The publicly traded Carl Zeiss Meditec AG (AFX DE) is a medical technology subsidiary of the privately held Carl Zeiss AG and has no involvement with ASML. The ASML relationship sits with Carl Zeiss SMT GmbH, a separate subsidiary of Carl Zeiss AG that supplies precision optics for EUV lithography machines — and in which ASML holds a 24.9% minority stake.
Dylan isn’t just saying “AI demand is huge.” He’s making a much more interesting argument about where the bottlenecks move over time.
First, better models are making even older GPUs more useful economically, not less. That makes those early long-term compute deals look way more important in hindsight.
Then the bottleneck story shifts. It’s not just power or datacenter buildout forever. If things keep scaling, the constraint moves upstream into manufacturing, maybe even EUV.
Once you look at it that way, a lot of the later parts of the conversation start to connect: HBM squeezing consumer electronics, Apple losing leverage at TSMC, space datacenters not being crazy but just too early, and Taiwan still being the real tail risk underneath all of this.
I’m not saying every specific number or timeline has to be right. But I do think the overall frame is strong. A lot of people still talk about AI like the question is just “how many GPUs do we need?” Dylan’s version is more like: “which physical layer breaks next, and who got there first?”
That feels like a much better way to think about the next few years.
The GPU depreciation bear case assumes that new chips will commoditize old ones. However, when demand grows faster than upstream bottlenecks can scale, the opposite occurs: old chips appreciate in value.
Michael Burry was modeling commodity economics. This is a bottlenecked infrastructure market where lithography and memory are the constraints, not production.
Aand guess what ASML is desperate for "Manufacturing Engineers"... Say it Dylan, not hard buddy. You think negative cash flow is bad for Oracle 😂 Dylan like Jekyll and Hyde, or the Hulk. Angry man on inside, calm on out. Only comes out now and then 👌🏽
Dylan is a GOAT. And if you ever get to meet him in-person, he's a really down-to-earth guy. You'd never think he's going to run a $B co by simply indexing on the shit he's obsessed with. Go figure.
They havent really touched on the elephant in the room - LLMs are so inherently error prone that they are only useful for specific boundaried tasks like coding. So, agents will not foreseeably be able to perform end to end projects like humans can. Yes they can improve productivity when used properly but the entire discussion above assumes almost human like capabilities which they do not have currently, and will not have in the forseeable future due to the inherent ungroundedness of LLMs. I just dont think the market is there. I think Yan Lecun has hit the nail on the head.
One point in the podcast that was pretty new to me was the idea that future robots will have a hybrid local + cloud intelligence. The local model in the robot is super fast and handles dexterity and movement, the cloud model is slow, powerful and handles higher order thinking. I guess a few science fiction worlds have suggested it (the Borg, the matrix), but I haven't heard it discussed in reality. I thought the slowness of current models would make robots impractical, but that is an elegant solution.
Isn't it a huge problem that AI data centers are going to make electricity and consumer electronics more expensive? What if the political backlash is so big that government starting banning or severely restricting AI data centers? It's already starting to happen at the local level.
I think the major players need to think really hard about this problem. They're already spending hundreds of billions, maybe spend a few billion more to expand fabs, memory and energy sources.
Then again, the incentives might be screwed up. Highly speculative, but Amazon, Google and Microsoft might be okay with governments curtailing new data centers since that will increase the value of existing compute and shut down potential new players.
The publicly traded Carl Zeiss Meditec AG (AFX DE) is a medical technology subsidiary of the privately held Carl Zeiss AG and has no involvement with ASML. The ASML relationship sits with Carl Zeiss SMT GmbH, a separate subsidiary of Carl Zeiss AG that supplies precision optics for EUV lithography machines — and in which ASML holds a 24.9% minority stake.
Dylan isn’t just saying “AI demand is huge.” He’s making a much more interesting argument about where the bottlenecks move over time.
First, better models are making even older GPUs more useful economically, not less. That makes those early long-term compute deals look way more important in hindsight.
Then the bottleneck story shifts. It’s not just power or datacenter buildout forever. If things keep scaling, the constraint moves upstream into manufacturing, maybe even EUV.
Once you look at it that way, a lot of the later parts of the conversation start to connect: HBM squeezing consumer electronics, Apple losing leverage at TSMC, space datacenters not being crazy but just too early, and Taiwan still being the real tail risk underneath all of this.
I’m not saying every specific number or timeline has to be right. But I do think the overall frame is strong. A lot of people still talk about AI like the question is just “how many GPUs do we need?” Dylan’s version is more like: “which physical layer breaks next, and who got there first?”
That feels like a much better way to think about the next few years.
The GPU depreciation bear case assumes that new chips will commoditize old ones. However, when demand grows faster than upstream bottlenecks can scale, the opposite occurs: old chips appreciate in value.
Michael Burry was modeling commodity economics. This is a bottlenecked infrastructure market where lithography and memory are the constraints, not production.
Aand guess what ASML is desperate for "Manufacturing Engineers"... Say it Dylan, not hard buddy. You think negative cash flow is bad for Oracle 😂 Dylan like Jekyll and Hyde, or the Hulk. Angry man on inside, calm on out. Only comes out now and then 👌🏽
2.5 hours of Dylan Patel on Dwarkesh is my version of a blockbuster movie. Dylan is so likeable and knowledgeable. Awesome guest (again).
Dylan is a GOAT. And if you ever get to meet him in-person, he's a really down-to-earth guy. You'd never think he's going to run a $B co by simply indexing on the shit he's obsessed with. Go figure.
I have met Dylan and can confirm this take
Query: If some of these players go bankrupt, will the survivors be able to purchase and use heir data centers?
They havent really touched on the elephant in the room - LLMs are so inherently error prone that they are only useful for specific boundaried tasks like coding. So, agents will not foreseeably be able to perform end to end projects like humans can. Yes they can improve productivity when used properly but the entire discussion above assumes almost human like capabilities which they do not have currently, and will not have in the forseeable future due to the inherent ungroundedness of LLMs. I just dont think the market is there. I think Yan Lecun has hit the nail on the head.
I greatly enjoyed this one, so many thought-provoking sections and interesting models for how to think about things. thanks for doing it, guys 💚 🥃
One point in the podcast that was pretty new to me was the idea that future robots will have a hybrid local + cloud intelligence. The local model in the robot is super fast and handles dexterity and movement, the cloud model is slow, powerful and handles higher order thinking. I guess a few science fiction worlds have suggested it (the Borg, the matrix), but I haven't heard it discussed in reality. I thought the slowness of current models would make robots impractical, but that is an elegant solution.
Isn't it a huge problem that AI data centers are going to make electricity and consumer electronics more expensive? What if the political backlash is so big that government starting banning or severely restricting AI data centers? It's already starting to happen at the local level.
I think the major players need to think really hard about this problem. They're already spending hundreds of billions, maybe spend a few billion more to expand fabs, memory and energy sources.
Then again, the incentives might be screwed up. Highly speculative, but Amazon, Google and Microsoft might be okay with governments curtailing new data centers since that will increase the value of existing compute and shut down potential new players.