Definition
Extreme ultraviolet lithography is how the smallest features on an advanced chip get printed, using light with a wavelength of 13.5 nanometres — light so energetic that it is absorbed by air, by glass, and by every lens ever made. That single physical fact dictates everything else about the machine: it has no lenses, it runs in a vacuum, its light comes from vaporising tin with a laser 50,000 times a second, and it is built by exactly one company on Earth (ASML).
Wafers, yield, nodes and the foundry model belong to Semiconductor Manufacturing; stacking and bonding the finished dies belongs to Advanced Packaging. This page is about the one step in that flow with no second supplier — and why that is a fact of physics and accumulated engineering rather than a fact about business.
How It Works
Why 13.5 nanometres, and not something more convenient
Printing a chip is projection: a pattern on a mask is shrunk and focused onto a light-sensitive
wafer. How small a feature that projection can resolve is governed by the Rayleigh criterion, which
ASML writes as CD = k₁ × λ / NA — where CD is the smallest printable feature, λ is the
wavelength, NA is the numerical aperture of the optics, and k₁ is a process factor whose
hard physical floor is 0.25
(ASML).
Only two of those three terms can move. So watch what happens when you exhaust one of them.
The previous generation used a 193 nm argon-fluoride laser, and the industry spent twenty years squeezing the other two variables. Immersing the wafer in water raised the numerical aperture to 1.35, the highest in the industry (ASML). Plug it in at the physical limit:
CD = 0.25 × 193 ÷ 1.35 = 35.7 nm
About 36 nanometres. That is the floor, and no amount of engineering goes under it with 193 nm light. Below it you must print a layer with several exposures — double patterning, then quadruple patterning — which multiplies masks, multiplies process steps, and multiplies the ways the pieces can land in the wrong place: overlay error is around 0.6 nm for two equal exposures but ≈2.0 nm for self-aligned double or quadruple patterning, and the cost of a 193i layer roughly doubles with SADP and triples with SAQP (Wikipedia). The 2010s were spent paying that tax.
Now change λ instead. Drop to 13.5 nm and accept a much worse numerical aperture of 0.33 — mirrors cannot collect light the way a lens can — and the same equation gives:
CD = 0.25 × 13.5 ÷ 0.33 = 10.2 nm
The wavelength shrank by a factor of 14 and swamped the loss in NA. In production ASML rates its
low-NA EUV systems at 13 nm resolution, which corresponds to k₁ ≈ 0.32 — not the theoretical
floor, but comfortably past anything 193 nm can reach in a single shot. The High-NA generation raises
NA to 0.55, and the same k₁ predicts 0.32 × 13.5 ÷ 0.55 = 7.9 nm — against ASML's published
figure of 8 nm (ASML). The spec sheet
is just this equation with the constants filled in.
The price of 13.5 nm: no lenses, no air, and a light budget that barely survives
Here is the catch that makes EUV a twenty-year engineering project rather than a wavelength swap. All matter absorbs EUV radiation (Wikipedia); a 13.5 nm photon carries 91–93 eV, against 6.4 eV for the old 193 nm light, and it deposits that energy in whatever it touches. ASML puts it plainly: "most materials absorb EUV light, the lenses would absorb the light in the system" (ASML). So there are no lenses. There is no air. The source, the optics, the mask and the wafer all live in a vacuum, and the pattern is steered by mirrors — including the mask itself, which reflects rather than transmits.
But a mirror for EUV cannot be a polished metal surface either, because metal absorbs it too. It has to be a Bragg reflector: forty-odd alternating layers of molybdenum and silicon, each a few atoms thick, tuned so that the weak reflections from every interface add up in phase. ASML describes "over 100 layers of materials" per mirror, polished "to a smoothness of less than one atom's thickness — if the mirrors were the size of Germany, the tallest 'mountain' would be just 1 millimetre high" (ASML, above). And after all that, the best such mirror still returns only about 70% of the light that hits it (Wikipedia).
Seventy percent sounds fine. Compound it.
Worked example: the light budget
An EUV scanner's optical path involves roughly 11 reflections between the source and the wafer. Each one keeps 70% of what it receives, so the fraction arriving is:
0.70¹¹ = 0.0198 → about 2%
Which is exactly what the literature reports: with 70% per mirror and 11 reflections, less than 2% of the generated EUV light reaches the wafer (Wikipedia, above). Ninety-eight percent of the most expensive light ever manufactured is turned into heat inside the machine that made it.
Follow that budget all the way back and the number gets stranger. The NXE:3800E runs an EUV source of over 500 W (SemiWiki), and that source is driven by a 40-kilowatt CO2 laser built by Trumpf (TRUMPF). So:
- Drive laser into the tin: 40,000 W of infrared.
- EUV radiated out of the plasma and collected: ~500 W — a conversion efficiency of about 1.25%.
- EUV surviving eleven mirrors:
500 × 0.0198 ≈ 10 Wat the wafer.
Forty kilowatts in, ten watts out — roughly 0.025% of the drive laser's power ends up doing the printing. That is the deal EUV strikes with physics, and it is why a single scanner draws up to 1,170 kW from the wall, with High-NA machines expected to reach around 1,400 kW (TechInsights, via TechSpot) — a number that belongs in the energy budget of AI as much as in this one.
