stuff up, but I don't think you're going to end up doing things the way you've been doing them historically. (00:31:58) You'll very quickly realize that it's okay (00:32:01) to waste tokens. (00:32:02) It's okay to let your employees make stupid apps that last for a couple of weeks, but burn through billions of tokens. (00:32:09) If it's running on your own hardware, then all you're paying for is the electricity in your office or in your IT. (00:32:14) This is what I've been talking about here for, like, I'm totally fine, you know, for my kids to eat sugar if I don't have to deal with them, you know, like, go ahead. (00:32:28) Have you guys noticed that (00:32:30) in the last three days, we've now seen other people trying to get their own lock-in. (00:32:35) Microsoft just announced a two and a half billion dollar investment to stand up in FDE, or Amazon is spending a billion dollars. (00:32:44) For deployed engineers, for the people in the order for deployed engineers, they're making. (00:32:47) has one. (00:32:48) Open AI has one. (00:32:49) Yeah, and when they come knocking, Jamat, they're knocking like, hey, can I send my engineers to study your business and put it into my model? (00:32:55) I mean, people are going to be slamming the door on these. (00:32:56) It's like getting a Jehovah's Witness at your door. (00:32:58) Like, no, I don't want to be part of your cult. (00:33:00) I want to own this. (00:33:02) And that's what I've been saying. (00:33:03) Just let me make one quick point and I'll hand it to you for you. (00:33:05) This is what I've been saying with, when I was going on my Open Claw, which now is like Hermes and some other products. (00:33:11) Everybody in your organization is going to have a Mac Studio or a Dell with a massive amount of RAM, and you're going to spend $10, $20,000 per employee on local compute so that they can token max to a retard maxing level. (00:33:26) Who cares what they do on their local computer? (00:33:28) Who cares? (00:33:28) Let them rip. (00:33:29) And then you're going to give them a laptop that connects to it, and it syncs so you can control it. (00:33:34) It's literally going to be a server per individual in your company. (00:33:38) That's the way to model this in your brain. (00:33:40) Everybody has their own language. (00:33:41) model that they're crafting 100% of the time as they work. (00:33:45) And it's all local, so you don't have data leaks. (00:33:47) Go ahead, Freeberg. (00:33:48) I'll give you the final word. (00:33:50) I mean, I think one of the things that this reveals quite clearly is that there is no buttered, slippery slide to job loss. (00:34:01) I think everyone is realizing that, you know, I was wondering where you were going with that buttering it up. (00:34:05) So is Truman. (00:34:07) But I think, the idea that everyone had in their head two years ago, the narrative that was formed and everyone clutched onto, and by the way, you will not see the media reverse on this narrative. (00:34:17) I've realized the importance here, if you're a reporter or you're a journalist and you come up with some story that says something that's happening or is going to happen or has happened or is about the future, you destroy your own credibility (00:34:37) if the narrative shifts, you cannot ever let go of the narrative. (00:34:41) And I think this is one of the things that we can kind of acknowledge is going on with this job loss narrative problem, is that even as all the data comes out, as all of the reconfiguring of how enterprises are using AI, as it's revealing to all of them that they're actually not just going to cut costs, but they're going to grow revenue, and it's going to be a kind of clunky way of getting there, and it's not going to happen overnight. (00:35:01) It's not the slippery slope to job loss, to nihilism. (00:35:05) Everyone's going to kind of wake up and be like, wait, (00:35:07) This reality that we all thought we were living in is not really the reality that we are living in. (00:35:12) We're living in a reality where AI is clunky. (00:35:15) It takes some putting together. (00:35:17) It's a little more complicated than we thought. (00:35:19) It's definitely going to deliver value, but it's not about just turning off all the jobs and letting the genius AI solve all my enterprise problems and scale me into infinity without humans. (00:35:28) And I think that that's a big kind of narrative shift. (00:35:30) And you will not see the media accept that their narrative is wrong. (00:35:34) Because as soon as they have to acknowledge that they were wrong in what they were saying about job loss and all the other senators and people that are proclaiming job loss, job loss, job loss. (00:35:43) By the way, the reason they're making that proclamation is so that they can step in and control AI and they can drive their systems of socialism, which is what they're all looking to deploy. (00:35:52) But if they had to come in and say, look, the data doesn't map to the narrative, their credibility is (00:35:58) So they'll double down on it and they'll double down on it. (00:36:00) And I'm telling everyone that's listening, look at the data. (00:36:04) There is no job loss with AI. (00:36:06) It is an absolute scam to tell the world that AI is taking away jobs and destroying jobs and the world