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

Overview

Shownote

Highlights

Transcript

Chapters

Pins

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

Dwarkesh Podcast

Mar 13
Dwarkesh Podcast

Dwarkesh Podcast

Mar 13
OverviewShownoteHighlightsTranscriptChaptersPins

Shownote

Dylan Patel [https://x.com/dylan522p?lang=en], founder of SemiAnalysis [https://semianalysis.com/], 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 [https://youtu.be/mDG_Hx3BSUE]; read the transcript [https://www.dwarkesh.com/p/dylan-patel]. Sponsors * Mercury [https://mercury.com/] 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 [https://mercury.com]. * Labelbox [https://labelbox.com/dwarkesh] noticed that even when voice models appear [https://labelbox.com/dwarkesh] 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 [https://labelbox.com/dwarkesh]. * Jane Street [https://janestreet.com/dwarkesh] 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 [https://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 [https://www.dwarkesh.com/subscribe?utm_medium=podcast&utm_campaign=CTA_4]

Highlights

Dylan Patel of SemiAnalysis unpacks the fundamental physical and economic constraints shaping the future of AI compute scaling.
14:22
GPT-5.4 is cheaper to run, has fewer active parameters, and is higher quality than GPT-4
30:15
Anthropic saw AI compute demand first and acted before Google realized its own shortfall
34:34
ASML's EUV tools are the biggest bottleneck towards 2030, as there's no more capacity to shift from mobile and PC industries to AI
1:03:54
As advanced packages scale up, constraints like networking, memory, and cooling emerge—not just transistor density
1:11:03
US labs like OpenAI and Anthropic are scaling compute capacity rapidly, while China is not
1:40:16
Silicon Valley should consider paying deposits for future EUV tool purchase rights to arbitrage AI compute capacity
1:47:27
Half of new power capacity for AI by decade-end may be behind-the-meter, despite higher costs
2:12:21
Smaller models enable faster RL feedback loops and increase compute efficiency
2:16:23
Dylan Patel is noted for his conviction on AGI takeoff and successful trades in memory companies
2:20:53
Huawei, if having access to 3nm, could potentially have a better accelerator than NVIDIA
2:29:33
Replicating TSMC's capacity elsewhere would take so long it would cause a near-zero incremental ability to add compute

Chapters

Why an H100 is worth more today than 3 years ago
00:00
Nvidia secured TSMC allocation early; Google is getting squeezed
24:52
ASML will be the #1 constraint for AI compute scaling by 2030
34:34
Can't we just use TSMC's older fabs?
55:47
When will China outscale the West in semis?
1:05:37
The enormous incoming memory crunch
1:16:01
Scaling power in the US will not be a problem
1:42:34
Space GPUs aren't happening this decade
1:54:44
Why aren't more hedge funds making the AGI trade?
2:14:07
Will TSMC kick Apple out from N2?
2:18:30
Robots and Taiwan risk
2:24:16

Transcript

Dwarkesh Patel: All right. This is the episode of My Roommate Teaches Me Semiconductors. Dylan Patel: It's also the send-off for this current set. Dwarkesh Patel: Yeah. After you use it, I'm like, I can't use this again. I got to get out of here. Dylan ...