Socialists Sweep NYC, China Catches Up in Coding, AI Memory Crunch, Micron's Blowout Quarter

“No. Well, one, DRAM is the most important bottleneck. There's a whole segment of people on X who are very focused on bottleneck— bottlenecks. I call them the bottleneck bros. You know, they'll, they'll do some work with Claude, find some esoteric Japanese company. The bottleneck that matters is DRAM and DRAM and HBM DRAM. This is the most important bottleneck simply because memory capacity and bandwidth are foundational to the performance of every AI model. So this is the most important bottleneck. Elon is focusing the Terafab on memory because he sees it as the most important bottleneck. You know, not lasers, not capacitors, not power supply semiconductors, not NAND flash, not HDDs, DRAM. And I think this bottleneck is going to be with us for a while. And it is kind of astonishing. So I think, so a few thoughts, like what was important about the quarter? They announced that they have these SCAs, these supply chain agreements. That have a floor and a ceiling for prices with increasingly large group of large customers. And this covers essentially 50% of their revenue, I think, which is 4 customers. And the floor pricing in these new contracts is ahead of prior cycle peaks from a gross margin perspective. And so this is really I think maybe end up being very transformational for the industry. Most other parts of the semiconductor supply chain have rerated. Lam Research, the wafer fab equipment suppliers, they all trade at huge premiums to DRAM relative to prior cycles, and their business models have improved. But so has the industry structure and business models of DRAM because HBM DRAM is increasingly a customized chip. But as far as other people being able to do this, look, CXMT is going public in China. They're going to— they may be the cure for Apple's ills. They will flood the market with, to some degree, cheap consumer-grade DRAM. But for the DRAM you need in these AI servers, there are 3 companies that can make it. It's really hard to do. This is as close to magic as science can get. And I think TeraFab is going to be an important part of this solution. But these stocks still trade or cross-sectionally cheap relative to the rest of AI. Something I've been thinking about memory is DRAM is probably going to be 30% to 40% of all hyperscaler CapEx next year. Every hundreds of billions of dollars that are spent, it's going straight to DRAM. It's wild. But this may actually be very valuable for society because it is probably going to inflate the cost of building a gigawatt data center to the point where even for the hyperscalers, economics matter, we're caught in this prisoner's dilemma. And this may give us as a society time to adapt, to adapt what our friend Brad Gerstner calls the social contract. So the high iPhone prices, you know, one, CXMT is coming for consumer-grade DRAM, but two, this may be good for AI. It may be good for us as a society.”

“Yeah, in the hosted version, you're not going to get a great answer of what agenda you should use for tourism in the great country of Taiwan or your visit to Tiananmen Square. All right, let's keep moving here on the doc at Micron smashed their earnings. If you don't know Micron, they're one of only 3 companies on Earth that make high-bandwidth memory. These are specialized chips. They sit on top of the NVIDIA GPU, and their entire 2026 supply is sold out and has been for some time. SK Hynix and Samsung also make HBM. Micron smashed earnings, revenue up 4x, 4x year over year, $9 billion to $42 billion. Beat expectations by 16%. Big jump in guidance for Q4, $50 billion versus $43 billion. Their stock is up 10x. Shout out to Gavin. In our 2025 prediction show, he gave a call on HBM makers like Micron as the best performing asset. Since that time, Micron up 14x. I'm not crying in my soup. You're not crying in your soup. I got a ton of information here. I think this, I'll just end on the Apple price increases. Everybody knows Apple has been really, uh, been a beneficiary of the run local models movement that I'm part of, and, and people are buying 128GB, 256GB MacBook Pros, Mac Studios. But the gig is up apparently, because now Apple, which had not passed on those costs to customers, is having to pass those increases on. So Everything from, you know, the new MacBook Neo, which is their $699 laptop, you know, kind of competing with Chromebooks, is now $799, up 15, 14%. And Mac Studio up 25%. The costs are just going to be very significant. Inflation has come to the desktop. Your thoughts, Gavin, on Micron and the impact on the industry? And is this a temporary bottleneck, or does this mean everybody has to get into this business quickly?”

“No. Well, one, DRAM is the most important bottleneck. There's a whole segment of people on X who are very focused on bottleneck— bottlenecks. I call them the bottleneck bros. You know, they'll, they'll do some work with Claude, find some esoteric Japanese company. The bottleneck that matters is DRAM and DRAM and HBM DRAM. This is the most important bottleneck simply because memory capacity and bandwidth are foundational to the performance of every AI model. So this is the most important bottleneck. Elon is focusing the Terafab on memory because he sees it as the most important bottleneck. You know, not lasers, not capacitors, not power supply semiconductors, not NAND flash, not HDDs, DRAM. And I think this bottleneck is going to be with us for a while. And it is kind of astonishing. So I think, so a few thoughts, like what was important about the quarter? They announced that they have these SCAs, these supply chain agreements. That have a floor and a ceiling for prices with increasingly large group of large customers. And this covers essentially 50% of their revenue, I think, which is 4 customers. And the floor pricing in these new contracts is ahead of prior cycle peaks from a gross margin perspective. And so this is really I think maybe end up being very transformational for the industry. Most other parts of the semiconductor supply chain have rerated. Lam Research, the wafer fab equipment suppliers, they all trade at huge premiums to DRAM relative to prior cycles, and their business models have improved. But so has the industry structure and business models of DRAM because HBM DRAM is increasingly a customized chip. But as far as other people being able to do this, look, CXMT is going public in China. They're going to— they may be the cure for Apple's ills. They will flood the market with, to some degree, cheap consumer-grade DRAM. But for the DRAM you need in these AI servers, there are 3 companies that can make it. It's really hard to do. This is as close to magic as science can get. And I think TeraFab is going to be an important part of this solution. But these stocks still trade or cross-sectionally cheap relative to the rest of AI. Something I've been thinking about memory is DRAM is probably going to be 30% to 40% of all hyperscaler CapEx next year. Every hundreds of billions of dollars that are spent, it's going straight to DRAM. It's wild. But this may actually be very valuable for society because it is probably going to inflate the cost of building a gigawatt data center to the point where even for the hyperscalers, economics matter, we're caught in this prisoner's dilemma. And this may give us as a society time to adapt, to adapt what our friend Brad Gerstner calls the social contract. So the high iPhone prices, you know, one, CXMT is coming for consumer-grade DRAM, but two, this may be good for AI. It may be good for us as a society.”

“No. Well, one, DRAM is the most important bottleneck. There's a whole segment of people on X who are very focused on bottleneck— bottlenecks. I call them the bottleneck bros. You know, they'll, they'll do some work with Claude, find some esoteric Japanese company. The bottleneck that matters is DRAM and DRAM and HBM DRAM. This is the most important bottleneck simply because memory capacity and bandwidth are foundational to the performance of every AI model. So this is the most important bottleneck. Elon is focusing the Terafab on memory because he sees it as the most important bottleneck. You know, not lasers, not capacitors, not power supply semiconductors, not NAND flash, not HDDs, DRAM. And I think this bottleneck is going to be with us for a while. And it is kind of astonishing. So I think, so a few thoughts, like what was important about the quarter? They announced that they have these SCAs, these supply chain agreements. That have a floor and a ceiling for prices with increasingly large group of large customers. And this covers essentially 50% of their revenue, I think, which is 4 customers. And the floor pricing in these new contracts is ahead of prior cycle peaks from a gross margin perspective. And so this is really I think maybe end up being very transformational for the industry. Most other parts of the semiconductor supply chain have rerated. Lam Research, the wafer fab equipment suppliers, they all trade at huge premiums to DRAM relative to prior cycles, and their business models have improved. But so has the industry structure and business models of DRAM because HBM DRAM is increasingly a customized chip. But as far as other people being able to do this, look, CXMT is going public in China. They're going to— they may be the cure for Apple's ills. They will flood the market with, to some degree, cheap consumer-grade DRAM. But for the DRAM you need in these AI servers, there are 3 companies that can make it. It's really hard to do. This is as close to magic as science can get. And I think TeraFab is going to be an important part of this solution. But these stocks still trade or cross-sectionally cheap relative to the rest of AI. Something I've been thinking about memory is DRAM is probably going to be 30% to 40% of all hyperscaler CapEx next year. Every hundreds of billions of dollars that are spent, it's going straight to DRAM. It's wild. But this may actually be very valuable for society because it is probably going to inflate the cost of building a gigawatt data center to the point where even for the hyperscalers, economics matter, we're caught in this prisoner's dilemma. And this may give us as a society time to adapt, to adapt what our friend Brad Gerstner calls the social contract. So the high iPhone prices, you know, one, CXMT is coming for consumer-grade DRAM, but two, this may be good for AI. It may be good for us as a society.”

“Microsoft raised the price of the Xbox. You know, it's coming for the Switch, it's coming for the PlayStation. You know, there's demand destruction 'cause the prices in consumer, whereas AI demand is relatively price insensitive. David, I would modify one statement. It's hard to build a Micron, it's hard to build a new fab in a deep blue state. You can build fabs.”

“I mean, come on, you got to be able to see through it. He was making Knicks references and it was the most get out of your seat, standing ovation cheer speech I've ever heard about the Knicks. And I was infuriated. This fucking guy is such a good orator. You got programmed. I hate him. We're going to have to deprogram you now. No, no, no. Hypnotized and I just pulled myself out. I got like pulled in for a second. Next topic. Next topic for sure. Oh my God, great, good, strong first topic. I didn't know you guys were going to go all in, so to speak. Topic 2: Chinese open source models appear to be catching up with the US frontier models. Let's start with GLM 5.2 released by China's Z.AI. This is a frontier class open source free to download Anywhere model, 744 billion parameters, 1 million token context window, and it's under the MIT license. If you don't know what that is, open source licenses have very— Super open source. It's super open source. Thank you. The most open source. The most open source of all open source. If you open it up, you can use it however you like. You can fork it, you can build your own company based on that. No regional restrictions, no API, Fully self-hostable. No, uh, no Daria. It's just yours.”

