TSMC’s Deputy Co-Chief Operating Officer Kevin Zhang told Bloomberg his company has no plans to buy ASML’s newest generation of high numerical aperture lithography machines. That’s it. No hedging, no “evaluating the roadmap,” no polite corporate fog. The largest chip manufacturer on the planet looked at a $400 million machine built for the AI era and said, essentially, we’re good.
ASML’s stock took the hit. Then came the bigger one, when the company told the market it “cannot confirm” growth for 2026 and roughly $30 billion in value evaporated. For a company that just posted full-year 2025 net sales of $39.16 billion and net income of $11.5 billion, that’s a strange place to be. Record numbers, record AI demand, and a shrug about next year.
What a $400 million machine still can’t do
The technical problem is almost comically basic once you strip the jargon away. ASML’s top-end High NA systems use sharper optics to etch smaller transistors than the previous generation. Physics being what it is, that improved resolution comes with a smaller printable area. The machine cannot print the largest chip designs in a single exposure. Anything bigger has to be split up and stitched together across multiple steps.
Multi-step processes are not free. They’re slower, they need more precision at each stage, and every extra step is another place where yield goes to die. So the customer buying the most expensive tool in semiconductor manufacturing gets handed a workflow tax on the exact designs that need the most help.
Now think about which chips are the biggest. AI accelerators. The giant reticle-hungry monsters powering every model you’ve ever prompted. Demand for those parts is the reason ASML had the year it had. It’s also the category most exposed to this limitation. Strong AI demand is making the constraint worse, not better, because the designs keep growing in exactly the direction the machine can’t stretch.
The fix arrives sometime next decade
A solution is planned. It’s expected between 2031 and 2033.
Read that timeline against the pace of the AI hardware cycle. Chip designs are being redrawn on multi-year cadences, and the fix for the field-size problem lands two full architecture generations out at best. If you run a fab, that’s not a roadmap item, that’s a rumor with a date attached. Zhang’s answer starts to look less like a snub and more like arithmetic.
I review AI tools for a living, and this is a pattern I recognize. A vendor ships an expensive upgrade that improves one dimension while quietly regressing another, then asks buyers to absorb the workaround cost until a future release cleans it up. In software, that release is six months out and often still doesn’t fix it. Here it’s six years out and involves optics, precision stages, and a supply chain no competitor can replicate. Same shape, much higher stakes.
Why this matters if you never touch a wafer
Everyone building on top of AI infrastructure is downstream of this. When the largest AI chips require costly multi-step patterning, that cost has to land somewhere. It lands in the price of accelerators, then in the price of compute, then in the per-token cost of the agent framework you’re evaluating this quarter. The abstraction layers are thick enough that nobody notices, but the floor under inference pricing is made of things like this.
It also explains why the “compute gets cheaper forever” assumption baked into so many AI product pitches deserves more suspicion than it gets. ASML has close to a monopoly on the machines at the leading edge and it still can’t promise growth next year. That’s not a demand problem. Demand is the one thing nobody is questioning. It’s a delivery problem, and delivery problems in this industry take a decade to solve because there is no way to parallelize physics.
The uncomfortable read
ASML is not in trouble. A company earning $11.5 billion on $39 billion in sales, with Q4 revenue of $11.62 billion, is not a company anyone should feel bad for. But the story sold to investors was that AI demand would push customers up the ladder into High NA systems on a predictable schedule. TSMC just declined to climb, and the reason isn’t sentiment, it’s that the tool doesn’t yet do the thing the biggest chips need it to do.
The honest summary is that the most advanced manufacturing tool humans have built has a size limit, the workaround is expensive, AI chips keep getting bigger, and the real answer is somewhere in the early 2030s. Every projection about cheap abundant AI compute between now and then is running on the assumption that this gets worked around gracefully.
I’d hold that assumption loosely.
🕒 Published: