An interview with SemiAnalysis CEO Dylan Patel on logic, memory, and power bottlenecks in scaling AI compute, Nvidia securing TSMC N3 allocation early, and more
Plus, why an H100 is worth more today than 3 years ago — Dylan Patel, founder of SemiAnalysis, provides a deep dive …
Dwarkesh PodcastDwarkesh Patel
Context & Ripple Effects
The discussion places AI-compute scaling in a supply-chain context: more capable logic alone does not remove memory and power constraints. It also highlights how early access to an advanced TSMC node can matter alongside GPU design.
Related coverage later focused on Nvidia’s supply-chain moat and competition from ASICs, while coverage of SemiAnalysis points to growing demand for specialist research on AI infrastructure constraints.
First-order effects
Nvidia’s early TSMC N3 allocation gives it a potentially stronger planning position for products tied to that process capacity, while rivals face a tighter path to matching leading-edge supply.
The H100’s reported increase in value underscores that deployed AI hardware is being priced around scarce usable compute, not simply its age as a product.
Second-order effects
AI system builders must treat memory availability and power delivery as co-equal constraints with accelerator procurement, increasing the value of designs that balance all three.
Competitors have greater incentive to diversify through custom ASICs, alternative accelerator architectures, or other manufacturing paths rather than rely solely on catching Nvidia at the leading edge.
Third-order effects
If logic, memory, and power remain jointly binding, AI competition will increasingly be decided by integrated infrastructure execution—chips, manufacturing capacity, memory, and energy—not peak-chip performance alone.
The pattern supports a more heterogeneous compute market, though Nvidia’s supply advantage will depend on whether competitors can secure comparable capacity and deliver viable alternatives.
The trend: AI infrastructure is shifting from a GPU-centric procurement race toward competition over the full stack of leading-edge silicon, memory, and power capacity.
.@dylan522p gives a deep dive on 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. 0:00:00 - Why an H100 [v…
Narrative violation from Dylan on Dwarkesh: H100s are worth *more* today than they were 3 years ago. There's a sentiment that data center buildouts are priced into the risk of rapidly depreciating GPUs. But the models want to learn. Token prices are falling so fast that you can […
.@dylan522p lays out how we know the hard upper bound on how much compute can be produced annually by 2030: around 200 GW/year. That's a crazy number (there's about 20 GW of AI deployed in the world right now), but it's nowhere near enough to satisfy Sam/Elon/Dario/Demis's [video…
The AI supply chain has the craziest value cascade of any industry in the world. thinks that over the next five years, the biggest bottleneck to deploying AI will be EUV machines. ASML sells EUV machines for $300-400 million. You need about three and a half machines, so $1.2 [vid…
This podcast with @dylan522p is a terrific rebuttal to the Citrini doomer scenario by playing through the real world constraints of a fast-ish takeoff (I know it wasn't intended as such). The constraints to producing enough AI tokens to be disruptive to society will slow it down
Every AI bubble argument assumes the compute requirements keep going up forever. The actual trend is going the other way. GPT-4 required cutting-edge H100s to run at scale. Newer models at the same or better quality level run on hardware that is two to three generations older.
If you've been wanting to understand AI inference and hardware economics... this is the best single place for a current take. It's long, but highly recommended for those interested!
Excellent listen. Key takeaways: 1) $ASML caps at 200GW by 2030 2) Memory eating 30% of Big 7 CapEx 3) H100s appreciate vs. depreciate 4) Neoclouds (and their agents such as $GLXY) control bottleneck 5) Early contracts ($CRWV 98% locked) print vs 50% spot markup @dylan522p 🐐
This matches my current world model. An H100 GPU is worth more today than 3 years ago, not less. People viewing GPUs as rapidly depreciating assets are missing that an older GPUs can and will do economically valuable work. The total economic value of work that GPU can do for you
The logic squeeze that is happening (Dylan thinks it's going to get worse) ensures that Intel after its N2P orders are filled for Nova Lake will be forced back to internal manufacturing The idea of Intel flexing ~20% to TSMC doesn't work in this world. That's good news
this completely fucking breaks the AI Bubble narrative. a 3 year-old gpu is MORE valuable today because it serves higher-quality ai tokens FOR CHEAPER. translation: gpt 5.4 runs BETTER on an OLD GPU than gpt-fucking-FOUR read that again. a newer, better model runs more
if you're interested in the race to agi you have to watch this. much more in depth on the super cluster build out. how much compute / gw's can you get online quickly. sam (the dealmaker) altman's conviction and acceleration is paying off again. whilst anthropics sbf
The people I find most insightful/interesting on the economics of AI are non-economists. On hardware/macro topics, it is very hard to beat Dwarkesh and Dylan - I'm always enlightened by listening to them and humbled by their breadth and depth of knowledge. Definitely going to