EUV Lithography: How We Print Chips With Light You Can't See
Every advanced chip — your phone's processor, the GPU in your laptop, the AI accelerators in data centers — was manufactured using extreme ultraviolet lithography. EUV is the most complex optical system ever built for manufacturing, and exactly one company in the world can build the machines that do it. Here's how it works, why it took 30 years to commercialize, and what comes next.
The Basic Principle
Chip manufacturing patterns features onto silicon wafers using light projected through masks (stencils). Smaller features need shorter wavelengths of light — the same reason you can't paint fine details with a wide brush. Previous-generation lithography used deep ultraviolet (DUV) light at 193nm wavelength. EUV uses 13.5nm — about 14× shorter.
That wavelength reduction enables printing features as small as 8-13nm in a single exposure, versus the multiple exposure passes (multi-patterning) that DUV required for similar dimensions. Multi-patterning worked — TSMC reached 7nm using DUV multi-patterning — but it's slow, expensive, and introduces alignment errors with each additional pass.
Inside an EUV Machine
An ASML EUV lithography system (the NXE:3600E, for instance) is a marvel of engineering. The key components:
Light source: A high-powered CO2 laser fires 50,000 pulses per second at tiny droplets of molten tin, each about 25 micrometers in diameter. The laser impact heats the tin to 500,000°C, creating a plasma that emits 13.5nm EUV light. Yes, really — the light source works by shooting tin droplets with a laser fifty thousand times a second.
The conversion efficiency from laser power to usable EUV light is only about 5-6%. The laser consumes approximately 30 kilowatts, producing about 250-300 watts of EUV light at the source. After losses through the optical system, perhaps 20-30 watts actually reaches the wafer. This abysmal efficiency is why EUV machines consume enormous power — over 1 megawatt per system including all support equipment.
Related reading: HBM4 Memory Technology: Architecture, Bandwidth, and the AI .
Optics: EUV light is absorbed by everything, including glass and air. There are no EUV-transparent lens materials. Instead, the optical system uses multilayer mirrors — alternating layers of molybdenum and silicon, each 7nm thick, stacked to form a Bragg reflector that reflects about 70% of incident EUV light. The system contains 11 mirrors (6 in the illumination path, 4 in the projection optics, plus the mask). Each mirror absorbs 30% of the light that hits it, which is why so little power reaches the wafer.
These mirrors are manufactured by Carl Zeiss SMT to surface accuracy of 25 picometers — about 1/10 the diameter of a silicon atom. The mirror surfaces are the smoothest objects ever manufactured by humans.
Vacuum: The entire optical path operates in near-perfect vacuum because air absorbs EUV light within millimeters. Managing the vacuum in a system with moving wafers, a plasma light source, and contamination from tin debris is an ongoing engineering challenge.
The Mask and Pellicle Problem
EUV masks are reflective (not transmissive like DUV masks) — the pattern is etched into a multilayer reflective coating on an ultra-flat glass substrate. These masks cost $500,000-1,000,000 each, and a complex chip design requires 80-100 masks.
Protecting these expensive masks from contamination during use requires a pellicle — a thin transparent membrane that sits above the mask surface. But making a membrane that's transparent to EUV light and mechanically strong enough to survive the vacuum and radiation environment inside the scanner is extremely difficult. Current EUV pellicles are made from polysilicon membranes just 50nm thick. They transmit about 85-90% of EUV light — better than the 0% transmission of thick pellicles, but every percent of light lost reduces throughput.
We covered a related topic in Interconnect Scaling: Why On-Chip Wiring Is the New Bottlene.
Throughput and Cost of Ownership
A modern EUV scanner processes about 160-185 wafers per hour at standard exposure doses. That sounds fast until you consider each wafer passes through the lithography step multiple times (for different layers), and a modern processor has 80+ lithography layers, of which 15-25 might use EUV.
Each EUV system costs approximately $200 million. A high-volume fab needs 10-20 EUV scanners (plus many more DUV scanners for non-critical layers). At $200M each, the lithography tools alone represent $2-4 billion of a $20+ billion fab investment. Operational costs (power, maintenance, consumables) add another $50-100 million per year per system.
Despite these staggering costs, EUV is actually cheaper per patterned feature than DUV multi-patterning for advanced nodes. A single EUV exposure replaces 3-4 DUV exposures, so even though the tool costs much more, the total patterning cost per wafer can be lower.
ASML's Monopoly
ASML is the only company that makes EUV lithography machines. This monopoly isn't artificial — it's the result of consolidating all the world's EUV development expertise into a single company over decades. Canon and Nikon, the other major lithography equipment makers, tried to develop EUV and gave up due to the immense technical challenges and investment required.
ASML's supply chain itself is concentrated. The optics come from Zeiss (no alternative supplier). The laser system comes from Trumpf (no alternative). If either company had a production problem, the entire advanced semiconductor industry would stall. There's no Plan B.
This connects to the ideas in Backside Power Delivery: Why Routing Power Under the Transis.
Order backlog for EUV systems extends 18-24 months. ASML can produce about 50-60 EUV systems per year, and they're working to increase that. But with every advanced fab in the world needing more EUV capacity, demand consistently exceeds supply.
What Comes After: High-NA EUV
ASML's next-generation High-NA EUV system (the EXE:5000 series) increases the numerical aperture from 0.33 to 0.55, enabling finer resolution. It'll be needed for the most critical layers at 2nm and beyond.
High-NA brings its own challenges: the higher aperture creates anamorphic (non-square) imaging that requires new mask infrastructure, and each machine costs $350-400 million. Only a handful of fabs will use High-NA in the near term — it's for the most critical layers where standard EUV can't achieve the required resolution.
The first High-NA systems were delivered to Intel for development in 2024, with TSMC and Samsung expected to receive production systems by 2026-2027. Whether High-NA EUV extends optical lithography to 1nm and beyond, or whether an entirely different approach is needed, remains an open question in the industry.