is shifting. (00:36:13) It is clunky. (00:36:14) It is valuable. (00:36:15) It is going to take some time and it is going to create far more jobs than it is destroying. (00:36:20) We're a bunch of monkeys. (00:36:21) We've been given this new tool. (00:36:22) We're going to solve more problems with the new tool. (00:36:25) There will be job displacement. (00:36:27) And the problem is people do not understand the nuance between these two terms. (00:36:32) Certain jobs will be retired and they're going to be retired at a faster rate than you can imagine. (00:36:37) But other jobs will happen. (00:36:39) And you pointed this out, Chamath, hey, it's the most, I don't know if you did it here or on your personal channel on YouTube, but you know, it's the most empowering tool ever. (00:36:47) If you. (00:36:47) Jason, what jobs are getting replaced? (00:36:50) But you're saying some of them are going to go away faster than you can see. (00:36:55) What are those jobs that are going to go away faster than you can see? (00:36:57) Because I'm still not seeing the categories. (00:37:00) Totally. (00:37:00) Customer support jobs, those are going to go away very quickly. (00:37:03) Where? (00:37:04) You're just making shit up because it's not actually happening. (00:37:06) I see it in enterprises all the time. (00:37:08) Where do you see it? (00:37:10) Show me the customer support enterprise shutdown that's happening. (00:37:13) Just show it to me. (00:37:14) Is it in the room with us right now? (00:37:15) Like, where is it? (00:37:17) I will literally, you know, I'll do that research and (00:37:19) I'll give you some research. (00:37:22) Hold on. (00:37:22) Can I provide some research? (00:37:23) We had this debate before. (00:37:24) If you ask me a question, just let me finish my answer and then you can take it. (00:37:28) You asked me the question. (00:37:29) Customer service jobs are going to go away because consumers prefer to talk to the AI and it's perfect at that. (00:37:36) Entry-level jobs around data entry, around business process outsouring, those are going to go away. (00:37:41) So will driving cabs and those kind of jobs. (00:37:43) Those ones are obvious low-hanging fruit, just like we had the typing pool and messengers and other products (00:37:49) go away when we had word processors on every desk. (00:37:52) Those jobs will be displaced 100%. (00:37:55) And those jobs will be in the middle. (00:37:56) You're 20 of cab drivers going away, dude. (00:37:58) Like when Uber came out, everyone said all the driving jobs were going to go away. (00:38:01) It's all over. (00:38:04) Waymo is stuck at 3,000. (00:38:05) Tesla's stuck at like 30 cars, but that's all going to change in the very near future. (00:38:11) Everything you're saying is perspective. (00:38:13) This is my point. (00:38:14) You are still being perspective in the future with everything you're saying. (00:38:18) And every point that we turn, every time we turn a page, it's like, wait, do you think that self-driving is real? (00:38:24) Is that your premise? (00:38:25) Is that self-driving is not going to get rid of cab drivers? (00:38:27) Is that your premise? (00:38:28) I think it's real, but I think it's real for sure. (00:38:30) Do you think it will get rid of cab drivers? (00:38:32) You're arguing that. (00:38:33) I get in the car every day and my car drives me and I get on my phone and I do work. (00:38:41) Okay, but do you think cabs and cities are not going to be self-driving? (00:38:45) Do you think Zipline's not going to deliver? (00:38:46) You're doing the Mountain Bailey thing again, J. (00:38:48) Cal. (00:38:48) It's not Mountain Bailey. (00:38:49) I see this from my investments. (00:38:51) I have a company, Auto Lane, that is doing this right now. (00:38:54) But just go to customer support again, because you've been saying, yeah. (00:38:56) I literally have had this debate verbatim and displacement, not job loss, is my only point. (00:39:03) is pointing out is that whenever anyone pushes back on the fact that you don't have any data to support this in the present, you say you're talking about the future. (00:39:11) Okay, fine. (00:39:12) It's a prediction. (00:39:13) You know, there's no data to support it in the present. (00:39:16) That's fine. (00:39:16) You're saying it's going to happen in the future. (00:39:18) No, I'll give you the data point. (00:39:19) You can dismiss what I want. (00:39:20) I will. (00:39:20) Yes. (00:39:21) If you talk to Uber and Waymo, in the cities where Waymo is present and has gotten past a couple of 100 cars, they've stopped recruiting drivers. (00:39:28) Drivers are either static or going down in those markets. (00:39:31) So that is absolute evidence that this is happening. (00:39:34) And the rollout is going to be fast and furious. (00:39:36) So it is not a future prediction. (00:39:37) You can talk to the CEO of Waymo, you can talk to the CEO of Lyft and Uber, and they will tell you this explicitly. (00:39:41) Let's call it marketing. (00:39:43) We interviewed Dar and he said jobs were increasing at Uber. (00:39:47) No, you're delivering more. (00:39:49) Like those, because productivity is going up with delivery, productivity goes up. (00:39:52) So these people drive more. (00:39:54) I'm 100% lying. (00:39:56) You're being disingenuous on stage.