“Distillation is when you have, you know, a, a, like, you know, we all have seen videos of these Chinese iPhone farms. Just picture a farm like that. Tens of thousands of phones, iPads, and computers that are asking the Claude API through masked accounts, very specific questions. And then these, what's called reasoning traces, are being harvested. Because if you're on the API, you know, you get to see every token. And those reasoning traces are then fed back into the model during the reinforcement learning process and probably during the pre-training process. And that is a way that you can get really, really close to the frontier at a fraction of the cost. And this is for sure going on. And has gone on for a long time. I do think it's a cheat sheet.”

“I think that on a certain level it is what they wanted because they've been advocating to have a federal regulator, basically a new agency. In fact, Dario posted a blog just a few weeks ago saying he wants an FAA for AI. They wanted government approval, a government approval process for AI models. And so in a sense, they've gotten exactly what they wanted. Now, that being said, I don't think they're happy about the fact that Fable has been rolled back. So in a sense, you could say that Dario got hoisted on his own petard here, or it could be a F-A-F-O situation. But look, my view on it is we should not reward Dario by giving him exactly what he's always craved, which is some sort of labyrinthine government approval process that does reward regulatory capture. So I hope that very soon now, I do think that as long as Anthropic has resolved the jailbreak issue, then I do think they should be allowed to come back to market. And similarly for OpenAI, I don't think we should be delaying them unnecessarily. We do not have months to give away in this race. And let me just say one other thing, which again, it's something I've been saying for months, which is with respect to risks like cyber, It is undoubtedly a risk, but what is the response to that? The only thing you can do is go out and find all the vulnerabilities first yourself. The white hats, have the white hats find all the vulnerabilities and do a big upgrade cycle, roll out the patches before they can be exploited. The reality is that if you just clamp down in a way that doesn't even allow these models to be used, The Chinese are going to have these capabilities imminently anyway. You know, they're already at Opus 4.8 level. And the founder of Z.AI, he said that before Q1, they'll have Fable level capability. I believe them because look, the Chinese have been, I'd say, 9 months behind our models, plus or minus 3 months depending on capability. But it's when they know there's been a breakthrough around something like cyber, they can deploy more resources against that particular problem and catch up faster.”

“So let me make a point about that. So on the silicon, there's been a huge push in China by the government to push their AI labs to develop and train on Huawei chips. And look, you can take these claims with a grain of salt, maybe they're not true, but it was claimed that DeepSeek V4 was trained on Huawei chips, and now Z.AI, they are saying that the GLM-5 family was trained entirely on clusters of Huawei Ascend 910B chips. So now look, maybe they're lying, maybe they smuggled in some Nvidia chips, but the claim is that this was all done on indigenous chips. And what I believe is that China is engaged in a strong indigenization push right now. They want to prop up Huawei as the national champion. They want all their companies using Huawei chips. They still need to scale some of the manufacturing, but they're going to do that pretty quickly. And then what they're going to do is they're going to take these Huawei chips, they're going to take these Huawei-optimized models. Remember, that GLM 5.2, the inference is optimized for the Huawei chips. Okay, we know that. And they are basically going to package these things up, they call it AI in a box, they're going to sell it at a fraction of the cost globally, which is what they do with every technology, right?”

“Yeah. And that's another thing is I've been saying since the beginning of this administration, we have to be pro-export because China is going to be there within 1 or 2 years. I said we're going to be kicking ourselves because we could have had the whole global market to ourselves. We invented reasons not to sell abroad to our friends and partners, and now China is going to be there imminently.”

“So to stand up a 1 gigawatt data center $35 billion in semiconductors, NVIDIA semiconductors, and then it's $25 billion of power and cooling equipment. And that is clearly inflationary because a lot of that $25 billion is the human labor required to install it. So the calculation that needs to be done for orbital compute is it's $35 billion of silicon in each space, and in literally outer space, in orbit, and on ground. But if you can get the cost of launch significantly below that $25 billion, then the math starts to really math. And when Starship is reusable, it's going to cost $5 billion to put a gigawatt of compute into space. And something that drives me crazy is people picture these Pentagon-sized data centers. No, it's racks in space linked with lasers. It's, it's kind of a virtual data center in space.”

“$5 billion of launch cost. Now you're $40 billion to put the gig into space, you're at $60 billion terrestrially, and the $25 billion that is power and compute is clearly inflationary. And so it may be that in 3 or 4 years, it's $70 billion versus $40 billion, and that 5, as Starship becomes rapidly reusable, is likely deflationary. So this is the economics that underpin orbital compute from first principles. And then on an ongoing basis, you are, you're maybe paying $1 billion a year for the power to run those chips and cool them.”

“So to stand up a 1 gigawatt data center $35 billion in semiconductors, NVIDIA semiconductors, and then it's $25 billion of power and cooling equipment. And that is clearly inflationary because a lot of that $25 billion is the human labor required to install it. So the calculation that needs to be done for orbital compute is it's $35 billion of silicon in each space, and in literally outer space, in orbit, and on ground. But if you can get the cost of launch significantly below that $25 billion, then the math starts to really math. And when Starship is reusable, it's going to cost $5 billion to put a gigawatt of compute into space. And something that drives me crazy is people picture these Pentagon-sized data centers. No, it's racks in space linked with lasers. It's, it's kind of a virtual data center in space.”

“You know, one other, uh, point here that I guess is, it might be relevant to SpaceX AI, although it's, don't have to limit it to this, is, um, I think there's an assumption that over time it would get cheaper and easier to stand up new data centers, right? But what you're saying is actually it might be getting harder, it might be getting more expensive, right? Because there's competition for these components. The memory is getting more expensive. I'm not sure that the GPUs are getting any cheaper.”

“You have a couple of issues right now to turn on compute terrestrially. So assume you have land, that's relatively straightforward. Assuming you can get it zoned, less straightforward. Assuming you can get power, very difficult. Then you have a very critical design decision. So all of these folks publish these things called the basis of design, and your, your BODs essentially tell you, here's the Anthropic spec, here's the OpenAI spec, here's the CoreWeave spec, here's the AWS spec, here's GCP, and you get these 50 and upwards of 500-page documents of all these technical details. The issue that we have is, I don't know if you guys have used OpenAI or Anthropic recently where you get the whole thing of come back later, right? Yes. That come back later is completely unacceptable. It just means that they have no compute.”

“And to riff on what Thomas said and all of this, one, I mean, there is actually a startup that is trying to put 4 GPU units with kind of a battery on people's houses and give them a discount on their power. And then you can do inference for that neighborhood from those 4 GPUs and it's like locked sealed so nobody can get in. But there's another dynamic that I think we should talk about with all of this and can play into the megapods and other people are working on data centers. Crusoe is working on modularly assembling data centers, like a data center and think of it as like an 18-wheeler. What do you call those things? The 18-wheeler, you know, shipping container, whatever it is. But that is the disaggregation of inference into prefill and decode. When you, when a model is answering your question, it's doing two things. The prefill part is understanding the question and its answer thus far. And think of, think of that as the more you can remember, the bigger your, you know, your memory capacity, Literally, the more words you can remember, the better. Decode is the process of generating the next token. And that is a memory bandwidth bound problem. And think of it as the faster you can speak, the better. And these two types of inference are increasingly being disaggregated. And Chamath was an investor in Groq, which NVIDIA bought, and they're gonna use this. Cerebras is the other solution today that is available. And you can put Groq or Cerebras decode-optimized chips. They would both say you can do more than decode on them, and that's true, in front of old NVIDIA GPUs like H100s. So you can lift H100s, A100s out of some old data center, put them in one of these megapods, a rack in a shipping container, Put a Groq or a Cerebras in front of it and you can get a very competitive solution. And so I do think the disaggregation of inference, we're going to be using GPUs for 7 years, 10 years, 12 years, and that's great because it lowers the cost to finance them, which makes this AI revolution more financeable. Chamath, you want to riff on Groq? Should we, should we, should we?”

“This is 100% true, but everything that cuts against training works for inference.”

“And to riff on what Thomas said and all of this, one, I mean, there is actually a startup that is trying to put 4 GPU units with kind of a battery on people's houses and give them a discount on their power. And then you can do inference for that neighborhood from those 4 GPUs and it's like locked sealed so nobody can get in. But there's another dynamic that I think we should talk about with all of this and can play into the megapods and other people are working on data centers. Crusoe is working on modularly assembling data centers, like a data center and think of it as like an 18-wheeler. What do you call those things? The 18-wheeler, you know, shipping container, whatever it is. But that is the disaggregation of inference into prefill and decode. When you, when a model is answering your question, it's doing two things. The prefill part is understanding the question and its answer thus far. And think of, think of that as the more you can remember, the bigger your, you know, your memory capacity, Literally, the more words you can remember, the better. Decode is the process of generating the next token. And that is a memory bandwidth bound problem. And think of it as the faster you can speak, the better. And these two types of inference are increasingly being disaggregated. And Chamath was an investor in Groq, which NVIDIA bought, and they're gonna use this. Cerebras is the other solution today that is available. And you can put Groq or Cerebras decode-optimized chips. They would both say you can do more than decode on them, and that's true, in front of old NVIDIA GPUs like H100s. So you can lift H100s, A100s out of some old data center, put them in one of these megapods, a rack in a shipping container, Put a Groq or a Cerebras in front of it and you can get a very competitive solution. And so I do think the disaggregation of inference, we're going to be using GPUs for 7 years, 10 years, 12 years, and that's great because it lowers the cost to finance them, which makes this AI revolution more financeable. Chamath, you want to riff on Groq? Should we, should we, should we?”