The light source is a plasma engine, not a lamp
Nothing emits usefully at 13.5 nm on demand, so the machine makes its own light by destroying matter. Molten tin droplets about 25 microns across are fired into a vacuum chamber, and each is hit twice by the CO2 laser: a low-intensity pulse flattens the droplet into a disc, then a far more powerful pulse blows it into a plasma that radiates at 13.5 nm. "This process is repeated 50,000 times every second" (ASML). Trumpf's laser fires the matching 50,000 pulses per second, amplifying a seed of a few watts by more than 10,000× along a beam path up to 30 metres long (TRUMPF, above).
Every one of those droplets must be hit, twice, in the right order and the right place, fifty thousand times a second, for years. That is merely the light source. The rest of the machine runs to roughly 100,000 parts, and shipping one takes 40 freight containers, three cargo planes and 20 trucks (ASML).
Real-World Applications
Every leading-edge AI chip. The logic dies behind GPU computing, custom accelerators and the DRAM in High Bandwidth Memory are patterned on these machines. There is no alternative path to the leading edge and no second vendor to switch to: if you are running a frontier model on hardware built in the last few years, EUV printed it.
Export control, as a consequence of the chokepoint. Because exactly one company can build the machine, and because building it depends on Zeiss's optics and Trumpf's laser, an EUV scanner cannot be shipped anywhere without a Dutch export licence — and none has ever gone to China. In June 2026 ASML restated publicly that it had delivered neither a complete EUV system nor components developed specifically for EUV systems to Chinese customers (Verdict). Set the politics aside and look at the mechanism, because the mechanism is the lesson: where there is one supplier, one licensing decision in one country determines who can manufacture advanced chips at all. Nothing else in the AI stack has that shape.
Capacity planning. ASML recognised revenue on 48 EUV systems in the whole of 2025 — 44 low-NA plus 4 High-NA, against 44 systems in 2024 (ASML 2025 Annual Report) — while carrying a backlog of €38.8 billion (ASML). Roughly four dozen machines a year, worldwide, each needing months to install, is the physical ceiling on how fast leading-edge capacity can grow. Every forecast of AI compute supply is, underneath, a forecast about this.
Key Concepts
- Wavelength beats everything else in the resolution equation. NA and k₁ are bounded — k₁ cannot go below 0.25, and NA is limited by what mirrors can collect. λ is the only term with room in it, which is why the industry paid an almost absurd price to move it, and why there is no obvious next move after 13.5 nm.
- Reflectivity compounds, and 70% is not a good number. One mirror at 70% is a rounding error; eleven of them throw away 98% of the light. Almost every hard problem in EUV — source power, resist sensitivity, throughput, the megawatt power draw — descends from that exponent.
- A single supplier is not the same as a monopoly on an idea. The barrier is not a patent that expires. It is a plasma source, an optics house, a laser company and thirty years of failures that taught them what does not work. Nikon abandoned EUV development in 2011; the technology reached high-volume manufacturing in 2019. A newcomer starting today would be starting that clock over.
Challenges
What breaks: believing a chip shortage is a money problem. It is the most common mistake made about this industry, and the arithmetic above is the refutation. When leading-edge capacity is short, the binding constraint is not silicon, not fab buildings and not capital — it is scanners, at roughly 48 a year for the entire planet, with lead times over a year. And ASML cannot simply build more, because its own output is gated by how fast Zeiss can polish and coat mirrors to sub-atomic tolerances and how fast Trumpf can build 40 kW lasers. You cannot pay a supply chain to skip a learning curve it took decades to climb. Money buys a place in the queue; it does not shorten it.
Source power is the throughput. Because only ~2% of the light survives the optics, a scanner's productivity — wafers per hour, and therefore cost per chip — is downstream of how many watts the tin plasma can produce and how long the collector optic survives being coated in vaporised tin. More power means more tin debris, and more debris means a shorter-lived collector. Throughput is won in that trade, not in the printing.
Photon shot noise. A 13.5 nm photon carries about 14 times the energy of a 193 nm one, so for the same energy delivered to the resist there are about 14 times fewer photons. In a feature a few nanometres wide, the number that happen to land is small enough that ordinary statistical fluctuation shows up as ragged edges and randomly missing contacts — a defect mode that is not a manufacturing error but a consequence of counting. Shorter wavelengths make it worse, which is one reason "just go shorter again" is not the roadmap.
High-NA is not obviously worth it. Raising NA to 0.55 shrinks the field the scanner can expose, so large dies must be split across two exposures and stitched — and the machine costs roughly twice as much: Samsung bought two High-NA scanners for $773 million, about $386 million each (TechPowerUp), against roughly $180 million for a low-NA system (Wikipedia). Whether that beats two exposures on a cheaper machine is a live argument, not a settled one.
Future Trends
The direction of travel is more numerical aperture, not less wavelength — High-NA at 0.55, with Hyper-NA studies beyond it — precisely because there is no comfortable wavelength below 13.5 nm. Everything already absorbs EUV; going shorter makes the mirrors worse, the photon count lower and the shot noise louder. The industry chose the last wavelength it could build a machine for.
The more consequential trend is that the chokepoint is not loosening. Each generation is harder to build than the last, depends on the same two irreplaceable suppliers, and is produced in smaller numbers per year than the demand curve for AI data centers implies. When people call compute the strategic resource of this decade, the honest version of the claim is narrower and stranger: a few dozen bus-sized machines a year, from one town in the Netherlands, with optics from one company in Germany.