The Tin Droplet That Etches Every Chip

Inside a clean room outside Veldhoven, a machine shoots a stream of 25-micron tin droplets across a vacuum chamber at roughly 70 meters per second [S1]. A CO₂ laser hits each one twice. The first pulse — the "pre-pulse" — flattens the spherical droplet into a pancake [S1]. The second pulse vaporizes the pancake into a plasma at around 220,000°C [S1]. The plasma emits a flash of 13.5-nanometer extreme-ultraviolet light. Then the next droplet arrives. The cycle repeats fifty thousand times per second [S1].

Every transistor inside every modern phone, every AI accelerator, every data-center GPU made at advanced nodes is etched by that flash.

Moore's Law was suffocating at 193 nanometers. Old deep-UV lithography had been stretched, immersed in water, and multi-patterned for years; going shorter meant going to extreme-UV at 13.5 nm. The problem is that 13.5-nm light is absorbed by air, glass, and every transparent material ever made [S3]. There is no lens you can put it through. So the entire optical path lives in hard vacuum, and instead of lenses you bounce the light off mirrors [S3].

The mirrors are not mirrors in any familiar sense. They're stacks of 40 to 80 alternating layers of molybdenum and silicon, each layer about 3 to 4 nanometers thick — a few dozen atoms [S3]. Stacked correctly they form a Bragg reflector that returns roughly 70% of the 13.5-nm light hitting it [S3]. Each surface must be polished to a roughness of around 50 picometers RMS — sub-0.1 nanometer [S3]. Zeiss's own analogy: scale one of their mirrors up to the size of Germany, about 360,000 square kilometers, and the tallest bump — the Zugspitze, as they put it — would be 0.1 millimeters [S3]. That is the smoothest surface humans have ever made, and a single EUV machine has about a dozen of them, aligned to picometer stability in vacuum [S3].

Here is the trick almost nobody notices: 13.5 nanometers wasn't chosen because it's the perfect wavelength for printing tiny features. It was chosen because Mo/Si happens to be the only multilayer pair that reflects efficiently, and 13.5 nm is just where that material's reflectivity peaks [S3]. Alternative wavelengths at 11 nm and 17.2 nm have been studied but never adopted at production scale. The whole modern chip industry runs on a wavelength reverse-selected from the mirror chemistry. The optics chose the light, not the other way around.

Producing the light was harder than reflecting it. In the early 2000s, ASML evaluated several potential suppliers for the EUV source [S10]. The one whose laser-produced-plasma tin source actually scaled to production power was Cymer, a San Diego startup founded in 1986 by two college friends, Robert Akins and Richard Sandstrom [S10]. In 2013, ASML bought Cymer outright for €1.95 billion — roughly $2.5 billion — because there was no other way to source the photons [S10].

The part the technology histories don't quite get across: in 1996, Congress voted to terminate DOE funding for EUV research [S4]. In 1997, a SEMATECH industry task force ranked EUV dead last of four candidate technologies, behind X-ray proximity lithography, e-beam, and ion projection [S4]. Intel rescued it. Intel led EUV-LLC — a consortium that also included AMD and Motorola — funding a roughly $250 million, three-year cooperative agreement with the DOE labs at Sandia, Berkeley, and Lawrence Livermore [S4][S5]. Two years after the dead-last ranking, the same industry community reversed and voted EUV the most probable path forward [S4]. The US national labs essentially invented productionizable EUV; ASML, in Veldhoven, productionized it. That history is also why the US government claims a legitimate veto over where the machines can be shipped.

A finished low-NA EUV machine — the workhorse NXE:3600D — runs roughly $150 to $200 million per unit, with high-NA EXE:5000 systems reported in the $350 to $400 million range [S6]. Each machine is assembled from modules manufactured at about 60 ASML sites around the world and shipped to Veldhoven for integration [S6]. Delivered, it arrives in about 40 freight containers and is installed on-site by ASML engineers [S6]. The TRUMPF CO₂ laser alone is a separate room-sized object [S1]. ASML is now roadmapping a jump from 50,000 droplets per second to 100,000, with three pulses per droplet — two for pre-shaping and one for ignition — for a total of 300,000 pulses per second, targeted to lift throughput around 50% by 2030 [S2].

The political consequence arrived on June 30, 2023, when the Netherlands imposed export licensing on three of ASML's most advanced DUV immersion machines — the NXT:2000i, 2050i, and 2100i — for shipment to China [S7]. EUV had already been informally blocked for years; the 2023 restrictions extended the perimeter to the older 193-nm immersion systems SMIC was using to multi-pattern its way to 7-nm chips [S7]. China was 24% of ASML's revenue that quarter, roughly tripled from a year earlier [S7]. Dutch MPs complained that economic jewels were being thrown away [S7]. The reason the US can force the Netherlands to make that call — and the reason Japan has signed similar restrictions — is that there is no second supplier on Earth. You either buy from ASML or you don't print at these nodes.

Or you build a fundamentally different machine. Huawei is reportedly testing an EUV prototype at its Dongguan facility that uses laser-induced discharge plasma, or LDP — tin evaporated between two electrodes and ionized by high-voltage discharge, no flying droplets, no CO₂ laser [S8]. LDP was one of the candidate source architectures Europe explored and abandoned in the 2000s in favor of LPP [S8]. Trial production was reportedly targeted for the third quarter of 2025 with rollout in 2026, but the prototypes emit only 50 to 100 watts of EUV power versus the 250 watts or so needed for commercial-throughput production [S8]. Analysts split on whether this is a year away, a decade away, or vaporware. The premise of the export-control regime is that the monopoly is permanent. The premise of Huawei's prototype is that it isn't.

Then there's the problem nobody has solved. EUV at 13.5 nm delivers each photon at 92 electron-volts — high enough that when it's absorbed in the photoresist, it doesn't just expose a single spot; it ejects a cascade of low-energy secondary electrons that scatter through the resist with an exponential blur distribution reaching as far as 30 nanometers [S9]. At a 3-nanometer pitch, that blur is bigger than the feature. Unexposed stochastic defects dominate at low laser doses; exposed stochastic defects dominate at high doses; there is a defect-density floor that cannot be driven to zero by tuning exposure [S9]. Frederick Chen, a former IC engineer who writes extensively on the physics, calls it a "stochastic valley of death" [S9]. The optical part of EUV — the tin, the lasers, the mirrors, the vacuum — works. The chemistry of what happens when the light arrives at the wafer may turn out to be the actual ceiling at sub-2-nm nodes [S9].

Step back and the bet looks vertiginous. A wavelength chosen because of the mirror material it could be reflected off of. A light source bought outright from a San Diego startup whose tin-droplet trick everyone else had given up on [S10]. Mirrors polished to half an atom's roughness by the only optics firm that ever got it to work [S3]. A capital program kept alive in the 1990s by Intel writing checks to DOE labs working on a technology a Congressional vote had just defunded [S4][S5]. And on top of that stack: the AI boom, the smartphone economy, and an export-control regime that depends on no government on the planet being able to replicate any of it.

The machine ships in 40 freight containers and takes months to assemble in a customer fab [S6]. It is the most consequential piece of industrial equipment most people will never see, etching the world's transistors out of vaporized metal pancakes, 50,000 times a second [S1].