ASML built the most expensive machine in the chip industry, and the biggest AI chips are too big to fit on it.
That’s the whole story in one sentence, and if you work anywhere near AI agents, model training, or inference costs, it’s worth understanding why. Because the machine in question sits at the very bottom of the stack that everything you use runs on.
What ASML actually sells
If you’ve never had reason to care about lithography equipment, a quick orientation. ASML is a Dutch company that makes the machines that print circuit patterns onto silicon wafers. Not the chips. The machines that make the chips. Every advanced processor in every data center training every model traces back to ASML gear at some point in its life.
Its newest, priciest product line is called High NA, short for high numerical aperture. These systems cost more than $400 million each and use shorter light beams to etch transistors than the previous generation. Shorter light means finer detail, which means more transistors in the same space. That’s the entire game in chipmaking, and it has been for decades.
Business has been good. ASML posted full-year 2025 net sales of $39.16 billion and net income of $11.5 billion, with fourth-quarter revenue of $11.62 billion. AI demand drove a lot of that.
Now the awkward part
The $400 million machine cannot print the largest chip designs in a single exposure. The area it can pattern at once is smaller than the biggest chips people want to build. To make a large chip on this equipment, you print it in pieces and stitch them together, which means extra steps, extra time, and extra cost.
Think of it as a very expensive printer with a smaller paper tray than the poster you’re trying to print. You can tape two sheets together. It works. It’s just slower, fussier, and more likely to go wrong.
ASML has a fix planned. It isn’t expected to arrive until somewhere in the 2031 to 2033 window. That’s a long wait in an industry where product cycles are measured in months.
Why AI makes this worse instead of better
You’d expect an AI boom to be the best possible news for the company selling the most advanced chipmaking tools. Mostly it is. But AI accelerators trend toward being physically enormous. The chips that train large models and serve them at scale are among the largest commercial silicon designs anyone produces. Demand is pushing hardest in exactly the direction the machine handles worst.
So the boom that should be pure tailwind comes with a catch. The more the market wants very large AI chips, the more that single-exposure limit stings.
Customers have noticed. TSMC’s Deputy Co-Chief Operating Officer, Kevin Zhang, told Bloomberg that the company has no plans to purchase the newer generation of ASML’s high numerical aperture machines. When your most important customer publicly passes on your flagship product, that’s a signal.
Markets reacted accordingly. ASML shares dropped 11% on a cautious 2026 growth outlook. The company said it “cannot confirm” growth in 2026 despite beating on current earnings, and roughly $30 billion in market value came off the table.
What this means if you build with AI agents
I write for people who use AI tools rather than manufacture them, so let me connect the dots.
- Compute costs have a physical floor. When people talk about inference getting cheaper forever, they’re implicitly assuming chip manufacturing keeps improving on schedule. Constraints like this one are where that assumption gets tested.
- Multi-step patterning means higher costs per chip. Those costs flow through chipmakers, then cloud providers, then eventually your API bill. Not overnight, and not in a way you can trace line by line, but the direction is real.
- A 2031 to 2033 fix is not a near-term answer. Whatever gets built between now and then gets built around the limitation, not past it.
- Hardware is not a solved background detail. The AI agent space talks endlessly about models and frameworks. The layer underneath has its own physics, its own delays, and its own bad quarters.
The honest read
None of this suggests ASML is in trouble. A company earning $11.5 billion in net income with no real competitor at the top end is not fragile. What it does show is that the AI buildout runs into hard physical limits that money alone cannot rush.
The industry’s usual answer to a problem is to spend more. Here the answer is a machine that doesn’t exist yet, arriving sometime in the early 2030s. Until then, chipmakers will keep taping sheets of paper together and paying for the privilege.
Worth remembering the next time someone tells you AI progress is limited only by imagination. Sometimes it’s limited by how much silicon fits under a lens.
🕒 Published: