"All right, this is the episode where my roommate teaches me semiconductors. It's also the send off for this current set. It is. After you use it, I'm like, "I can't use this again. I gotta get out of ..."

"Dylan is the CEO of SemiAnalysis. Dylan, here’s the burning question I have for you. If you add up the big four—Amazon, Meta, Google, Microsoft—their combined forecasted CapEx this year that you publi..."

"And second, what are the labs raising all this money for if the yearly price of a one-gigawatt data center is $13 billion? So when you talk about the CapEx of these hyperscalers being on the order of ..."

"So with roughly 20 gigawatts deployed this year, a big chunk is hyperscalers, and a chunk is not. For all of these companies, their biggest customers are Anthropic and OpenAI. Anthropic and OpenAI are..."

"In a sense, Anthropic needs to get to well above five gigawatts by the end of this year. It's going to be really tough for them to get there, but it's possible. Can I ask a question about that? If Ant..."

"But in reality, he's screwed the pooch compared to OpenAI, whose approach was, "Let's just sign these crazy fucking deals." OpenAI has got way more access to compute than Anthropic by the end of the y..."

"They've gone to random companies, or companies one would think are random, like SoftBank Energy, who has never built a data center in their life but is building data centers now for OpenAI. They've go..."

"Anthropic was a lot more conservative. They were like, "We'll sign contracts, but we'll be principled. We'll purposely undershoot what we think we can possibly do and be conservative because we don't ..."

"Who built the spare capacity such that it's available for Anthropic and OpenAI to get last minute? What is the concrete advantage that OpenAI has gotten if they end up at similar compute numbers by 20..."

"To acquire excess compute, yes, there is capacity at hyperscalers. Not all contracts for compute are long-term, five-year deals. There's compute from 2023 or 2024, or H100s from 2025, that were signed..."

"I’ve seen deals where certain AI labs—I'm being a little bit vague here for a reason—have signed at as high as $2.40 for two to three years for H100s. If you think about the margin, it costs $1.40 to ..."

"So if I want Hopper, there is some capacity out there. Also, while most of the capacity at an Oracle or a CoreWeave is signed for a long-term deal in terms of Blackwell, anything that's going online t..."

"Or in the case of Anthropic, they don't have to have all the compute directly. Amazon can have the compute and serve Bedrock, or Google can have the compute and serve Vertex, or Microsoft can have the..."

"Right, there's a trade-off there. But at the same time, for a solid four months, everyone was saying to OpenAI, "We're not going to sign deals with you." That sounds crazy, but it was because, "you do..."

"There are not that many incremental buyers of compute yet, because Anthropic hit the capability tier first where their revenue is mooning. That’s interesting. Otherwise you might think having the best..."

"Maybe this is an obvious point. But at least until recently, people had made this huge point about the depreciation cycle of a GPU. The bears, the Michael Burrys or whoever, have said, "Look, people a..."

"But in fact you’re pointing out that maybe the depreciation cycle is even longer than five years. If we're using Hoppers—especially if AI really takes off and in 2030, we’re saying, "We have to get th..."

"I think Anthropic will be able to get to five gigawatts-ish, maybe a little bit more by the end of the year through themselves as well as their product being served through Bedrock, Vertex, or Foundry..."

"But in the chip supply chain, there's a number of different bottlenecks. There's memory, logic wafers from TSMC, and the fabs themselves. Construction of the fabs takes two to three years, versus a da..."

"The bottlenecks, as we've scaled, have shifted based on what the supply chain is currently not able to do. It was CoWoS, power, and data centers, but those were all shorter lead time items. CoWoS is a..."

"There's a spec called overlay, which is: you did all this work, you drew these lines on the wafer, now I want to draw these dots. Let's say I want to draw these dots to connect these lines of metal to..."

"You say we're producing 60 of these machines this year and then 70, 80 over subsequent years. What would happen if ASML just decided to double its CapEx or triple its CapEx? What is preventing them fr..."

"Let's talk about the source specifically. What does the source do? It drops these tin droplets. It hits it three subsequent times with a laser perfectly. The first one hits this tin droplet, it expand..."

"When we think about a lens, it's in a shape, and it focuses the light. This is like a mirror that's also a lens, so it's pretty complicated. Any defect in these super thinly deposited stacks will mess..."

"Each of these things is a wonder and marvel of chemistry, fabrication, mechanical engineering, and optical engineering, because you have to align all these things and make sure they're perfect. All of..."

"If you just think about it, you're talking about two physically moving objects that are the size of a wafer, and it has to be accurate to the level of single-digit nanometers or even smaller because t..."

"Nvidia's had to do a lot to get the entire supply chain to even deliver the capacity they're going to make this year. When you go talk to Anthropic, they're like, 'We're short of TPUs, we're short of ..."

"What does Elon want? He wants 100 gigawatts a year in space by 2028 or 2029. Sam Altman wants 52 gigawatts a year by the end of the decade. Anthropic probably needs the same, and Google needs that. Yo..."

"The movement of data between racks is on the order of hundreds of gigabits a second, 400 gig or 800 gig a second, so roughly 100 gigabytes a second. So you have this huge ladder: on-chip communication..."

"Can I ask you a very naive question? The B200 now has two dies on a single chip, so you can get that bandwidth without having to go through NVLink or InfiniBand. Next year, Rubin Ultra will have four ..."

"The way Nvidia is doing it is CoWoS. Google, Broadcom, MediaTek, and Amazon's Trainium are all doing CoWoS. But actually you can go look back at what Tesla did with Dojo, which they cancelled and rest..."

"By 2030, it's possible that they do. But to date, all of China's 7 nm and 14 nm capacity uses ASML DUV tools. The amount that they can import from ASML is large. But the vast majority of ASML's revenu..."

"They will really scale up production and kick it into high gear. They have more engineers working on it and more desire to throw capital at the problem. So by 2030, will they have fully indigenized DU..."

"It's a great question. It's a bit of a challenge to look into this supply chain especially. We try really hard. In some instances, they're buying stuff from Japanese vendors. If they want a fully indi..."

"Elon says he wants to do it and China is obviously going to do it. TSMC is trying to do that. The memory makers may get to a million wafers a month as well, but not in a single fab. It's mind-boggling..."

"If you have this transformational thing that is commanding tens or hundreds of trillions of dollars of economic growth and token output, but it happens in 2035? What does that imply for the West versu..."

"As you go out to 2035, things are just so radically different. Your error bars get so large it's hard to make an estimate. But at the end of the day, if takeoff or timelines are slow enough, I don't s..."

"At some point, when you can't distill the learnings from these labs into the Chinese models, plus with this compute race that OpenAI, Anthropic, Google, and Meta are all racing on, they end up getting..."

"There's nearly a trillion dollars of CapEx being invested in data centers in America this year, roughly. What's the return on invested capital here? You and I would think the return on invested capita..."

"The memory crunch will continue to get harder, and prices will continue to go up. This affects different parts of the market differently. Are people going to hate AI more and more? Yes, because smartp..."

"It took another three to six months for the memory vendors to start building fabs. Those fabs take two years to build. So we won't have really meaningful fabs to even put these tools in until late 202..."

"I interviewed Elon recently, and his whole plan is that they're going to build this TeraFab and they're going to build the clean rooms. I won't even ask you about the dirty rooms thing, but let's say ..."

"Two, does that even matter if, in two years, your view is that we're not bottlenecked on clean room space, but on the tooling? As with any complex supply chain, it takes time, and constraints shift ov..."

"The thing about Elon is he has a tremendous capability to garner physical resources and really smart people to build things. The way he recruits amazing people is by trying to build the craziest stuff..."