“Yeah. I mean, a lot of people who had large SpaceX positions were large buyers on the IPO. On Cerebras, so Cerebras has had a tough 2 days since they reported their first quarter as a public company. And I think there are two things that are worth discussing here. One is there is a whole generation of portfolio managers. There's a lot of people who are advocating for squeezing the blood out of the stone on IPO prices. And the flip side of that is that there are a lot of portfolio managers who, if a stock breaks deal price, they sell it. No matter what. They consider it a promise that was broken. And so this is what has happened with Cerebras to some degree over the last 2 days. And this may seem irrational, but there are people who run giant funds who I know personally, where if a stock breaks deal price, they sell no matter what. And so if stock breaks deal price, it can sometimes go to places you wouldn't think it would go. And this means that shorts, if a stock gets close to deal price, they short it because they want to break deal price. And then, you know, they make a quick 10 or 20%. So it becomes a pile-on because you have this price-insensitive selling that can be triggered. And this is what has happened to Cerebras. You know, people talk about hate sale, hate selling, but they broke deal price. And so just if you're going public and you're listening to this, tell your bankers, price this in such a way that we're not going to break deal price in our first 9 months as a public company. And that's what I always advise everyone to do. And I think it's important, but I also think it takes companies a while to learn how to tell their story and communicate to public markets. It's a very different audience than VCs. And the way I would've respectfully told the Cerebras story, because what happened to Cerebras is They reported a quarter and they're growing fast, but relative to the rest of AI, they're not growing that fast in the March quarter. So what I would have said is we signed this transformational, you know, $20, $25 billion, I don't know the exact number, contract with OpenAI in December, December of 2025. We immediately ordered more wafers from TSMC. It takes 100, it takes 4 months. From when we make that order, TSMC, you know, says yes, they start producing, takes 4 months to make the chip. Then we— then it takes us 2 months plus or minus to turn that chip into a server. And then if we're lucky and we can find the power, it takes us a month to energize that chip and start making tokens with it. So the first time you're going to see the impact —of this OpenAI deal at the earliest is probably around Labor Day. So you'll see a little bit of it in the third quarter, but then it really starts to build. And just like really simple math. So let's just use some rough numbers. Let's take some NVIDIA numbers. It takes them $35 billion to bring on a gigawatt and $15 billion that you can generate $15 billion in token revenue and cloud revenue out of that gigawatt.. And somebody on the call talked about adding 50 megawatts a month. If they could add 50 megawatts a month in 2027, forget '26, that means they exit the year at roughly a $9 billion cloud computing run rate. And we're at less than $40 billion of market cap. Now that is going to be really hard to do, and they've never done anything like that before. I would focus, as an investor, I think what matters here is not where they sit competitively, not what new demand they can bring on, but just how quickly can they bring on power. Listen, outside of the hyperscalers, the only companies that have ever brought on more than a gigawatt I think are CoreWeave, Crusoe, and xAI. So bringing on 600 megawatts, it's really hard. And that is what I'm focused on as an investor. How many megawatts can they bring on? Because we know what they're going to monetize at, and that is the question, and we'll see. Yeah.”

“So guys, I've got 500 properties with lots of energy, lots of mechanical cooling systems, and lots of LNG access, or sorry, not LNG, natural gas, sorry, already piped in. So on all the stuff I'm doing on robotics, AI, physical AI, We're literally looking at putting some of our compute into our kitchens.”

“If OpenAI, which launched their Jalapeño chip this week and announced it being built by, I believe, Broadcom, and they are saying, hey, fuck you to Jensen and NVIDIA, and they already were full contact with him, don't be surprised if NVIDIA says, "You know what? We kind of like the area you're operating in now that you're going to make chips," and maybe OpenAI sells those chips to other people. Don't be surprised if NVIDIA starts an OpenAI competitor. You heard it here first on All In. Sacks, did you want to jump in there?”

“No. Well, one, DRAM is the most important bottleneck. There's a whole segment of people on X who are very focused on bottleneck— bottlenecks. I call them the bottleneck bros. You know, they'll, they'll do some work with Claude, find some esoteric Japanese company. The bottleneck that matters is DRAM and DRAM and HBM DRAM. This is the most important bottleneck simply because memory capacity and bandwidth are foundational to the performance of every AI model. So this is the most important bottleneck. Elon is focusing the Terafab on memory because he sees it as the most important bottleneck. You know, not lasers, not capacitors, not power supply semiconductors, not NAND flash, not HDDs, DRAM. And I think this bottleneck is going to be with us for a while. And it is kind of astonishing. So I think, so a few thoughts, like what was important about the quarter? They announced that they have these SCAs, these supply chain agreements. That have a floor and a ceiling for prices with increasingly large group of large customers. And this covers essentially 50% of their revenue, I think, which is 4 customers. And the floor pricing in these new contracts is ahead of prior cycle peaks from a gross margin perspective. And so this is really I think maybe end up being very transformational for the industry. Most other parts of the semiconductor supply chain have rerated. Lam Research, the wafer fab equipment suppliers, they all trade at huge premiums to DRAM relative to prior cycles, and their business models have improved. But so has the industry structure and business models of DRAM because HBM DRAM is increasingly a customized chip. But as far as other people being able to do this, look, CXMT is going public in China. They're going to— they may be the cure for Apple's ills. They will flood the market with, to some degree, cheap consumer-grade DRAM. But for the DRAM you need in these AI servers, there are 3 companies that can make it. It's really hard to do. This is as close to magic as science can get. And I think TeraFab is going to be an important part of this solution. But these stocks still trade or cross-sectionally cheap relative to the rest of AI. Something I've been thinking about memory is DRAM is probably going to be 30% to 40% of all hyperscaler CapEx next year. Every hundreds of billions of dollars that are spent, it's going straight to DRAM. It's wild. But this may actually be very valuable for society because it is probably going to inflate the cost of building a gigawatt data center to the point where even for the hyperscalers, economics matter, we're caught in this prisoner's dilemma. And this may give us as a society time to adapt, to adapt what our friend Brad Gerstner calls the social contract. So the high iPhone prices, you know, one, CXMT is coming for consumer-grade DRAM, but two, this may be good for AI. It may be good for us as a society.”

