More Trillion Dollar IPOs, Anthropic $3T, Zuck's Price War, China Ends Open Source?, Trump Accounts

“Well, we're going to get into that. That's on the docket for sure. But let's start with the IPO update. You know, there's a trillion-dollar IPO rush to the exits, and, you know, this was a big topic of discussion, Brad, at the Liquidity Summit last month. And we'd never seen a trillion-dollar IPO. We had one, this year already. SpaceX trading right about where it went public, so it was priced, I guess, to perfection. And theoretically going to see two more, uh, Brad has the inside information, so I'll try to get it out of him. OpenAI and Anthropic are slated to go out. Let's, uh, just go quickly over what happened with SpaceX. It ran up to $200 a share, uh, it's been down a bit, it's at $150 a share, as I said, that's right at the IPO price. So it's trading at that $2 trillion market cap, currently 7th largest company in the world. And Anthropic confidentially filed on June 1st. I, I don't know why they call this confidential filing when it immediately comes out, uh, but I guess the information is confidential. Polly Market says 65% chance Anthropic's IPO will happen this year on light volume, 360K. And, uh, 2 weeks ago Gavin Baker, another bestie, said he thinks They're gonna end 2026 with over $100 billion in revenue and very profitable. He said, mm-hmm, a couple of us guessed on the program that he thinks it would trade at $3 trillion right now if it went public. Chamath, you made a great call on the pod. You said, hey, good idea for Elon to get out first. What are the chances here, Chamath, that these other two get out this year or maybe in like, you know, say 9 months in the, in, in the first quarter of next year?”

“Well, but just to take on Brad's point for a second, I actually agree with what I think you're saying, Brad, which is the market today seems to be pushing towards duopoly. Or it has become a duopoly, certainly measured in terms of revenue. If you're to look at market share based on token revenue, there's only two companies making meaningful revenue. Anthropic's at what, $60-something billion ARR? OpenAI at $40-something billion ARR. I don't know if anybody else even registers. And it may be the case that the more tokens that Anthropic and OpenAI produce— I mean, we've got to remember, every token that they're serving up is on behalf of a use case, right? So they themselves are learning from that, and they're getting better at then providing whatever offering that is. And so who knows, like, the gap may be growing. A year ago it seemed like we had 5 major labs, you know, now it seems like there's a top 2 and then everybody else. So I mean, look, I could see AI easily becoming another tech market that becomes a duopoly, which, which by the way is the trend. The historical trend is like monopoly or duopoly in most tech categories, for better or worse.”

“Yeah, I think, I think it's very high. But let me, let me first say, you know, the SpaceX IPO where we were also investors and we also bought in the IPO, I mean, it was textbook. It was a hugely successful IPO. They raised $75 billion at 1.75 trillion. Okay, so it went out below where we are today. It's up 25%, you know, and let's call it on $35 billion of forward revenue. So if you think about that revenue multiple, it's trading at $2 trillion on roughly $35 billion of, of forward revenue. It's an incredible achievement. I think it was textbook. I think Anthropic and OpenAI were watching very closely because frankly, we had not had an IPO of that size. And to Elon's credit and to the team's credit, Bret and Gwen, they really pioneered some really smart and interesting things as part of that IPO. So, you know, you heard from Gavin, Anthropic's rumored to be, you know, trending over $100 billion in revenue compared to the $35, right? If they exit the year at $100, that means their GAAP revenue next year could be well over $100. So based on the SpaceX success, I think it would be a blockbuster IPO. And I think SpaceX has shown them the way on things like the total raise, pricing, liquidity, inclusion into the indexes, how to do the lockup. Like, I think they've gone to school.”

“That the, the peak to trough drawdown in the 6 months post-IPO is 50%. We've seen a pretty big drawdown here from the peak to trough as well. So you don't want to jam it into an index at the peak and then have a 30% drawdown on top of people, which often happens in IPOs because people get excited, it runs ahead of itself. But they didn't do that here. There was fear that that was going to happen. So both the, the exchanges and the indexes, they looked at this and they made some modifications because the, the other side of the argument is it's so damn big and important that it needs to be part of the index, right? Right. And so the reason the rules had previously existed is because most companies coming public were younger, earlier, less tested, less revenues, less profitable, all the things weren't as important in the overall scheme of things. So I think that they pioneered some really smart things. It's worked well, it's traded well. And so I think that that provides a bit of a blueprint for Anthropic. But just in, in, in terms of the enthusiasm, Is an Altimeter, is a Fidelity, is a T. Rowe an enthusiastic buyer of Anthropic based upon the things we know today around profitability and model improvement and revenue growth, etc.? Yes, everybody would be pigpiling in. Everybody would be trying to get into the top of the book. And, you know, the last I heard, you know, again, rumor that they would like to get out this year on OpenAI. Everybody knows that Anthropic kind of passed OpenAI on a revenue trajectory. But I will tell you, OpenAI's kind of got its swagger and mojo back. It's coming out, you know, just today with a whole new set of models. We know GPT-6, you know, there's a lot of talk of that coming out within the next 30 days, a whole new generation of models. I think their revenue has really ticked back up. The most recent kind of rumors I see on Twitter is around $70 billion ARR at the year. So, Just as a reminder, $70 billion may not be over $100 billion that's rumored in Anthropic, but it's still twice, uh, you know, where, where the revenue of SpaceX is at. So can they get out at over a trillion on that type of revenue growth, being one of the two frontier premier labs? I think the answer to that is yes. Um, I'm not sure there's a huge race between the two of them to get out first. I think they'll both go out when it's, when it's time. I think OpenAI has a little bit more complexity just associated with the corporate restructuring that they have to go through, et cetera. So I would be surprised if they go out before Anthropic, but the fact of the matter is, I don't know. But today, as I sit here today, Altimeter would be a buyer at scale and at size in both of those IPOs.”