"How surprised would you be if in 2030 there just happened to be some total disruption where we're not using EUV? What if we're using something that has much better effects, is much simpler to produce,..."

"Something that's very simple and easy to scale, I assign a very, very low probability. There are a number of companies working on effectively particle accelerators or synchrotrons that generate light ..."

"Right now, everyone's roadmap goes from the current 6F cell, to a 4F cell, and then finally 3D DRAM by the end of the decade or early next decade. There's still a lot of R&D, manufacturing, and integr..."

"But as we've seen across time, lithography demand as a percentage of wafer cost has trended up. Around the 2014 era, it was 17% of the wafer cost, and it's gone to 30% over the last fifteen years. For..."

"It's a funny dynamic we saw in 2023, 2024, and 2025. People who saw the energy bottleneck before others asymmetrically went to Siemens, Mitsubishi, and of course GE Vernova, and bought up turbine capa..."

"Right now we're at 20 or 30. This is critical IT capacity, by the way, which is an important thing to mention. When I'm talking about these gigawatts, I'm talking about critical IT capacity. Server pl..."

"The other thing is that the grid is scaled so we don't cut off power at peak usage on the hottest day of the summer. But in reality, that's a load spike that is 10-20% higher than the average. If you ..."

"Labor is a big constraint. It's a humongous constraint in this. People have to be trained. Likewise, we'll probably start importing the highest-skilled labor. It makes sense that a really high-skilled..."

"The main factor for reducing the number of people is going to be modularizing things and making them in factories in Asia. Unfortunately for America, places like Korea, Southeast Asia, and in many way..."

"Permitting is a challenge, but America is a big place with 50 states, and things will get done. There are a lot of small jurisdictions where you can just transport in all the workers you need for a te..."

"Other than the economic argument making less sense once you consider that energy is a small fraction of the total cost of ownership of a data center, what are the other reasons you're skeptical?"

"Because we're so capacity-constrained, that compute is theoretically most valuable in the first six months you have it. We're more constrained now than we will be in the future."

"A lot of models fit within one scale-up domain, but many times you split them across multiple scale-up domains."

"Space data centers effectively are not limited by their energy advantage. They are limited by the same contended resource."

"But Elon doesn't win by doing 20% gains. He never wins that way."

"If you can run a very hot chip, it allows a lot of—"

"For Nvidia, previously this meant an H100 server had eight GPUs, and those eight GPUs could talk to each other at terabytes a second."

"If you pre-train a five-trillion-parameter model, how many rollouts do you have to do in RL?"

"But at the end of the day, in isolation, you almost always want to go with a smaller model that gets RL'd faster and gets deployed into research and development earlier."

"I think there are a lot of people making money in many ways."

"is industry. So AI labs, data center companies, hyperscalers, semiconductor companies, the whole supply chain across AI infrastructure. But 40% of our revenue is hedge funds. I'm not going to comment ..."

"They're not going to kick Apple out entirely. What they're going to do is when Apple orders X, they might say, "Hey, we project you only need X minus one, and so that's what we're going to give you, X..."

"What's also interesting is traditionally, Apple has been the first to a process node. 2 nm is actually the first time they're not. Well, that’s besides Huawei. Huawei, back in 2020 and before, was the..."

"Potentially, yeah. Huawei was the first with a 7 nm AI chip as well. They were the first with a 5 nm mobile chip, but they were the first with a 7 nm AI chip. The Huawei Ascend was two months before t..."

"There are a lot of difficulties with the VLMs and VLAs that people are deploying on robots. But to some extent, you don't need to have all the intelligence in the robot. It would be much more efficien..."

"That's an interesting fact about the future that I think people might not appreciate. I think Elon recognizes this, which is why he's going to different places for his chips. He signed this massive de..."

"If we believe that tools are the ultimate bottleneck, how much of Taiwan's place in the AI semiconductor supply chain could we de-risk simply by having a plan to airlift every single process engineer ..."