“The top and the bottom, just like any other Saturday night. Here's the trademark for Megapod that came out. This is a, uh, filing date of 6/18/2026, so a very recent June 18th. Modular data center hardware for artificial intelligence computing comprised of network— of computer servers, computer hardware for artificial intelligence processing, computer network hardware, electric power distribution units and cooling systems sold as a unit. Self-contained modular computing hardware systems for artificial intelligence workloads, yada yada yada.”

“Yeah, I think that is roughly where it would probably trade as a public company. And wow. Yeah, I mean, look, they're going to do— they're going to end this year—”

“by the— David, can I ask, do you think Dario got exactly what he wanted? It seems to me there's some chance this has been a very calculated strategy to provoke the US government into doing what they just did, and this is what he wants. He has a regulatory moat now. He can keep his future models behind this, you know, give it out to Glasswing, use it to distill it for him for themselves. Do you think this is what they wanted?”

“Honestly, I think that we are losing the script and part of it is because we've been our own worst enemy. I'll just keep saying this, that I think AI is a very good prism into this problem. I think AI is the greatest economic leveler thing that we'll ever find in our lifetime. I think it's the thing that can create the greatest amount of equality. I think that it can even the starting line for every single person on Earth. But we've done such a poor job in representing it, in bringing it to market, in talking about it. We've let all of our own personal trials and tribulations and insecurities and fights spill out into the open. As a result, Silicon Valley has lost even more credibility with the people at large. And in that vacuum, what other people can paint is a picture of how anything other than what capitalism looks like today is a better version of what they see. And this is why you're seeing, I think, a lot of these people get a lot of momentum. I think if you look at some of the key congressional races, they were a referendum on AI. And the good news is we were able to hold the line. In some of these key places, but just barely. In Utah and in New York, there was a couple of very important races where it was essentially Anthropic-funded anti-AI groups, which is again insane, against in some ways OpenAI-funded pro-AI groups, and the pro-AI groups won.”

“Yeah, zero-zero is, uh, the— how the Knicks talked about the next game in every series. They're like, it's zero-zero, we come into this as if it's like the first game of the series even if we're up 2 games or 3 games. The socialists have swept New York in the congressional Democratic primaries. On Tuesday, New York City Mayor Adams went 3 for 3 in the candidates which he endorsed, and they all won their primaries. 10th District leader Brad Lander defeated two-term incumbent Dan Goldman. 10th is one of New York's richest districts, includes the West Village, all those townhouses, uh, Wall Street, Dumbo, Cobble Hill, Carroll Gardens, Park Slope. That's some weird geography put together there. In New York's 13th, Chevalier beat a five-term incumbent who was backed by House Speaker Hakeem Jeffries, and apparently the socialists are coming for him. 13 is one of the poorest districts, Harlem and the West Bronx, the boogie-down Bronx. She's a 32-year-old Democrat socialist with a history of spicy remarks. New York's 7th District, Claire Valdez won the open seat over the handpicked successor, the incumbent. 7 is a DSA stronghold, Bushwick, Williamsburg, Long Island City, Greenpoint, known as the Commie Corridor. A lot of hipsters and baristas with suspenders in that neighborhood. According to our partners, Polymarket, this was a pretty big upset. The Adams sweep chances were just 26% before Election Day. Yeah, that would be like the trifecta there if you were gambling. These candidates, just like Adams, did a lot better with younger, college-educated, and high-income folks— the folks who can afford to be socialists. And these are all safe Democratic seats. The DSA will very likely win. So the DSA caucus—”

“Well, and by the way, I, I, I don't think it's just New York. I mean, that race for mayor in LA where, was it, Rahmen somehow beat Karen Bass thanks to ballot harvesting after voting day, or I should say votes that were found and counted after Election Day. I mean, that will be a test of the DSA because they are highly organized and they have learned how to take advantage of all these rules, these ballot harvesting rules and all these types of things. The DSA, I think they've got something like half the city council seats now in LA and they're growing. So especially in these low turnout elections, and look, this was a Democratic primary in New York, which is a strongly blue state. So I think maybe 17%. Turned out. So it was very low, but this is where the DSA really thrives and excels because they care passionately and they're highly organized and they know how to take advantage. This is why they want all these like ranked choice voting and all these types of things. They know how to manipulate and take advantage of those kinds of systems. So I think that you're gonna see this in lots of other jurisdictions wherever they're organized. I think LA will be a really interesting test. So I don't think we can just chalk this up to Zohran's popularity. This is a national movement and we're going to see it in a lot of places, but there's no question that Mamdani is now kind of the spiritual leader. I mean, a lot of these people, they think AOC is a sellout or Bernie's a sellout. They are way more radical than even those types.”

“I think there is some truth to that. I mean, I think the choices of the future are gonna be communism, or if you want to call it socialism, of the Democrat Party, or nationalism in the Republican Party. I mean, that is where we're headed. Those are the two populist directions. But let's look at what these DSA candidates stand for. So let's look at what their platform is. They actually say they want to abolish the Senate, They want to abolish the carceral state. That means basically police forces and prisons. They want to abolish ICE and grant amnesty for all. They do not support any deportations whatsoever. They want to replace the president and Supreme Court with an executive and judiciary that is chosen by and subordinate to Congress, which basically now I guess just means this house. And with respect to House elections, they want to abolish the Electoral College. They want to replace the two-party system with a multi-party democracy, and they want to expand the House of Representatives, implement proportionate representation, and ranked choice voting in all elections. So this would be a total makeover of our constitutional system. They want a free Palestine. They want public ownership of major corporations. They want to defund the Department of War. This is a very radical organization, and you would laugh at a lot of these types of proposals, but you can't really laugh at it anymore because these guys are taking over the Democratic Party. And you can see the Democratic establishment is in complete panic right now because they have lost control of the party to Adams, um, Donnie and his, his allies. So, Jason, like you said, I mean, let's take this one race here. New York 13, you've got this ally of Hakeem Jeffries, longtime incumbent Congressman Espaillat, I guess is his name. He is the chair of the Congressional Hispanic Caucus, and he was defeated by an unemployed 32-year-old PhD candidate. She's never had a job. She's been in college for 10 years, I guess, writing this PhD thesis.”

“Yeah, I think the specific stat is like 0.1% commits 70% of the crimes. And if you just dealt with the 0.1%, you'd have effectively no crime.”

“And even the establishment wing of the Democratic Party is now bending the knee. And so what you're going to see is regardless of how many DSA candidates actually get elected, the rest of the party is now responding to this and they're going to bend, they're going to blow in this direction. Because they don't want to get challenged in a primary. I mean, think about this. You had 3 major congressional races where the Mamdani candidate won, 2 of them unseated, you know, really strong incumbents. These were big upsets. So you gotta think now that every congressional race in a pretty blue district, those members are now gonna have to take into account that they could get primaried and they're gonna have to tilt their voting and their views and their rhetoric in this DSA direction. Because they don't want to have happen to them what just happened to Dan Goldman in the New York 10th District. And just to make one last point on that, so J. Cal, you mentioned the Israel issue, and I actually think that is a hugely important and salient issue now in the Democratic Party, in Democratic primaries. Obviously, you saw that as part of the DSA platform, one of the legs is Free Palestine. This defeat of Dan Goldman, two-time congressman, he led the impeachment effort against Trump. He had all the right progressive credentials. He checked the box on all the left-wing policies. He was on the right cable news channels all the time. No one expected him to lose. He lost to Brad Lander really just over this issue of Israel. Dan Goldman is very pro-Israel. He basically defended Israel's actions over the past few years, whereas Lander, who like Goldman is also Jewish, so this was again, you know, White Jewish congressman against white Jewish longtime New York politician. So on paper, they're very similar. It was just in this issue of Israel where they disagreed. And Lander went to a mosque in order to denounce what he called the genocide in Gaza. So this was really as close as you can get to a straight up vote on that one issue in this primary. And Lander won pretty handily. Now, the reason for this is if you look at polling, 80% of Democrats now say they disapprove of Israel. So the approval/disapproval rating, Israel used to have high approval ratings pretty much across the board. I mean, it was sort of a consensus, both Democrats and Republicans. Now, 80% of Democrats say they disapprove. You really can't underestimate how much of a motivator this is for young Democrats.”

