Terafab, Free-Electron Lasers and ASML

Two stories collided in the first week of August 2026. SpaceX and Tesla confirmed Terafab, a semiconductor megafactory in Grimes County, Texas. Separately, free-electron laser lithography moved from research paper to funded programme. The internet promptly welded them together into "Musk builds fab that kills ASML".

The physics is real and the funding is real. The link between them is mostly inference. Worth separating those before going further.

What is confirmed, and what is not

The official Terafab announcement says nothing about lithography, EUV or light sources. The FEL connection comes from commentary on X and from the fact that both stories broke in the same week. Treat it as a plausible thesis, not a plan.

What is genuinely happening: xLight is building an FEL EUV source with US federal money, and ASML is working with them on it.

How ASML makes EUV light today

ASML's method is laser-produced plasma (LPP), and it is a controlled explosion repeated 50,000 times a second.

1. A generator fires ~30-micron droplets of molten tin into a vacuum chamber.
2. A high-power CO2 laser hits each droplet twice. A pre-pulse flattens it into a disk, a main pulse turns it into plasma hotter than the surface of the Sun.
3. The ionised tin emits a broad spectrum with a peak at 13.5 nm.
4. A collector mirror gathers what it can and feeds it into the scanner.

It works, and it is the reason EUV shipped at all. The problems are structural:

• Wall-plug efficiency below 0.1%. Almost all the energy heats tin rather than making photons you can use.
• Tin debris. Atoms, ions and particles coat the multilayer collector mirror. It degrades, it gets replaced, the tool stops.
• Power ceiling. Current sources deliver a few hundred watts, recently pushed toward 1 kW. That caps throughput around 200 wafers per hour.
• Photon scarcity. Broadband, unpolarised light with too few photons per exposure produces stochastic defects, random patterning errors that get worse as features shrink.

How a free-electron laser does it instead

No target, no plasma, nothing vaporised. Just electrons and magnets.

1. Electrons are accelerated to relativistic speed in a linear accelerator, usually an energy-recovery linac (ERL).
2. The beam passes through an undulator, a long section lined with alternating magnets.
3. The alternating field forces the electrons to wiggle side to side. Accelerating charges radiate, and at relativistic speed the emission is Doppler-shifted into the extreme ultraviolet.
4. The light co-propagates with the electrons and drives them into microbunches spaced exactly one wavelength apart, which produces coherent amplification. This is SASE, self-amplified spontaneous emission.
5. Tune the beam energy and the undulator period and you set the wavelength. 13.5 nm today, shorter for future nodes.

The ERL is the economic trick. Spent electrons hand their energy back to the incoming beam instead of being dumped, so you get high current and high repetition rate for the same electricity bill.

Side by side

The tunability line matters most long term. Tin emits at 13.5 nm because that is what tin does. Below EUV, LPP has nowhere to go. An FEL just needs a different magnet setting.

EUV Light Source: Is the Future in a Particle Accelerator?

IEEE Spectrum on KEK's energy-recovery linac work

https://spectrum.ieee.org/euv-fel

Japan's KEK has been on this longest. Their compact ERL runs at 17 MeV, produced 20-micron infrared SASE in 2023, and targets 10 kW serving multiple scanners. Their 2021 estimate for a prototype was 40 billion yen, roughly $260M, plus 4 billion yen a year to run.

xLight, the company actually building it

The US commercial vehicle, and the reason this stopped being a physics-department story.

Their claims: 4x the EUV power of an LPP source, a 50% cut in EUV cost, 20% lower total wafer cost. They argue EUV is 40% of the cost of a wafer and that current sources supply only a quarter of the light the tools could use.

xLight

FEL EUV source, Pat Gelsinger as executive chairman

https://www.xlight.com

Commerce finalizes CHIPS incentives with xLight

$150M award for the Albany prototype, June 2026

https://www.nist.gov/news-events/news/2026/06/...xlight-support

So does this disrupt ASML?

Less than the framing suggests. Three reasons.

xLight is not attacking ASML. It wants to be a supplier. The stated plan is to sell the source into ASML's tools, and Fouquet has said ASML is collaborating with xLight on technology demonstrations. A vendor swap inside the supply chain is not a competitor.

The source is not where the moat is. ASML's monopoly is the scanner: the Zeiss optics, the wafer stage, overlay accuracy, metrology, and thirty years of integration. Replacing the light bulb does not replace the machine. ASML bought its own source in the first place rather than inventing it, taking Cymer for €1.95B when announced in October 2012, closing at a reported $3.7B in May 2013.

The numbers are smaller than they look. ASML's FY2025 net sales were €32.7B, of which EUV systems were €11.6B across 48 systems, a blended ASP around €242M. The source is a fraction of that ASP. The real exposure is the annuity: tin debris means collector mirrors, service visits and consumables, and a clean source removes that revenue permanently. Painful, not existential.

Where it would actually hurt

The genuine threat is not the technology, it is the business model that becomes possible with it. If one FEL feeds ten scanners, a fab can buy EUV light as a utility from a party that is not ASML. That decouples the light from the tool, and a decoupled subsystem is the first step toward a second scanner vendor. That is a decade-plus scenario requiring someone to solve the optics and stage problem, which nobody has.

For now the arrow points the other way. Terafab at 14A needs EUV, EUV means ASML, and industry estimates put a leading-edge logic fab at 80 to 100 scanners for 20,000 to 30,000 wafer starts a month. Terafab is shaping up to be one of ASML's largest customers, not its executioner.

ASML

Sole supplier of EUV lithography systems

https://www.asml.com

What would have to be true

For the disruption thesis to land, all of these need to happen:

1. xLight's Albany prototype works in 2028, roughly on time.
2. It runs at fab-grade uptime. A research accelerator that drops out for a day is fine. A source feeding ten production scanners is not.
3. The beam transport works. Getting kilowatts of EUV down a distribution line to multiple tools without cooking the mirrors is unsolved at production scale.
4. Somebody accepts single-point-of-failure risk. One source, ten scanners, one accelerator fault stops the fab.
5. The economics beat a depreciated LPP fleet, not just a new one.

ASML looked at FEL years ago and picked LPP because it carried less risk. That judgement was correct for 2010. Whether it stays correct through 2030 is the actual question.

My take

The physics case is not in doubt. Sub-0.1% efficiency and a debris-generating source are embarrassments the industry tolerates only because there was no alternative. An accelerator that converts electron kinetic energy straight into coherent photons is obviously the better engine.

The interesting bet is Gelsinger. A former Intel CEO chairing the company that supplies ASML's most critical subsystem, funded by the CHIPS Act, prototyping in Albany, is a US industrial-policy play as much as a technical one. The goal is not to kill ASML. It is to make sure the next generation of EUV has an American component in the critical path.

Terafab is the customer in this story, not the weapon. Anyone claiming otherwise is reading an X post, not a filing.

Further reading

• xLight
• IEEE Spectrum on FEL EUV
• Commerce and NIST finalise xLight's CHIPS award
• Electrek on the Terafab confirmation
• Tom's Hardware on ASML's talks with Musk
• ASML and KEK