ASML CFO Roger Dassen told Reuters something that should have been a victory lap and instead reads like a warning label: the company’s existing EUV machines, at roughly $200 million each, are effectively sold out through 2027. Sold out. Years of production, spoken for, before most AI startups have figured out what their inference bill will look like next quarter.
My reaction, as someone who spends most days reviewing AI tools that promise the moon and ship a spreadsheet: that is not a flex. That is a bottleneck wearing a bow tie.
The $400 million machine that needs a second pass
Here is the part that gets glossed over in the breathless coverage of ASML’s High NA EUV systems, the ones priced around $400 million. They cannot print the largest chip designs in a single exposure. The biggest, most ambitious AI accelerator layouts do not fit inside one shot. So you stitch. You do multiple exposures and align them, which means extra steps, extra cost, extra chances to ruin an expensive wafer.
A fix is planned. It arrives somewhere in the 2031 to 2033 window.
Read that timeline against how fast AI hardware roadmaps churn. Chip designers are not asking for slightly larger dies out of vanity. Big AI silicon keeps getting physically bigger because that is one of the few reliable ways to get more compute and more memory bandwidth into one package. The demand curve is pointed straight at the exact limitation that has a seven-year repair schedule.
Why this is worse than a normal supply problem
Most supply crunches respond to money. Throw enough capital at a factory and eventually more units come out the other end. ASML’s near-term output does not work that way. It is capped by physics and clean-room space. You cannot venture-fund your way past optics, and you cannot summon a qualified clean room in a quarter.
The numbers ASML has laid out for full-year 2026 are instructive: roughly 65 Low NA EUV systems and about 130 DUV immersion systems, alongside a planned 30% capacity increase. Those are real, meaningful figures. They are also finite in a way that AI compute demand simply is not. Every hyperscaler, every sovereign AI project, every startup that decided it needs custom silicon is drawing from the same small pool of machines.
What Samsung and TSMC signing on actually tells you
On 8 September 2026, ASML announced that Samsung and TSMC had committed to using High NA EUV systems in high-volume manufacturing. Intel has given the machines a vote of confidence too.
The optimistic read: the technology is ready enough that the most demanding manufacturers on earth are betting production lines on it.
The read I trust more: the largest players are locking in access to a constrained resource, and they are doing it early because they have done the arithmetic on what happens if they do not. When the three biggest names in advanced manufacturing all move at once on hardware that is sold out for years, that is not a technology endorsement. That is a scramble for shelf space.
The part that matters if you buy AI tools instead of lithography systems
You are not purchasing a $400 million machine. So why should you care?
- Compute prices have a floor set by physics, not by competition. When the machines that make the chips are output-capped and the biggest chips need multiple expensive exposures, that cost travels downstream. It lands in your inference bill.
- Multi-step patterning is a tax on ambition. The most capable AI chips are precisely the ones that do not fit in a single exposure. The frontier is the expensive part, by definition.
- Any AI product whose margins assume compute gets cheap fast is making a bet on a fix scheduled for the early 2030s. That is a long time to be wrong.
I review a lot of agent platforms and AI tools that quietly assume token costs collapse on a predictable curve. Their pricing pages depend on it. Their unit economics depend on it. Almost none of them have looked at the manufacturing constraint sitting underneath the whole stack, where the fix has a delivery date most of their startups will not survive to see.
A useful kind of honesty
What I actually appreciate about ASML in all this is the lack of theater. The limitation is acknowledged. The timeline for addressing it is stated, unflatteringly far out. Nobody is pretending the constraint is a feature.
Compare that to the average AI product launch, where a known limitation becomes an “opinionated design decision” and the roadmap fix is always next quarter.
The company selling the most valuable machines in the industry will tell you plainly what those machines cannot do. That is a higher standard of candor than most software vendors manage about a login flow. The industry building on top of that hardware could stand to borrow some of it.
🕒 Published: