The EUV light source is the plasma-based subsystem that generates extreme-ultraviolet photons inside a lithography scanner, and because its power and uptime set wafer throughput it sits underneath what analysts now call the single durable ceiling on AI compute scaling.
Summary
Extreme-ultraviolet lithography prints the smallest features in modern logic and memory by exposing wafers with photons far shorter in wavelength than the deep-ultraviolet light used previously. The hard part is making those photons in industrial quantity. EUV is absorbed strongly by essentially all materials, which is why the field has almost no transmissive optics at these wavelengths and why optics research is still working on basic components: a 2026 demonstration of the first broadband transmissive EUV metagrating describes strong material absorption as the limiting factor and reports phase-based operation up to 50 eV photon energy with an octave-spanning 25 eV bandwidth. The production answer to photon generation is laser-produced plasma (LPP): a high-power drive laser strikes a target to create a hot plasma that radiates in the EUV band ref.
The parameters that decide the technology are source power (photons per second delivered to the wafer, which sets exposure time and therefore wafers per hour), collector and optics efficiency, and availability, because a scanner costing tens of millions of euros is worthless while its source is down. None of those numbers appear in the sources supplied here, and that is the single biggest gap in this assessment: the sources establish the industrial and geopolitical position of EUV very well and its source-level engineering not at all.
There are alternative EUV generation architectures in the physics literature. A 2026 experiment demonstrated a fully coherent echo-enabled harmonic generation free-electron laser driven by a single sub-microjoule seed, removing the hundred-megawatt peak-power laser systems and dual-laser synchronisation that previously restricted such machines to kilohertz repetition rates. That work targets time-resolved ultrafast spectroscopy, not lithography, and the sources make no claim about FEL sources reaching lithographic power or cost.
Evidence base
- SemiAnalysis pins the long-run ceiling on AI compute to ASML producing fewer than roughly 100 EUV machines a year, with the binding constraint shifting from CoWoS packaging to EUV post-2028 (14 Mar 2026) ref.
- ASML guides 51-53% gross margin for 2026 with a 2030 target of 56-60%, on 2026 revenue guidance of EUR 36-40bn and China normalising to ~20% of sales from 33% in 2025 (15 Apr 2026) ref ref.
- TechInsights confirmed SMIC N+3 in volume production at a 5 nm-class node using DUV multi-patterning only, no EUV (11 Dec 2025) ref; a follow-up teardown puts it at ~32.5 nm M0 pitch and ~113 MTr/mm2 at steep cost in complexity and mask count (14 Jun 2026) ref.
- IBM announced Nanostack at a 0.7 nm node with ASML High-NA EUV as a named ecosystem contribution, ~100bn transistors on a fingernail-sized chip and a production pathway within about five years (25 Jun 2026) ref.
- A 392-dot, 7x7 silicon quantum dot array was fabricated in a 300 mm CMOS process patterned by EUV lithography, reaching threshold-voltage variability below 63 mV standard deviation at 17 nm oxide thickness (14 May 2026).
- The first broadband transmissive EUV metaoptic reports 0.04 deg/nm angular dispersion, directionality up to 5.8 and phase-based operation to 50 eV, framed against strong material absorption that limits transmissive optics at EUV wavelengths (8 Jul 2026).
- US export controls have blocked EUV NXE platforms to China across the Oct 2022, Oct 2023 and Dec 2024 BIS rounds 2024 Bis Export Controls China Semis; the proposed MATCH Act would additionally ban maintenance of previously sold machines (24 Jun 2026) ref.
Open questions
- What is the current delivered in-band source power, conversion efficiency and availability, and does High-NA EUV require more source power per exposure than the 0.33-NA generation?
- Is the light source, rather than optics or precision stages, the binding constraint on ASML’s sub-100-unit annual output?
- Could a free-electron-laser source of the kind demonstrated with sub-microjoule seeding ever reach lithographic average power and cost per photon, or is the repetition-rate and stability gap structural?
- If the MATCH Act passes and bans maintenance of previously sold machines ref, how quickly does an unserviced installed base lose usable source uptime?
Assessment drafted 2026-08-31 from up to 18 KB sources using the technology-scorecard framework; scores are a draft read pending review.