$400 million. That’s the sticker price on ASML’s High NA lithography systems, the machines that etch the smallest transistors humanity knows how to make. And there’s a category of chip design they can’t print in a single exposure.
Not “can’t print efficiently.” Can’t print in one shot. The largest designs have to be split across multiple exposures and stitched together, which adds steps, adds cost, and adds places for things to go wrong. The fix ASML has planned isn’t expected to land until somewhere in 2031 to 2033.
I review AI tools for a living, so I spend most of my time three or four abstraction layers above this. But this one is worth understanding, because it’s the same failure pattern I see in software toolkits every single week.
The spec sheet says yes, the workflow says no
Here’s the pattern I keep running into when I test AI tooling: a product nails the headline metric and quietly fails on the shape of real work.
A coding agent that writes beautiful functions but can’t hold a whole repository in context. A document tool that handles any file you throw at it, until the file is 400 pages. A vector database that’s fast at a million records and falls over at fifty million. The capability is real. The envelope is smaller than the marketing.
ASML’s situation is that same story with a nine-figure price tag. The resolution is there. The transistors get smaller. But the printable area per exposure doesn’t cover the biggest designs, and the workaround is a multi-step process that costs real money.
What makes this sting is the timing. AI accelerators are exactly the chips that want to be enormous. Bigger die, more compute per package, fewer bottlenecks between parts. The demand curve is pushing straight at the one dimension the machine doesn’t scale.
Demand this strong and growth is still a question mark
The financials make the tension obvious. 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. Strong AI demand drove those numbers.
Then the company told the market it “cannot confirm” growth for 2026. Shares dropped 11% on the cautious outlook. Roughly $30 billion in market value went away.
That is a strange combination. Record-adjacent results, ferocious end-market demand, and a company unwilling to promise the next year goes up. Add TSMC’s Deputy Co-COO Kevin Zhang telling Bloomberg that TSMC has no plans to buy the newer generation of high numerical aperture machines, and the picture gets clearer. The most important customer in the industry looked at the tool and said not yet.
Why buyers say no to better tools
This is the part I find genuinely useful for anyone evaluating tools at any scale. A customer declining the more capable, more expensive option is rarely about the capability itself. It’s about whether the total workflow gets better.
If the new tool is sharper but needs extra steps to cover your actual use case, and the older tool already handles your use case with a process you’ve tuned for years, the sharper tool loses. Not on specs. On arithmetic.
I run into this constantly with AI toolkits. The newer model is smarter and the newer framework is cleaner, but migrating costs three weeks and the current setup already ships. Teams stay put, and they’re usually right to. Better-in-isolation is not the same as better-in-context.
The 2031 problem
The timeline is what I’d stare at longest. A solution planned for 2031 to 2033 is a solution roughly six to eight years out. In semiconductor manufacturing, that’s a reasonable roadmap. In AI, it’s an eternity. Nobody credibly knows what accelerator designs will want by then.
That gap between physical tooling timelines and software timelines explains a lot of the weirdness in this market. Software teams iterate in weeks. The machines that make the silicon underneath iterate across decades of R&D. When AI demand accelerates, the software layer absorbs it fast and the physical layer absorbs it slowly, and pressure builds at the seam.
Right now the seam is the size of a single exposure.
What I’d actually take from this
Two things, and they apply whether you’re buying lithography systems or picking an agent framework.
- Test the tool at your maximum, not your average. The limit is where the cost lives, and the limit is almost never in the marketing material.
- When a serious buyer publicly passes on the most advanced option available, that’s real information. Find out what they know about their own workflow that the spec sheet doesn’t cover.
ASML still makes machines nobody else on Earth can make. The limitation doesn’t change that. It just proves that the hardest engineering problem and the customer’s actual problem aren’t always the same problem, and that a solid roadmap six years out doesn’t help anyone shipping this quarter.
🕒 Published: