How ASML's EUV scanner became the bottleneck on AI
Every leading-edge AI chip in production passes through a machine that costs $200 to $400 million, weighs 180 tons, and is built by a single Dutch company. The engineering miracle that makes 5 nm and below economical, and the monopoly that came with it.
EUV scanner production rate and rising cost per machine together cap how fast leading-edge wafer capacity can grow. ASML is the gate; the AI accelerator supply curve is a derivative of EUV scanner shipments, with a lag.
EUV is what makes 5 nm and below economical
Below the 7 nm node, DUV lithography requires multi-patterning to fake a finer pattern, with each extra pass adding cost, defect risk, and time. EUV's 13.5 nm light collapses many of those passes back into a single exposure, restoring economics that multi-patterning was eroding. Every leading-edge logic node from TSMC N5 onward, plus the leading-edge DRAM that feeds HBM stacks, depends on it.
The light source is the engineering miracle
Producing 13.5 nm light at industrial power is not something nature provides. The ASML scanner does it by firing about 50,000 droplets of molten tin per second into a vacuum chamber. A high-power CO₂ laser hits each droplet twice: first to flatten it, then to vaporize it into a plasma that emits EUV. The machine then collects that light with a mirror system polished to atomic flatness, because EUV is absorbed by air, glass, and conventional optics.
From the plasma, the light reflects through roughly thirteen mirrors before reaching the mask. Every reflection loses 20 to 30% of the power. By the time the photons arrive at the wafer, only a few percent of what the plasma originally emitted is still there. The full scanner is about the size of a city bus, weighs around 180 tons, ships in dozens of crates on multiple Boeing 747s, and is assembled on-site at the customer fab.
Source: ASML technical briefings; "Chip War" by Chris Miller, 2022
The mirrors are as flat as Germany
EUV mirrors are not ordinary mirrors. They are stacks of forty to fifty alternating molybdenum and silicon layers, each a few nanometers thick, and the surface has to be polished to a flatness so extreme that, as Chris Miller put it in Chip War, "if the mirrors were scaled to the size of Germany, their biggest irregularities would be a tenth of a millimeter."
Only one company, Carl Zeiss, knows how to make them at production scale. That single supplier is the reason ASML's lithography monopoly is durable. A competing scanner program would not just have to invent a tin-plasma light source from scratch; it would need to invent a Zeiss too. Decades, not years.
Source: Chris Miller, "Chip War", 2022
ASML is a monopoly because nobody else solved the optics
EUV lithography took ASML, Zeiss, the US Department of Energy national labs, and Cymer (the laser maker ASML later acquired) several decades to make work at production volume. The original research was largely US-funded; the technology was sold to ASML in the 1990s before modern export-control regimes existed to stop it. The result is a single supplier for the most consequential piece of equipment in modern semiconductors, headquartered in Veldhoven, Netherlands.
There is no second source. China cannot buy the latest scanners under US and EU export controls. The one US-soil leverage point is Cymer: the light source must be manufactured stateside, and the US has veto power over who ASML can ship to. A competing scanner program would take a decade to ramp and would still depend on the same Zeiss optics. The bottleneck is real and durable.
High-NA EUV pays the half-field tax
High-NA EUV uses a larger numerical aperture (0.55 versus the standard 0.33) to resolve finer features, enabling nodes below 2 nm. The catch is anamorphic optics: the new mirror geometry means each exposure covers only half the wafer area the previous tool could. The buyer needs not one $400 million machine but two of them to pattern the same surface. ASML compensates by accelerating the scanner mechatronics to fighter-jet pace with nanometer precision, but the trade-off is real and expensive.
Intel took the first High-NA deliveries; TSMC will follow on its own timeline. The rumored next step, Hyper-NA, would push numerical aperture higher again at a cost approaching a billion dollars per tool. The cycle repeats: more advanced lithography unlocks more advanced chips, which unlocks more capable models, which raises demand for more advanced chips. The slow link in that chain is the machine in Veldhoven.
Source: ASML quarterly reports, 2024-2026
The strategic read
When a forecaster projects AI compute supply five years out, the binding number is not how many fabs are announced. It is how many EUV scanners ASML can build, who is allocated each one, and whether High-NA hits its volume targets on schedule. Geopolitics, export controls, and capacity announcements all matter; they all run downstream of one Dutch supplier and one set of mirrors.
The cost curve is the next variable to watch. Two startups, xLight and Substrate, are betting they can bend it from first principles.