“That the, the peak to trough drawdown in the 6 months post-IPO is 50%. We've seen a pretty big drawdown here from the peak to trough as well. So you don't want to jam it into an index at the peak and then have a 30% drawdown on top of people, which often happens in IPOs because people get excited, it runs ahead of itself. But they didn't do that here. There was fear that that was going to happen. So both the, the exchanges and the indexes, they looked at this and they made some modifications because the, the other side of the argument is it's so damn big and important that it needs to be part of the index, right? Right. And so the reason the rules had previously existed is because most companies coming public were younger, earlier, less tested, less revenues, less profitable, all the things weren't as important in the overall scheme of things. So I think that they pioneered some really smart things. It's worked well, it's traded well. And so I think that that provides a bit of a blueprint for Anthropic. But just in, in, in terms of the enthusiasm, Is an Altimeter, is a Fidelity, is a T. Rowe an enthusiastic buyer of Anthropic based upon the things we know today around profitability and model improvement and revenue growth, etc.? Yes, everybody would be pigpiling in. Everybody would be trying to get into the top of the book. And, you know, the last I heard, you know, again, rumor that they would like to get out this year on OpenAI. Everybody knows that Anthropic kind of passed OpenAI on a revenue trajectory. But I will tell you, OpenAI's kind of got its swagger and mojo back. It's coming out, you know, just today with a whole new set of models. We know GPT-6, you know, there's a lot of talk of that coming out within the next 30 days, a whole new generation of models. I think their revenue has really ticked back up. The most recent kind of rumors I see on Twitter is around $70 billion ARR at the year. So, Just as a reminder, $70 billion may not be over $100 billion that's rumored in Anthropic, but it's still twice, uh, you know, where, where the revenue of SpaceX is at. So can they get out at over a trillion on that type of revenue growth, being one of the two frontier premier labs? I think the answer to that is yes. Um, I'm not sure there's a huge race between the two of them to get out first. I think they'll both go out when it's, when it's time. I think OpenAI has a little bit more complexity just associated with the corporate restructuring that they have to go through, et cetera. So I would be surprised if they go out before Anthropic, but the fact of the matter is, I don't know. But today, as I sit here today, Altimeter would be a buyer at scale and at size in both of those IPOs.”

“Right? The accredited investor laws are insane that we have in this country and keeps people from participating in these things. But it is what it is. Right? So they're coming public at over $1 trillion. I still think there's a lot of meat on the bone on SpaceX, on Anthropic, on OpenAI. But you're not going to have things that are— I don't expect that they're going to be priced in a way where you're going to get a 50 to 100% durable bounce out of the IPOs. If so, that would mean they were probably mispriced right into the IPO. But I do think that these things can be compounders. They're going to compound at the rate they compound revenue. And I think all of these companies are going to compound revenue. At well over 30% for the next many years.”

“And I said, okay, so my costs are doubling every 45 days. My upside is essentially flat. And he said, basically. And I said, well, explain why that is. And he said, honestly, what we're finding out is that you need to use a lot more tokens to get to this next iteration of improvement because we've effectively already asymptoted. And I said, so what should we do? And he said, honestly, we have to figure this out. And so we're going to take a step back and try to figure out what to do. I don't know how many other companies will actually go through this reckoning now, but the point is everybody in the next 3 or 4 years will for sure go through it. So I suspect that if you can get out now, you should get out now before all of that starts to seep into the water table, because I think that's probably what allows you to get out at a huge price and raise a huge amount of money.”

“Anthropic's new model. I first asked it, what is the lift of the S&P 500 earnings per share growth since 2024 from AI? And they answered, oh, it's 50%. So then I looked through it and I said, well, no, you're including the money that Nvidia makes from selling chips to So I said, okay, I asked a different question, which is then what was the EPS growth of the S&P 493? And the answer was 9%. And I said, okay, well that's different. And I said, unpack that. And the overwhelming majority of that was from pricing power sitting on top of inflation. And then the other 3% was from buybacks. And so the answer, as far as all publicly available data, was that the actual ROI was somewhere between 0% and 2%. So I don't know. I mean, I think that enterprise looks really good. The problem is that very smart investors like Brad Stone and Gavin Wood and others at some point will start asking companies, what's your ROI? What's the actual EPS lift? And if the answer is, well, I don't really know, or I'm not sure, and you know, you don't necessarily have the pricing power to continue to raise prices. Enterprise is probably a little bit more brittle because there are fewer buyers and they're more demanding. Consumer, on the other hand, then all of a sudden becomes an incredible safe harbor because you have tens of millions of buyers., and having those 2 orders of magnitude more buyers at a much smaller price point inoculates you from the vicissitudes of an ROI discussion. So it all really depends on what the actual ROI is of this money being spent. I think that we're in the phase of just being astonished, as Brad said, about the scale of the revenue growth. Yeah, but at some point you'd have to be an idiot not to ask Well, who is paying you this, and can they sustain paying it to you? I just don't know what the answer to that question is. And at some point— it may not be now— at some point, people will have to answer that question. And interestingly, a million dollars a year on tokens, and that million dollars a year is doubling and tripling and quadrupling. At some point, you're going to have to show an ROI that's above the risk-free rate of return. Otherwise, you're going to have some angry investors on your hands.”

“Anthropic's new model. I first asked it, what is the lift of the S&P 500 earnings per share growth since 2024 from AI? And they answered, oh, it's 50%. So then I looked through it and I said, well, no, you're including the money that Nvidia makes from selling chips to So I said, okay, I asked a different question, which is then what was the EPS growth of the S&P 493? And the answer was 9%. And I said, okay, well that's different. And I said, unpack that. And the overwhelming majority of that was from pricing power sitting on top of inflation. And then the other 3% was from buybacks. And so the answer, as far as all publicly available data, was that the actual ROI was somewhere between 0% and 2%. So I don't know. I mean, I think that enterprise looks really good. The problem is that very smart investors like Brad Stone and Gavin Wood and others at some point will start asking companies, what's your ROI? What's the actual EPS lift? And if the answer is, well, I don't really know, or I'm not sure, and you know, you don't necessarily have the pricing power to continue to raise prices. Enterprise is probably a little bit more brittle because there are fewer buyers and they're more demanding. Consumer, on the other hand, then all of a sudden becomes an incredible safe harbor because you have tens of millions of buyers., and having those 2 orders of magnitude more buyers at a much smaller price point inoculates you from the vicissitudes of an ROI discussion. So it all really depends on what the actual ROI is of this money being spent. I think that we're in the phase of just being astonished, as Brad said, about the scale of the revenue growth. Yeah, but at some point you'd have to be an idiot not to ask Well, who is paying you this, and can they sustain paying it to you? I just don't know what the answer to that question is. And at some point— it may not be now— at some point, people will have to answer that question. And interestingly, a million dollars a year on tokens, and that million dollars a year is doubling and tripling and quadrupling. At some point, you're going to have to show an ROI that's above the risk-free rate of return. Otherwise, you're going to have some angry investors on your hands.”

“Anthropic's new model. I first asked it, what is the lift of the S&P 500 earnings per share growth since 2024 from AI? And they answered, oh, it's 50%. So then I looked through it and I said, well, no, you're including the money that Nvidia makes from selling chips to So I said, okay, I asked a different question, which is then what was the EPS growth of the S&P 493? And the answer was 9%. And I said, okay, well that's different. And I said, unpack that. And the overwhelming majority of that was from pricing power sitting on top of inflation. And then the other 3% was from buybacks. And so the answer, as far as all publicly available data, was that the actual ROI was somewhere between 0% and 2%. So I don't know. I mean, I think that enterprise looks really good. The problem is that very smart investors like Brad Stone and Gavin Wood and others at some point will start asking companies, what's your ROI? What's the actual EPS lift? And if the answer is, well, I don't really know, or I'm not sure, and you know, you don't necessarily have the pricing power to continue to raise prices. Enterprise is probably a little bit more brittle because there are fewer buyers and they're more demanding. Consumer, on the other hand, then all of a sudden becomes an incredible safe harbor because you have tens of millions of buyers., and having those 2 orders of magnitude more buyers at a much smaller price point inoculates you from the vicissitudes of an ROI discussion. So it all really depends on what the actual ROI is of this money being spent. I think that we're in the phase of just being astonished, as Brad said, about the scale of the revenue growth. Yeah, but at some point you'd have to be an idiot not to ask Well, who is paying you this, and can they sustain paying it to you? I just don't know what the answer to that question is. And at some point— it may not be now— at some point, people will have to answer that question. And interestingly, a million dollars a year on tokens, and that million dollars a year is doubling and tripling and quadrupling. At some point, you're going to have to show an ROI that's above the risk-free rate of return. Otherwise, you're going to have some angry investors on your hands.”

“I think it's not that. I think it's more that when the iPhone was a novelty, everybody would keep upgrading because you expected that the new price was worth it. And then at some point there's a moment, and you can debate when it happened, where people said, you know what, I'm just going to keep the old phone because it's good enough and I just don't see the difference. And I think that there's going to be a moment like that. Like When I use Gable-5, the problem is that it's nerfed on a bunch of things that I would normally research. I was with somebody this weekend and he was telling me about some health thing and I put it into Gable and it's like, won't answer you. And I'm like, okay. So I think that everybody will get to a point, they'll get to it at different times where they just say, you know what? It shouldn't really matter what model I'm using if I get an answer that I think is reasonable and I can go about my day. Separately, I think when the corporate CFO gets involved, that'll be an entirely different conversation altogether. I think that what I can tell you after this UN commission that I joined with Benioff and Jensen and Brad Smith, there is not a single country in the world that is not trying to figure out its own sovereign AI strategy. And I don't think they believe using a closed-source American model is the answer. And so, you know, we— I think we have to keep in mind there's trends. One is just geographic penetration of humans, and there are still many, many, many more people that don't use it than do, which is an upside and an opportunity for everybody. And then the second is there is going to be the experimentation, as you said, that needs to transition to ongoing repeatable usage. And then the third is that all of that then needs to plug into the existing regulatory infrastructure that we use as societies to run the world. And I think when you put all of these things together, it's not clear to me who wins, except that you're going to have a lot of diversity of choice. Certain countries, I can tell you after this week, have no desire to subjugate themselves to any technical risk. And so they're willing to spend the money to have their own. Now, we can argue and debate whether that country has any chance, right? They would rather take an open-source model like Llama's, actually, and stand up their own stack, soup to nuts, for their own people and their own companies inside of their own country. And if the models are 99% as good or 95% as good, there's going to be a claim that some countries make, which is it's just good enough.”

“Well, look, I think that enterprise CTOs would like to shift their token consumption to cheaper models. For the obvious reason that that would be more efficient. And they are seeing their compute costs or their token costs just skyrocketing right now. So everyone's trying to figure out how do we put the brakes on this or at least control it, you know, make sure we're getting ROI. You also have the AI sovereignty issue that we discussed last week that Alex Karp talked about where they're worried about giving up the secret sauce or the alpha in their business to a frontier lab that may one day be competing with them. So there's no question that enterprises would like to diversify. They would like to get off of these frontier models when they can. The problem is, I think in most cases they don't have the technical ability to do it. I mean, Coinbase figured out how to do it. DoorDash figured out how to do it, which is to say they built a token routing system, a layer of middleware that allows them to sort of send frontier tasks to frontier models and non-frontier tasks to more mundane models., but I don't think your average enterprise has the technical capability to do that. So I think this is a case of the spirit is willing, but the flesh is weak. I mean, they are willing, they would like to diversify off of these closed models, but they are unable to do it. And so this is why the share of wallet of closed models, it actually increased. I think that open source went from from 19% last year to 11% this year. So open source as a share of enterprise spending is actually decreasing. Now, I don't think that means that usage is decreasing. I think usage is skyrocketing in both these categories. It also may be the case that because the whole point of using an open model is you just pay for the compute cost, you don't have to pay a lab. So it may be the case that it's hard to measure that usage in terms of spend. But nonetheless, I mean, anyone who's saying that these closed models are gonna lose or are somehow losing, you're just not seeing it in the data. Like Brad's saying, the revenue is skyrocketing. And I think the most you can say is that enterprises that are technically capable would like to gravitate towards hybrid architectures, but at the same time, it takes technical expertise and it is just phenomenally convenient whether you're a developer or an enterprise just to go with the Frontier Labs. And that's why their revenue is skyrocketing as well.”

“Yeah. In this case though, however, we're using revenue as the metric to determine the winner. Keep in mind, when you're doing open source, those are dark tokens. Those don't come up as revenue. So we don't know the utilization that's occurring at DoorDash when they're using an open source model. We do know their Fable and their Anthropic spend, right? And because we see that in the Anthropic revenue ramp, the more they deploy these things on their own hardware, using commoditized hardware, using the Neo Clouds, you don't see that. It doesn't come up as revenue, it comes up as free. The only thing you're paying for there is the hosting cost, you know, and that's— that will come up on Nvidia's balance sheet. So the gains you'll see there will be Cerebras, NeoClouds, CoreWeave Cloud, etc. So just keep that in mind when we're having this discussion. Let's talk a little bit back to sovereignty here. The CCP said that they might— or there's a report out according to Reuters, Reuters generally does a good job of this They dropped a couple of anonymously sourced reports about AI in China, and these were published about 15 minutes apart. The big scoop that CCP officials, Chinese Communist Party, are reportedly considering restricting overseas access to China's top models. So two Chinese regulators met with Alibaba, ByteDance, and Z.AI. They're the ones who are doing GLM 5.2 that we keep referencing. They're discussing limiting access to the top open and closed models outside of China. Why are they doing this? Well, they're, they're making any theft or leaks of AI research a national security offense, and they want to control who can fund Chinese AI labs. And we saw this with Manus, which was a Chinese company, tried to go to Singapore. The CCP pulled those employees from Singapore back to China. And so here is their main concern. The quote is that they're concerned about Mythos. Chinese authorities are deeply worried about the potential for Mythos to exploit software vulnerabilities and that Washington might deploy a model against Chinese interest. Sachs, last week I proposed the reverse to you in your previous position as czar of AI. Do you think the United States should be banning those models? Now we have the opposite. China's saying potentially, according to these reports, allegedly, that they might restrict them. So explain the game on the field here. If you're going to look into what China's thinking, why would they want us to not have those open-source models? And how is this chessboard developing?”

“To add to that, uh, Brad, just open source is very hard to implement when compared to just firing up Claude and having Claude already approved in your organization. The number of steps it took me in order to— and I, I'm pretty familiar with technology. It took me hours to configure my new setup to get onto this BitTensor network, to get OpenRouter going. And to your point, Chamath, it does dynamically route now. So I'm dynamically routing, and I'm, you know, GLM 5.2, and then if I fall back to Claude— but which Claude am I going to fall back to? And here's another piece of evidence to your point, Chamath. There are some organizations that just aren't capable of this. They don't have the the team that, you know, does this naturally. We just talked about the CTO of Uber. Now let's talk about another CTO. Andy Fang is the CTO of DoorDash. Shout out to Stanley. And so he, as you can see here in this, uh, tweet— I, I have a lot of people listening to All In over the last couple weeks are coming out and, as I said, explaining what they're doing to address this exact issue. He says, hey, with our internal coding benchmarks, we're able to confidently introduce open weight models into our AI code review without degrading code quality. Have the frontier model Claude from Anthropic to do the hardest work, delegate lower level work to KIMI 2.6, and they are now releasing their benchmarks. So another group releasing their benchmarks and saying, hey, we know this is an issue. The CTO has been charged, to your point, Chamath, again, CFO says, hey, make sure this is profitable. We get the ROI. They put that on the CTO. Here's another CTO from another leading tech organization that knows how to implement this.”

“But if you don't know exactly what you're gonna use 'em for, you want the most powerful general intelligence that you can get, right? So what he said is that for mature use cases, yeah, you want to go open, but for immature use cases, which are all the new things people are discovering right now, you're just gonna want to use the most capable general model that you can. And then once you figure out what the workflow is and what the workload is gonna be and exactly what you're trying to accomplish, then you can use a small, highly trained model. And I think he said something like—”

“I think it's not that. I think it's more that when the iPhone was a novelty, everybody would keep upgrading because you expected that the new price was worth it. And then at some point there's a moment, and you can debate when it happened, where people said, you know what, I'm just going to keep the old phone because it's good enough and I just don't see the difference. And I think that there's going to be a moment like that. Like When I use Gable-5, the problem is that it's nerfed on a bunch of things that I would normally research. I was with somebody this weekend and he was telling me about some health thing and I put it into Gable and it's like, won't answer you. And I'm like, okay. So I think that everybody will get to a point, they'll get to it at different times where they just say, you know what? It shouldn't really matter what model I'm using if I get an answer that I think is reasonable and I can go about my day. Separately, I think when the corporate CFO gets involved, that'll be an entirely different conversation altogether. I think that what I can tell you after this UN commission that I joined with Benioff and Jensen and Brad Smith, there is not a single country in the world that is not trying to figure out its own sovereign AI strategy. And I don't think they believe using a closed-source American model is the answer. And so, you know, we— I think we have to keep in mind there's trends. One is just geographic penetration of humans, and there are still many, many, many more people that don't use it than do, which is an upside and an opportunity for everybody. And then the second is there is going to be the experimentation, as you said, that needs to transition to ongoing repeatable usage. And then the third is that all of that then needs to plug into the existing regulatory infrastructure that we use as societies to run the world. And I think when you put all of these things together, it's not clear to me who wins, except that you're going to have a lot of diversity of choice. Certain countries, I can tell you after this week, have no desire to subjugate themselves to any technical risk. And so they're willing to spend the money to have their own. Now, we can argue and debate whether that country has any chance, right? They would rather take an open-source model like Llama's, actually, and stand up their own stack, soup to nuts, for their own people and their own companies inside of their own country. And if the models are 99% as good or 95% as good, there's going to be a claim that some countries make, which is it's just good enough.”

“let me throw something out I'd like to get your opinion on. You know, to a certain extent, there's this implied assumption in the world that, uh, there's going to be this convergence of intelligence, right? And if you look at the benchmarks today, seems like everybody, you know, on the benchmarks, they are converging. But yet if you look at the revenue distribution, it's not converging at all. One of the questions I have, will the model router itself be smart enough to overcome, David, the inherent intelligence advantages of the generalized, of the frontier labs? The non-consensus argument might be that intelligence is not converging at all, that superintelligence becomes fully self, uh, uh, you know, recursive. And as it becomes recursive, you actually extend the lead because the smarter your model gets, the more revenue you get, the more compute you can buy, the more compute you can buy, the better the model is that you can build. So I think there's a chance that over the course of the next 2 to 3 years, as we take on much more complex agentic tasks, that the distance between the frontier and everybody else doesn't converge, it actually extends. We shall see. But, you know, I think there's this implicit assumption in all the arguments today that everything's converging, I'm not sure that we've really run that to ground.”

“Yeah, here's the Mark Zuckerberg tweet, just so I queue it up for you there, Brad. And he's @FinkD, that was his old handle back when he was in college, F-I-N-K-D. And he's done more tweets today over this Llama Spark announcement than he's done in his history. So he's getting into the x.com conversation. "Today we're releasing Llama Spark 1.1, a strong agentic encoding model at a very low price. It's available through our new Meta Model API and in Meta AI." So he's coming out saying, "Hey, we got the strongest agentic." tool here, please come use it. He also wants to have his own, essentially, you know, he wants to jump into not the hosting space, but he wants to provide tokens as well.”

“Yeah. In this case though, however, we're using revenue as the metric to determine the winner. Keep in mind, when you're doing open source, those are dark tokens. Those don't come up as revenue. So we don't know the utilization that's occurring at DoorDash when they're using an open source model. We do know their Fable and their Anthropic spend, right? And because we see that in the Anthropic revenue ramp, the more they deploy these things on their own hardware, using commoditized hardware, using the Neo Clouds, you don't see that. It doesn't come up as revenue, it comes up as free. The only thing you're paying for there is the hosting cost, you know, and that's— that will come up on Nvidia's balance sheet. So the gains you'll see there will be Cerebras, NeoClouds, CoreWeave Cloud, etc. So just keep that in mind when we're having this discussion. Let's talk a little bit back to sovereignty here. The CCP said that they might— or there's a report out according to Reuters, Reuters generally does a good job of this They dropped a couple of anonymously sourced reports about AI in China, and these were published about 15 minutes apart. The big scoop that CCP officials, Chinese Communist Party, are reportedly considering restricting overseas access to China's top models. So two Chinese regulators met with Alibaba, ByteDance, and Z.AI. They're the ones who are doing GLM 5.2 that we keep referencing. They're discussing limiting access to the top open and closed models outside of China. Why are they doing this? Well, they're, they're making any theft or leaks of AI research a national security offense, and they want to control who can fund Chinese AI labs. And we saw this with Manus, which was a Chinese company, tried to go to Singapore. The CCP pulled those employees from Singapore back to China. And so here is their main concern. The quote is that they're concerned about Mythos. Chinese authorities are deeply worried about the potential for Mythos to exploit software vulnerabilities and that Washington might deploy a model against Chinese interest. Sachs, last week I proposed the reverse to you in your previous position as czar of AI. Do you think the United States should be banning those models? Now we have the opposite. China's saying potentially, according to these reports, allegedly, that they might restrict them. So explain the game on the field here. If you're going to look into what China's thinking, why would they want us to not have those open-source models? And how is this chessboard developing?”

