EUV Lithography: The $200 Million Machine Printing AI Chips

01 The Machine That Prints the Future

Inside a clean room in Eindhoven, Netherlands, sits a machine that weighs 150 metric tons, costs approximately $200 million per unit, and can fire pulses of light at a rate faster than a camera shutter could ever dream of capturing. This is an ASML extreme ultraviolet (EUV) lithography scanner, and it is the single most complex piece of manufacturing equipment ever constructed by human beings. Every advanced microprocessor on the planet — the Nvidia GPUs that train large language models, the Apple Silicon that powers iPhones, the Qualcomm Snapdragon chips inside Android flagships — owes its existence to a machine that only one company in the world knows how to build.

The concentration of power is staggering. ASML, a Dutch company with roots in a 1984 joint venture between Philips and the Dutch toolmaker ASM International, holds a monopoly on EUV lithography. There is no alternative supplier, no substitute technology at this resolution, and no imminent competitor. TSMC, Samsung, and Intel are the only three companies on Earth that operate these machines in production, and each pays a premium measured in hundreds of millions of dollars for the privilege.

02 From DUV to EUV: A Wavelength Revolution

Lithography is, at its core, a photographic process. Light passes through a mask — a stencil containing the circuit pattern — and is focused through lenses onto a silicon wafer coated in light-sensitive chemicals called photoresist. The shorter the wavelength of light, the smaller the features you can print. For decades, the industry used deep ultraviolet (DUV) light at a wavelength of 193 nanometers, squeezing ever more resolution through clever tricks like immersion lithography (putting water between the lens and wafer) and multi-patterning (exposing the same wafer multiple times with slightly shifted masks).

But multi-patterning is expensive and error-prone. Each additional exposure doubles the number of masks, increases alignment risk, and slows throughput. By the time the industry reached the 7-nanometer node, DUV multi-patterning required four exposures per layer. EUV lithography, operating at a wavelength of 13.5 nanometers — roughly fourteen times shorter than DUV — promised to print the same features in a single exposure. The problem was that generating, focusing, and controlling light at that wavelength is so difficult that the technology took nearly twenty years to commercialize.

03 How EUV Lithography Works

The engineering behind EUV is almost absurdly complex. To generate 13.5-nanometer light, the machine fires a high-power carbon dioxide laser at microscopic droplets of molten tin, falling through a vacuum chamber at 50,000 droplets per second. Each droplet is hit twice — first to flatten it, then to vaporize it into plasma that emits EUV light. The laser alone is the size of a small room. The entire optical system must operate in a vacuum because air absorbs EUV light almost instantly, and the lenses themselves cannot be glass — glass is opaque to EUV — so the system uses curved mirrors made of alternating layers of molybdenum and silicon, each with surface roughness measured in picometers.

Every component pushes the boundary of what is physically possible. The mirrors are made by Zeiss and require surface precision so fine that if one of them were scaled to the size of Germany, the largest bump would be less than one millimeter. The plasma that generates the light reaches temperatures of several hundred thousand degrees. The system consumes roughly one megawatt of power continuously during operation. And the resulting EUV light is so weak by the time it reaches the wafer — after bouncing off a dozen mirrors, each absorbing about 30 percent of the light — that the photoresist must be extraordinarily sensitive to capture it.

04 The Supply Chain Bottleneck

The supply chain for an EUV scanner spans multiple companies across at least four countries. ASML designs and assembles the final machine in Veldhoven, Netherlands. Zeiss in Oberkochen, Germany, manufactures the mirrors and optical systems. The laser is produced by Trumpf in Stuttgart. The light source assembly comes from Cymer, a San Diego company ASML acquired in 2013. The wafer stage — the precision platform that moves the silicon wafer during exposure — is built by ASML itself but relies on components from suppliers across Europe and the United States. A single scanner contains roughly 100,000 parts and takes approximately 18 months to assemble after the order is placed.

This concentration creates enormous geopolitical risk. The Netherlands government, pressured by the United States, has restricted ASML from exporting its most advanced EUV systems to China. The first High-NA EUV scanners, which improve resolution further by increasing the numerical aperture of the lens system from 0.33 to 0.55, began shipping to Intel in late 2023. China, cut off from EUV entirely, has been attempting to develop domestic alternatives but remains years behind, forcing Chinese foundries to rely on older DUV technology that caps their process capability at roughly 7nm — the same node where EUV becomes essential.

05 Why AI Chips Need EUV

The connection between EUV lithography and artificial intelligence is direct and unavoidable. Modern AI accelerators like Nvidia's H100 and Blackwell GPUs contain tens of billions of transistors packed into die areas of 800 square millimeters or more. The Hopper architecture uses a custom 4nm process from TSMC, which requires EUV lithography for its most critical layers. Without EUV, the chip would need to be significantly larger (reducing yield and increasing cost) or significantly less capable (reducing the performance that makes it worth buying). Every generation of AI improvement — from GPT-3 to GPT-4 to GPT-5 and beyond — depends on packing more transistors into the same silicon area, and that packing density depends on the resolution of the lithography system.

The economics are brutal. A single Nvidia H100 GPU costs roughly $25,000 to $40,000. Training a large language model like GPT-5 can require tens of thousands of these GPUs running for months, with total compute costs measured in the hundreds of millions of dollars. The entire pipeline — from the tin droplets in an ASML scanner to the trained neural network in a data center — is a single industrial chain, and EUV lithography is its weakest link. If ASML cannot produce enough scanners, or if a geopolitical crisis disrupts the Zeiss-ASML-Trumpf-Cymer supply chain, the pace of AI progress slows.

06 The Geopolitics of Extreme Light

The export controls on EUV equipment represent one of the most significant technology restrictions in modern history. The United States has progressively tightened restrictions on semiconductor technology transfers to China, culminating in rules that not only ban the sale of advanced lithography equipment but also restrict the export of advanced AI chips themselves. The logic is straightforward: whoever controls the lithography controls the chips, and whoever controls the chips controls the frontier of artificial intelligence. China has responded with massive domestic investment, reportedly committing over $100 billion to its semiconductor industry, but the gap in EUV capability remains years wide.

Meanwhile, the race for the next generation — High-NA EUV — is already underway. ASML shipped its first High-NA system (the EXE:5000) to Intel in late 2023, with a price tag reportedly exceeding $350 million. High-NA systems can print features down to 2nm and below, enabling the chip generations that will power AI models through the late 2020s. TSMC and Samsung have also placed orders, but Intel's early commitment gave it a potential first-mover advantage in the next process node race — a race where the prize is nothing less than the physical foundation of the AI economy.

07 What Comes Next: High-NA and Beyond

The roadmap beyond High-NA EUV is less clear. ASML's current High-NA systems (numerical aperture 0.55) are expected to carry the industry through the 2nm and 1.4nm nodes. Beyond that, the industry faces a choice between further increasing numerical aperture (hyper-NA, potentially 0.75 or higher) or moving to entirely new patterning technologies. Directed self-assembly, nanoimprint lithography, and even electron beam lithography have been explored, but none has demonstrated the throughput required for high-volume manufacturing. The most likely path is incremental: continued improvements to High-NA EUV, new photoresist chemistries, and advanced computational lithography that uses machine learning to optimize mask designs.

The irony is that the machine that prints AI chips is itself becoming an AI-powered system. ASML has begun using machine learning to optimize exposure parameters, predict maintenance needs, and correct for alignment errors in real time. The lithography scanner — the most precisely engineered machine in human history — is now learning from itself, using the same class of technology it exists to manufacture. This feedback loop between AI and its own manufacturing infrastructure may be the most important technological convergence of the 2020s, and it all starts with 13.5 nanometers of light.