EUV machines are the most complicated tools humans make. Their supply chain has over 10,000 individual suppliers, and any one of them not scaling fast enough can bottleneck the entire AI industry. An EUV tool fires lasers at a tiny tin droplet three times in precise sequence, blasting it... hard enough to emit EUV light. That light bounces off 18 multilayer mirrors onto the wafer. Meanwhile, the two platforms inside the machine - one holding the stencil, one holding the chip - are flying back and forth at 9Gs in opposite directions. The successive passes have to land on top of each other to within 3 nanometers. If any part of this is off, yield goes to zero. Take just one component. The mirrors are mostly supplied by Carl Zeiss, who have probably fewer than a thousand people working on them. In turn, Carl Zeiss rely on machines from Switzerland to deposit each of the layers, and use a coating process co-developed with a different German company. None of these companies have woken up. They’re gradually increasing production, but nowhere near the levels necessary for what the labs want by the end of the decade. Dylan Patel predicts production can't scale beyond about 100 EUV machines per year by 2030, no matter how much money gets thrown at the problem. In the medium term this is the key bottleneck on scaling.show more

People underestimate how much this comes down to a few hundred specialized engineers in Europe rather than just raw capital. You can't just "software update" your way out of a physical hardware bottleneck when the tolerances are that tight. It feels like the market is pricing in infinite compute while ignoring the fact that we're essentially reliant on a handful of master craftsmen.

i've been waiting for this day my whole life 🥲

True, but nothing new:

I would watch @dylan522p and @dwarkesh_sp try and convince a Swiss lens manufacturer to increase production live in Switzerland.

A LEAP-1A has about 19,000 individual parts. The hot section turbine blades operate at temperatures above their own melting point. They survive only because of internal cooling channels cast into single-crystal nickel superalloys using ceramic cores that dissolve after casting. The metallurgical tolerance on grain boundaries is effectively zero; one stray crystal and the blade fails. These blades are coated in thermal barrier ceramics deposited in electron-beam PVD chambers that cost tens of millions each. The supply chain is dominated by three casting houses globally (PCC, Howmet, ITP). Safran and GE have been shipping ~2,000 LEAP engines per year and targeting higher. Nobody writes Twitter threads about how jet engine production caps at 100 units because the ceramic core supplier in New Hampshire only has 400 people. My ai is wicked smaht

EUV is just one way to project a pattern on a wafer. Expect Elon to do it differently. Everyone in this thread seems stuck in a box.

@dylan522p you can learn more here if you want (not @dwarkesh_sp of course but people coming in this thread ;) )

Cymer no longer produces the EUV light sources. They have been made by Trumpf for a few years now

Fewer than a thousand people at Carl Zeiss are a bottleneck for the entire AI industry. Does that feel like a solvable recruiting problem or something structurally different?

Zeiss is the crazy rabbit hole

3/ missing the point. ASML knows the market and knows their own suppliers and their needs. It's just that not everyone is a bull-pilled starry-eyed believer ready to bet the farm that Christmas will happen 10 times per year in the next 10 years. These are reasonably cautious

1/ Yes, but in the real world, increasing capacity at all these suppliers would require enormous CAPEX that nobody can reasonably finance. And these players have to take into account the long term calculus (many have been burned before). You don't increase capacity like that just

@DanielleFong Na. We are golden here. We will run out of copper fast enough that the tire 2 suppliers of ASML should make sure to not overspent on the hype. Energy and energy distribution, let alone the whole capex and return on capital tropic, will be the bottlenecks.

How many droplets of tin per second you may ask? 50.000 And they are working on 100.000 tin droplets per second.

amazing how much of @dylan522p's alpha is "in order to make x chip we it needs y passes on an EUV machine and we can only make z of them". Genius is the ability to find alpha in plain sight by contemplating the obvious

And how do you make sure the positioning is within 3nm? Turns out it is just one company manufacturing interferometers in Santa Clara, CA that makes it possible. I now feel really bad we didn’t show that to the @SemiAnalysis_ team when we had a meeting on a separate topic

@dwarkesh_sp High-NA EUV is the peak of engineering, but lithography is only half the battle. Even at 2nm, the pJ/bit wall remains the ultimate ceiling. If we don’t innovate in 3D packaging/interconnects, we’re just making efficient transistors for a leaky energy bucket. 🏛️⚡

6/ In other words: Yes, adjacent "good old industries" (power generation, optics for ASML, etc.) are a new terrifying bottleneck for this crazy DC buildup. But these industries don't operate on the tsunami timeline and that's just how it is. At some point, the tsunami hits walls.

Fascinating breakdown. But I'd wager markets innovate around bottlenecks faster than expected when profits align. Constraints often accelerate solutions.

now that's what I call feeds and speeds gd

Delete like 9,000 of those suppliers. And then delete another 900. Radically simplify this machine like your lives depend on it. Because they do.

Chinas new approach does NOT need these machines. Please stop talking gibberish. You are embarassing yourself, do the research.

I always thought the US superpower was quickly allocating capital, and kinda shocked we don't have industrialists just rabidly copying all these firms, possibly by adopting the Chinese model and just trying to steal or buy outright all their engineers and processes

Could you have a chat with the TeraFab authors?

10,000 suppliers and most companies can barely map past tier 1. The bottleneck isn't just scaling, it's that nobody has visibility into who supplies the supplier. That's where risk actually lives.

time to think how to make suppliers efficient or reduce the number of suppliers needed by EUVs

It’s astonishing, few things humans have built operate at this level of precision, where physics, engineering, and coordination all meet at the very edge of what’s possible.

EUV is the bottleneck but not for the reason most think. More suppliers = more failure points, not more resilience. The AI race depends on 10,000 strangers all doing their job right.

Really? ASML has 10,000 suppliers for the EUV Lithography equipment?

The EUV lightsource is definitely advanced science and tech But how difficult is it to hit a droplet with lasers 3 times? Haha should i Vibecode this?

Guaranteed the most complicated piece of equipment humans ever built. No one will copy that. ASML’s biggest risk is a competitor developing entirely new tech to produce chips, instead of imaging silicon.

2/ for the next 3 years. And money to finance this expansion isn't free either. Yes, everyone is very "excited" these days by the current tsunami going through the industry. By in my humble opinion, stating that "None of these companies have woken up" is factually wrong and

When can we expect this kind of machine can be used by everyone? I dont even know any kind of euv machine

Dwarkesh Podcast Episode: "Dylan Patel — Deep Dive on the 3 Big Bottlenecks to…" Mar 13, 2026 · Clip Starts at 47:25 Full episode ↓