“Well, last week I explained why it would be harmful to the US to ban open models. So if you're China and you want to harm the US, maybe you would want to. I mean, it does kind of make sense because our, our companies are benefiting a lot from all this R&D that they're doing. Now, at the end of the day, I think the story is probably a little bit overstated. I think there are a few Chinese models that were open source that have gone closed source, but I don't think they're all— I'd be surprised, let's put it that way, if they all went closed. So for example, the number one model in China, as I understand it, is ByteDance's model, which is already closed. That's kind of like their ChatGPT equivalent, and it's always been closed. Then you've got Alibaba's Qwen, which was open and now I think is going closed. And Xipu, which has GLM 5.2, which we've talked about a couple weeks ago because it seemed to be catching up to what was then commercially available as the American Frontier at certain tasks. They, I think, are going closed too after having been open. And so this is, I think, the tactic is you stay open until you catch the Frontier or you get close to it. And then there's a really compelling incentive to go close because you want to capture all the value for yourself.”

“Sacks. Well, having, having spent some time in DC this week, I, and, and, and talking with put the White House and Treasury, etc., on this topic. What I can tell you is, while there may be some, you know, debates about regulation of U.S. models, the one thing there's absolute agreement on is doing everything to stay ahead of China. And, you know, and, and the president, all the way up to the president, very interested. How far are we ahead of China? What are the things we need to do to stay ahead of China? It is a unifying force in Washington And the idea that we were going to kind of take our frontier labs off the field, off the playing field, while letting Chinese open-source models run free, you know, and on top of that, distilling our models. I will tell you, GLM 5.2 has watermarks from Mythos all over it, right? So we know they were distilling, et cetera. And I think the US government's gonna take steps against distillation, which they should do. So I think that, you know, China doing this in some ways I don't think it hurts the United States. The United States can spin up open source models. We've got Reflection spinning one up. Obviously, we've got the good work going on at Nvidia with their open source models. The labs— I've talked to a couple of the frontier labs about open source models. I said, why aren't you guys making open source models? They're like, there's not a lot of demand for it. If there was a lot of demand for it, we would make it. And so I think the US is in a good position. I think this is probably more chess playing by China than actual threats. Because it would hurt them a lot more than it would, uh, hurt us.”

“Sacks. Well, having, having spent some time in DC this week, I, and, and, and talking with put the White House and Treasury, etc., on this topic. What I can tell you is, while there may be some, you know, debates about regulation of U.S. models, the one thing there's absolute agreement on is doing everything to stay ahead of China. And, you know, and, and the president, all the way up to the president, very interested. How far are we ahead of China? What are the things we need to do to stay ahead of China? It is a unifying force in Washington And the idea that we were going to kind of take our frontier labs off the field, off the playing field, while letting Chinese open-source models run free, you know, and on top of that, distilling our models. I will tell you, GLM 5.2 has watermarks from Mythos all over it, right? So we know they were distilling, et cetera. And I think the US government's gonna take steps against distillation, which they should do. So I think that, you know, China doing this in some ways I don't think it hurts the United States. The United States can spin up open source models. We've got Reflection spinning one up. Obviously, we've got the good work going on at Nvidia with their open source models. The labs— I've talked to a couple of the frontier labs about open source models. I said, why aren't you guys making open source models? They're like, there's not a lot of demand for it. If there was a lot of demand for it, we would make it. And so I think the US is in a good position. I think this is probably more chess playing by China than actual threats. Because it would hurt them a lot more than it would, uh, hurt us.”

“Brad, let me just say just one comment. I mean, look, I agree with you that from the president on down, everyone wants to win the AI race. And in fact, that was in the big AI policy speech the president gave about one year ago. That was the whole thrust of the speech was declaring that we were in an AI race and America had to win it. I think the big risk is more that, and this would not be at like the top level, you know, I think if the president could make every single decision, it'd be perfect. The issue is at a lower level in the bureaucracy, do people somehow do things that are counterproductive? Maybe they think it's gonna help us in the race against China, but they end up doing something that's ham-fisted. They just like ban something or without, you know, really truly understanding all the implications of it. So I think there's no question that the administration wants to win the AI race. The president definitely does. And at the top levels, they will all make smart decisions. The question is whether at lower levels of the bureaucracy you can get mistakes being made. And then you have the influence of Congress and whatever they want to do. Those guys, they're more responsive, I think, in a way to like the Doomer community that's creating a lot of political pressure right now.”

“They're spending $5,000. Well, if the average salary is $100,000, $150,000 at this organization, it's only an incremental 3, 4, 5% on top of their salary. So the way I look at it is, did it make that person 3, 4, 5 times more effective at their job? And I think the answer is yes. So that's why there's so much token maxing going on. And it's also a bottom-up type product. You can just get into this product for $20 a month and, you know, no CIO or CTO is like, oh no, you can't spend $20 a month on your corporate card for this technology. So it— when a bottom-up technology hits everybody at the same time, That's what would explain this revenue ramp that we're all having a hard time adjusting to. It applies to every single person. Like, who isn't impacted by the technology is my question to you, Chamath. Like, in what organization you're working with, with 80-90, is there a department that says, yeah, the intelligence on demand, not for us, we don't need it?”

“And our discussion here for the last couple of weeks on the pod has centered around that. And the industry has responded on the place where all the CTOs, CEOs, and capital allocators hang out, which is X.com, formerly known as Twitter. Here's Praveen, the CTO of Uber. And so when you ask how are they getting the ROI out of this, People are now bringing that conversation front and center and they're explaining it on X. And he talked, remember, Uber was also the one that ran through all their tokens in the first quarter. So then on the other side of the business, which is legal, operations, marketing, customer support, HR, and procurement, which he lists here, he says in this, today, 99% of our engineers use AI tools. Okay, great. Everybody's doing vibe coding and has coding assistants. More than 70% of pull requests are attributed to local or cloud agents. Our engineers have built 2,500 agentic skills. So how are we bringing agentic AI beyond engineering? And what they've decided to do is essentially, he talks about these agentic pods. And this to me seems directionally how this should be done, which is you find engineers and you, as we talked about, forward deployed engineers, fancy way of saying put an engineer, put them into departments and have them work with the department heads who understand systems thinking, how their process is done. And, um, he, uh, he says it's making— basically, long and short of it is they're making massive, massive progress on the operational side of the business. So Brad, you're pretty familiar with Uber and have been a long supporter of that. This is a company that knows how to deploy technology pretty well, and they're an operations machine run by an operations machine, Dara.”

“Yeah, we've seen 90% reductions in the price of tokens for each of the last 2.5 years. We've talked a lot about Jevons paradox, which I think you're referencing here, which is you're going to use a hell of a lot more when it happens. I think the central debate right now in AI is the one that Chamath keeps pointing us back in the direction of, which is for 18 months since the Deepseek moment, right? When the Deepseek moment happened, the markets fell 40% and there was a reason for that. Many started arguing that the frontier models were screwed, that open source was going to kill them, that they were closing the intelligence gap, that model routing was gonna make it easier and easier to route these tasks to cheap tokens. But despite all of those arguments, and now we're 18 months into this, and I had this back and forth with Gurley a lot. I love open source. I want all the competition in the world. Let's be very clear. But despite all of those arguments, the facts on the field are just the opposite. The share of economic value, right? There's this quote, there's this tweet this week from Jesse Zhang that we ought to pull up here. You know, the economic value, the share of wallet is actually increasing. To the Frontier Labs while the share of tokens, these commodity tokens, is obviously going up, you know, to the other guys. And I had a little back and forth this week with Nikesh on this, kind of trying to suss out why is that the case, right? Because what people would've thought is, oh, cheaper, pretty damn good. 90% is good enough to do all these tasks that you're talking about, Jason. So nobody's going to use the Anthropics and the OpenAIs of the world. But despite that, it looks like their share of wallet has gone up.”

“Okay, so finance, HIPAA, yeah, there's HR data. You're not allowed to put that to work just yet. What I'm finding is once you start using this and getting some gains, it's very addictive. And we were sitting here, Brad, I don't know, maybe in January, and I got that OpenClaw bug. And then I started playing with this Hermes agent, which is not a French company, by the way. They just use French names. It's Nouveau Research or whatever it is. I started playing with that. It's a very peculiar piece of software, but it's a very open piece of software. So I went to OpenRouter, I got my own keys. I've been playing with GLM. Then I talked a little bit about BitTensor on the program known as Tao, $TAO. It's a crypto project. Somebody who is creating a subnet that is putting GLM 5.2 and other models available at really cheap prices. So I all of a sudden experienced because they gave me an API key, having my token costs go down 95%. Yes. And when you have unlimited tokens as an exercise, which is going to come to everybody, eventually everybody's going to learn how to drop the price by 95%. And it's going to happen as well because people like Groq with inference, this is all inference, right? This is what people are using. They're using inference to do this. Well, inference is being impacted 3 or 4 different ways. The software's getting better, open source at the same time. You're going to have distributed networks like Tau, and you're going to have better chipsets from Groq and Cerebras, et cetera. All that's happening at the same time. Once I got down to 95% cheaper, I started setting my agents instead of doing daily runs to doing hourly runs. Then I took my agents from doing one task and I broke them up into 3 agents and have them doing 3 different things. On the hour. And when you start doing hourly tasks and then you wake up in the morning and like 14 jobs have been done, you're like, wait a second, this is completely different. As one example, I have it, has all the All In episodes, all the This Week in Startups episodes, and we set these cron jobs to go find what the new trends are in technology. I have a trend spotting agent running every hour informing me of the top 3 or 4 trends, and I just give it words. Really does change your thinking when costs go down. What do you think tokens are going to cost, Brad, in next year?”

