Dylan Patel — Deep dive on the 3 big bottlenecks to scaling AI compute

March 13, 2026

AI Summary

5 min read

The Semiconductor Bottleneck That Will Define AI's Next Decade

In 2023, a single gigawatt of AI compute required roughly 3.5 EUV lithography machines from ASML. Those machines cost about $1.2 billion combined. The data center built around that gigawatt costs roughly $50 billion. The most important bottleneck in the entire AI supply chain is a $1.2 billion piece of equipment that ASML can only produce about 70 of this year, and maybe 100 by 2030.

The Three Shifting Bottlenecks

The constraint on scaling AI compute has moved up the supply chain over time. Two years ago, the bottleneck was CoWoS advanced packaging. Last year, it was power and data center construction. This year and going forward, the bottleneck shifts to the semiconductor supply chain itself — specifically logic wafers, memory, and the tools that make them.

The reason is that the easy slack has been used up. Previously, AI could cannibalize capacity from mobile and PC chips. Now NVIDIA is the largest customer at both TSMC and SK Hynix. There is no more capacity to shift from consumer electronics to AI. Every new gigawatt of AI compute must come from new fabrication capacity, and that takes years to build.

The EUV Constraint

Continue reading the full summary in the app — free to try.

Read Full Summary →

Free • No credit card required

Listen to Audio Summary Open in App

Never miss an episode of Dwarkesh Podcast

Get every new episode summarized in your inbox — free, ~5 minutes to read.

No spam. Unsubscribe anytime.

What you'll learn

1 (00:00) **Episode Introduction & The Capex Puzzle** - Dylan Patel (SemiAnalysis CEO) joins; Dwarkesh frames the core puzzle: Big Tech's combined $600B capex and lab fundraises ($110B OpenAI, $30B Anthropic) seem to far exceed the cost of compute coming online this year.
2 (01:41) **The Setup Capex Explanation** - Dylan explains that a large portion of the massive capex is "setup capex" for future years (turbine deposits, data center construction, power agreements), not just compute coming online this year.
3 (02:54) **Anthropic's Compute Pinch & The Cost of Being Conservative** - Anthropic's revenue is exploding, requiring ~4 gigawatts of new inference capacity just to grow, but they were conservative on compute contracts.
4 (11:24) **GPU Depreciation & The Value of Old Chips** - Contrary to bearish takes, the value of an H100 is *rising* because newer, more valuable models (e.g., GPT-5.4) can run on it, generating more revenue per token.
5 (18:52) **The Alchian-Allen Effect & Model Margins** - A fixed cost increase in compute (e.g., GPU price rise) makes the relative price of the best model cheaper, pushing users toward higher-quality models.
6 (24:54) **NVIDIA's Lock on Logic & Memory** - How NVIDIA secured the majority of TSMC's N3 capacity and memory supply by being "AGI-pilled" and signing non-cancellable contracts earlier than competitors.
7 (34:38) **The Ultimate Bottleneck: The Semiconductor Supply Chain** - The bottleneck shifts from power/data centers to chips, and ultimately to ASML's EUV lithography tools.

+ Full timestamped outline available in the app

Show Notes

Dylan Patel, founder of SemiAnalysis, provides a deep dive into the 3 big bottlenecks to scaling AI compute: logic, memory, and power.

And walks through the economics of labs, hyperscalers, foundries, and fab equipment manufacturers.

Learned a ton about every single level of the stack. Enjoy!

Watch on YouTube; read the transcript.

Sponsors

* Mercury has already saved me a bunch of time this tax season. Last year, I used Mercury to request W-9s from all the contractors I worked with. Then, when it came time to issue 1099s this year, I literally just clicked a button and Mercury sent them out. Learn more at mercury.com.

* Labelbox noticed that even when voice models appear to take interruptions in stride, their performance degrades. To figure out why, they built a new evaluation pipeline called EchoChain. EchoChain diagnoses voice models’ specific failure modes, letting you understand what your model needs to truly handle interruptions. Check it out at labelbox.com/dwarkesh.

* Jane Street is basically a research lab with a trading desk attached – and their infrastructure backs this up. They’ve got tens of thousands of GPUs, hundreds of thousands of CPU cores, and exabytes of storage. This is what it takes to find subtle signals hidden deep within noisy market data. If this sounds interesting, you can explore open positions at janestreet.com/dwarkesh.

Timestamps

(00:00:00) – Why an H100 is worth more today than 3 years ago

(00:24:52) – Nvidia secured TSMC allocation early; Google is getting squeezed

(00:34:34) – ASML will be the #1 constraint for AI compute scaling by 2030

(00:55:47) – Can't we just use TSMC's older fabs?

(01:05:37) – When will China outscale the West in semis?

(01:16:01) – The enormous incoming memory crunch

(01:42:34) – Scaling power in the US will not be a problem

(01:54:44) – Space GPUs aren't happening this decade

(02:14:07) – Why aren't more hedge funds making the AGI trade?

(02:18:30) – Will TSMC kick Apple out from N2?

(02:24:16) – Robots and Taiwan risk

Get full access to Dwarkesh Podcast at www.dwarkesh.com/subscribe