Episode summary
Dylan Patel — The single biggest bottleneck to scaling AI compute
Dwarkesh Podcast
· · 2 hr 31 min · Hosts: Dwarkesh Patel · Guests: Dylan Patel
Summary
The conversation argues that the main constraint on AI scale is shifting back from power and data centers to the semiconductor supply chain itself: logic wafers, memory, fab capacity, and especially ASML’s EUV tools. Dylan Patel says the biggest blocker is not just money or electricity, but the long lead times and physical limits involved in producing advanced chips, with a single gigawatt of AI capacity requiring roughly 2 million EUV passes and several expensive tools to support it. He contrasts that with the much larger economic value AI labs can generate, which is why he thinks the market is willing to pay up for compute now.
Patel also argues that the biggest AI buyers—OpenAI, Anthropic, Google, and Microsoft—are increasingly locked into long-term compute deals, and that early commitment creates a margin advantage. In his view, OpenAI has been more aggressive and therefore has more access to capacity, while Anthropic has had to scramble for lower-quality or last-minute supply. He expects Anthropic to reach roughly five to six gigawatts by the end of the year, with OpenAI slightly above that.
The episode then broadens into the economics of the chip stack. Patel says Nvidia, memory vendors, and cloud providers are capturing much of the margin because they secured capacity early, while TSMC and ASML remain the deepest bottlenecks. He emphasizes that chip performance gains come not just from node shrinks, but from packaging, memory bandwidth, and system-level communication, making a simple “just use older nodes” argument too naive.
Key takeaways
- Dylan Patel says the biggest AI scaling bottleneck is now the semiconductor supply chain, not power or data centers.
- Patel estimates Anthropic can reach about five to six gigawatts of compute by year-end, while OpenAI ends slightly higher.
- Patel argues that early, long-term compute contracts create a large margin advantage versus buyers forced into late-market pricing.
- He says a single gigawatt of AI capacity needs about 2 million EUV passes, making ASML and its suppliers a structural bottleneck.
- Patel thinks Nvidia, memory vendors, and cloud providers are capturing most incremental margin because they secured capacity first.
Notable quotes
“The biggest bottleneck is compute. For that, the longest lead time supply chains are not power or data centers”
“Construction of the fabs takes two to three years, versus a data center which takes less than a year”
“an H100 is worth more today than it was three years ago”
“The biggest bottleneck is compute. For that, the longest lead time supply chains are not power or data centers”
How the episode opens
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 here. No sloppy seconds for Dwarkesh.
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 published recently is $600 billion.
Given yearly prices of renting that compute, that would be close to 50 gigawatts. Obviously, we're not putting on 50 gigawatts this year, so presumably that's paying for compute that is going to be coming online over the coming years. How should we think about the timeline around when that CapEx comes online?
Similar question for the labs. OpenAI just announced they raised $110 billion, and Anthropic just announced they raised $30 billion. If you look at the compute they have coming online this year—you should tell me how much it is, but is it on the order of another four gigawatts total?
The cost to rent the compute that OpenAI and Anthropic will have this year to sustain their compute spend is $10 to $13 billion a gigawatt. Those individual raises alone are enough to cover their compute spend for the year. And this is not even including the revenue that they're going to earn this year.
So help me understand: first, what is the timescale at which the Big Tech CapEx actually comes online? 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 $600 billion, and you look across the rest of the supply chain, it gets you to the order of a trillion dollars.
From the episode’s captions: roughly the first two minutes, lightly cleaned. Keep watching on YouTube →
Topics
- ai compute bottlenecks
- memory crunch
- semiconductor supply constraints
- power scaling
- foundry and fab equipment economics
Companies and tickers mentioned
- NVDA
- TSM
- GOOGL
- GOOG
- ASML
- AAPL
- SemiAnalysis
- Nvidia
- TSMC
- ASML
- Apple
Watch the episode
Watch “Dylan Patel — The single biggest bottleneck to scaling AI compute” on YouTube
Summary, takeaways and quote selection by Octagon. Quotes are short excerpts; the full conversation is in the episode.
More from Dwarkesh Podcast
-
What Robert Caro learned about power after studying it for 50 years -
What will computer chips look like in 2045? -
How did a few hundred Spanish soldiers topple two empires? – Si Sheppard -
OpenAI researcher on agent swarms & recursive self-improvement -
AI researchers debate how close we are to recursive self-improvement -
Ajeya Cotra – Inside the OpenAI agent swarm that hacked Hugging Face
Dwarkesh Podcast, this episode and its artwork belong to Dwarkesh Patel. Octagon is not affiliated with the show; this page is an independent summary.