“they're charging $200 an hour. The difference between spending $3 on a cheap model or $15 on an expensive model to replace a $200-an-hour consultant, it's just irrelevant. That inference cost difference is irrelevant if you're getting something that's bulletproof for $15. And so I think that's what we're seeing play out. The best evidence for all of this is just revenue growth, right? We can sit here and speculate all day long as to what this company—”

“Yeah. Yes, I, you know, first I would say Chamath is right. The only question is on what time frame. There's no doubt that there is a lot of money being spent today that is in the experimental bucket, right? Where I, I think there probably isn't direct ROI, Chamath, to your point, but I think we're so early, nobody cares. I think we're so early in terms of enterprise adoption. Remember, the total addressable market here is every single small, medium, large company on the planet. And so we've never seen revenue growth like this because we've never seen a TAM like this.. And if you look at the distribution of revenues across these businesses, it's not like it's concentrated with 4 or 5 customers. There are millions of customers independently, economically making the decision that is rational for them every day that it makes sense, like Praveen at Uber. And of course they're trying to find things on both right now, mostly the cost side, cost takeouts to justify the investments that they're making in, you know, in tokens. But I think we're on the verge. Of breakthroughs in intelligence that's going to dramatically change the revenue side of the equation for a lot of these businesses. Breakthroughs in life sciences, breakthroughs in, in product innovation, etc., where they could not divorce themselves from this even if they wanted to. For example, Jensen Huang has talked many times that all of his design work, all of his design work now at Nvidia, is using AI to design the next generation chip. The machine is building the machine. So you can't get rid of that even if you wanted to. And tiny intelligence advantages at the frontier where he sits are required. Like, there's no way I don't think that Jensen is going to use anything but the best models that he can to build out those capabilities. So I just think that we're not going to see that in the next few years. You're going to see it under the hood, of course. But that occurred at Snowflake. There was tons of optimization that occurred at Snowflake, but their revenue continued unabated. Their revenue growth continued unabated because they further penetrated use cases, further penetrated the enterprise. So let me be provocative here. If these guys end the year over $100 billion, I think that they're on a revenue trajectory that they could 3 to 5x again next year. We've never seen anything like this.”

“S&P 500. So when these accounts are created, all that money goes into the S&P 500. There's no cost. It's a free account for the lifetime of the recipient. And that was packaged into the Invest America Act, which was passed into law a year ago as part of the reconciliation bill. So that's what actually occurred on July 4th of this year. All those accounts were created for all of these kids. That's the reason the Trump Account app is number one in the App Store, because parents started hearing about this and saying, whoa, I need to go download and get this set up for my child. We had over a million and a half accounts created in the first 24 hours after the launch of this, we had over a billion dollars of deposits. So I was contributing money into the accounts of my nieces, my nephews, my kids, friends' kids. Every account app has a QR code, Jason. So somebody can just send you the QR code for your kid. You double, you know, Apple Pay on your phone, double-click, and you send them $25 or $50. So that is kind of the, the most essential part of it. But we also had a bunch of announcements around philanthropy.”

“And I, I think the market gap, Brad, correct me if I'm wrong, but basically the, the market gap that was created here is is that you can't get an IRA, which is basically a tax-advantaged savings account, until you have your first job, right?”

“Just to be clear, you can get access to that when you're 18, 19, 20 years old and start putting it towards school, or you can roll it into your IRA, Roth IRA, I guess, your, your, your retirement account. Obviously, Michael and Susan Dell were the, were the anchors here. It's over $6 billion, $250 for each of 25 million children, primarily lower and middle income kids. SpaceX's president, Gwynne Shotwell, she joined the party, put $350 million in her SpaceX shares, uh, and for children of lower income communities. So with this, there's a device or some way to do it so you can target specific communities by geo, uh, or by, I guess, their net worth.”

“Great. So as you guys know, the idea was very simple. $1,000 for every child at birth that could compound for their life in a privately owned investment account. So you're born, you get a Social Security number. And you get an investment account. And if you do that and you start with $1,000 and somebody matches that and you save $10 a week, that's $50,000 at age 18. And the idea—”

“That's right, that's right. All the numbers you hear me quote, the $50,000 and the $200,000, it doesn't assume maxing. You know, that just assumes people are adding $50 a month because I've been focused as Michael and others have really on the families who, who don't have the capacity to save today. We're getting all of them into the game. And the president directed us. He said, listen, we have 529 accounts that already help the top 10%. That's not who we're focused on. This is about the Main Street agenda. This is about all the families that he ran for to feel left out and left behind. And we're reconnecting them to the American dream through universal ownership. They all have their own account. They all have a private account on their phone. It's a game changer for the country. It's the largest change to our social contract since 1935 and Social Security. And importantly, I think it couldn't come at a better time. You know, we have this— we have this fight.”

“Uh, we, we— our intention is to get all 50 to 70 million accounts created over the course of the next 90 days using all of this data. But, you know, listen, we got to work through Treasury, the White House, Social Security, etc. They're definitely—”

“Well, and the throttle. You know, paradoxically to all of this might not be the software, might not be the chips, it might be energy. Yeah, Chamath, I mean, when you look at your data center projects and the other ones that are going out there, if we need more tokens, if people need more inference, we have a gating factor in the United States, which is, which is energy.”

“There's, uh, an analysis that my team put together, which I think is quite staggering. If you just look at the load growth that's expected between now and 2050, We are about 3 entire Californias worth of energy short, and that's just assuming regular consumption of devices and cars, fridges, televisions, and computers. So yeah, we have a— we have an enormous problem in the United States with respect to electrons.”