“Yeah. I mean, a lot of people who had large SpaceX positions were large buyers on the IPO. On Cerebras, so Cerebras has had a tough 2 days since they reported their first quarter as a public company. And I think there are two things that are worth discussing here. One is there is a whole generation of portfolio managers. There's a lot of people who are advocating for squeezing the blood out of the stone on IPO prices. And the flip side of that is that there are a lot of portfolio managers who, if a stock breaks deal price, they sell it. No matter what. They consider it a promise that was broken. And so this is what has happened with Cerebras to some degree over the last 2 days. And this may seem irrational, but there are people who run giant funds who I know personally, where if a stock breaks deal price, they sell no matter what. And so if stock breaks deal price, it can sometimes go to places you wouldn't think it would go. And this means that shorts, if a stock gets close to deal price, they short it because they want to break deal price. And then, you know, they make a quick 10 or 20%. So it becomes a pile-on because you have this price-insensitive selling that can be triggered. And this is what has happened to Cerebras. You know, people talk about hate sale, hate selling, but they broke deal price. And so just if you're going public and you're listening to this, tell your bankers, price this in such a way that we're not going to break deal price in our first 9 months as a public company. And that's what I always advise everyone to do. And I think it's important, but I also think it takes companies a while to learn how to tell their story and communicate to public markets. It's a very different audience than VCs. And the way I would've respectfully told the Cerebras story, because what happened to Cerebras is They reported a quarter and they're growing fast, but relative to the rest of AI, they're not growing that fast in the March quarter. So what I would have said is we signed this transformational, you know, $20, $25 billion, I don't know the exact number, contract with OpenAI in December, December of 2025. We immediately ordered more wafers from TSMC. It takes 100, it takes 4 months. From when we make that order, TSMC, you know, says yes, they start producing, takes 4 months to make the chip. Then we— then it takes us 2 months plus or minus to turn that chip into a server. And then if we're lucky and we can find the power, it takes us a month to energize that chip and start making tokens with it. So the first time you're going to see the impact —of this OpenAI deal at the earliest is probably around Labor Day. So you'll see a little bit of it in the third quarter, but then it really starts to build. And just like really simple math. So let's just use some rough numbers. Let's take some NVIDIA numbers. It takes them $35 billion to bring on a gigawatt and $15 billion that you can generate $15 billion in token revenue and cloud revenue out of that gigawatt.. And somebody on the call talked about adding 50 megawatts a month. If they could add 50 megawatts a month in 2027, forget '26, that means they exit the year at roughly a $9 billion cloud computing run rate. And we're at less than $40 billion of market cap. Now that is going to be really hard to do, and they've never done anything like that before. I would focus, as an investor, I think what matters here is not where they sit competitively, not what new demand they can bring on, but just how quickly can they bring on power. Listen, outside of the hyperscalers, the only companies that have ever brought on more than a gigawatt I think are CoreWeave, Crusoe, and xAI. So bringing on 600 megawatts, it's really hard. And that is what I'm focused on as an investor. How many megawatts can they bring on? Because we know what they're going to monetize at, and that is the question, and we'll see. Yeah.”

“Holy shit. What's the '28 number? '28 number, is it $200? Is it $300 billion? It's probably not going to trade at 10 times that number, and it will be very profitable at that scale because it'll be inference dominated. And people reporting they have 85% gross margins on inference. But in terms of the market absorbing this, the market's already absorbed it. It's just shifting from private to public. And so in the scale of global capital markets, these seem like really big numbers, but you're just moving from the private markets to the public markets, which are even bigger. As far as SpaceX specifically, I think one of the more important things is everybody who's a SpaceX investor employee has had a chance to sell every 6 months for the last 10 years. So there may not be the wall of liquidity that some people are thinking about. I read this New York hedge fund short, short report that you could just short SpaceX on the lockup because so many people are going to sell. Really? Well, everybody who's on the cap table, they had an opportunity to sell. And almost half the employees at SpaceX bought on the IPO.”

“Now, I do think Cerebras But Gavin, one thing though, I got to say, so I've had SpaceX since 2018. What, what? But their little liquidity thing every year, like last year it was like 350 billion last year. So you got an 8x in one year. You could have a lot of people selling, right? For sure. They were doing little 20% up, 30% up clips for many years. And then an 8xer could create that liquidity.”

“What we did was every person who wanted to be involved had to fill out a sheet of how much money they put it at, 10, 11, 12, 13, 14, all the way up to 20. And then we just did the Dutch auction. We said we want to clear 1.5 billion. We just did the auction and cleared it, went back to people and said, hey, you're not going to get it. You have another shot. They update their Excel sheet and it just moves the number up a little bit and then you close it down. But yeah, it started, that round started at I think it was $9 or $10 or something like that and ended up at $17.”

“It's super awesome. Super awesome. Super awesome. Um, for your super awesome startup Adams, which, man, loads of people have been calling me to say how, like, they've been in the, you know, they've spoken to you and— yeah, investor, the, uh, your, your dog is hunting with investors, Travis. All right, fair enough, fair enough. Okay, you're going to have a router, and every query that somebody comes in, every task that needs to be done at your company, that router is going to send it to, you know, your URL SFT'd version of GLM 4.5 or Gemma. Yes. Then at some point in the workflow, a frontier model may or may not come in to kind of check it, add to it. And that's what I mean by a composable model when you have kind of, um, you know, kind of a symphony of models working together with kind of the frontier models being, you know, maybe the conductors. But that's what I mean when I say composable.”