WHO CONTROLS THE MACHINES THAT MAKE AI POSSIBLE?

Semiconductor manufacturing begins with design and ends in fabrication. A company like Nvidia designs processors: it specifies the transistor-level layout of a chip using electronic design tools, and hands that design to a manufacturer. The manufacturer — a foundry such as TSMC — turns the design into physical chips inside fabrication plants of extraordinary complexity. Neither step is the other. A design without fabrication is a file; fabrication without equipment is an empty building.

Inside the fab, the critical step is lithography: projecting a chip's circuit pattern onto a silicon wafer coated with light-sensitive material, layer after layer. Deep ultraviolet (DUV) lithography, the workhorse technology of the industry, uses light with wavelengths around 193 nanometers. Extreme ultraviolet (EUV) lithography uses light with a wavelength of just 13.5 nanometers — approaching the X-ray range — and prints features that DUV systems cannot. ASML describes EUV as a technology unique to the company: a system that vaporizes droplets of tin with CO2 laser pulses up to fifty thousand times per second to generate the light, guides it through a vacuum chamber using multilayer mirrors rather than lenses, and positions the wafer to within a quarter of a nanometer for each exposure.

Two clarifications anchor everything that follows. First, ASML does not manufacture Nvidia's processors and does not operate TSMC's foundries: it builds the equipment that manufacturers use. Second, EUV systems print the most intricate layers of a chip; the remaining layers are printed with DUV systems, which ASML supplies alongside competitors. The bottleneck examined in this investigation is specific: it is the most advanced end of lithography, where ASML's EUV position is unique.

ASML is the only company that manufactures and sells EUV lithography systems for commercial chip production. This position was not created overnight. ASML states that it invested more than six billion euros in EUV research and development over seventeen years, and that it acquired Cymer — a light-source specialist — to accelerate development. The first pre-production EUV system shipped in 2010; the first production system, the TWINSCAN NXE:3300, shipped in 2013; the hundredth EUV system shipped in early 2020.

The current generations define the frontier. NXE systems, with a numerical aperture of 0.33, print the foundation layers of 7 nm, 5 nm and 3 nm nodes and now support high-volume 2 nm logic production. EXE systems — High NA EUV, with a numerical aperture of 0.55 — print features with a resolution of 8 nm; the first system was delivered in December 2023. ASML reports annual research and development spending of 4.7 billion euros in its 2025 Annual Report, on total net sales of 32.7 billion euros.

The barriers to entry are not primarily financial. They are accumulated: decades of process knowledge, an installed base of expertise in optics, light sources, metrology and software, and service relationships that tie each installed system to its manufacturer for its operating life. Rivals in DUV lithography — Canon and Nikon among them — have not brought a commercial EUV system to market. The consequence is structural: at the most advanced nodes, there is no second source.

An EUV system is assembled from components that themselves represent decades of specialized engineering. The optical system — the multilayer mirrors that reflect 13.5 nm light with near-total efficiency — is produced by Carl Zeiss SMT in Oberkochen, Germany. The high-power CO2 drive lasers come from TRUMPF, also of Germany. The light-source technology traces to Cymer of San Diego, acquired by ASML in 2013. Around these core partners stand hundreds of specialized suppliers in precision engineering, advanced materials, metrology, software and service.

ASML's relationship with Carl Zeiss SMT is structural rather than transactional: ASML holds a minority equity stake in the optics operation and co-invests in its development capacity — an arrangement announced in 2016 and deepened since. This matters for sovereignty analysis: the most critical component of the most critical machine is made by a German company, integrated by a Dutch company, using a light-source lineage that runs through a California company — for foundries in Taiwan, South Korea and the United States.

The lesson DSC draws is that manufacturing sovereignty cannot be reduced to one corporation or one country. A nation cannot secure its access to advanced chips merely by courting one supplier. The dependencies run through an international network in which no participant controls the whole — and in which each participant can become a point of leverage.

The industry can be described as four layers. Equipment makers design and build fabrication tools. Foundries operate the fabs that use them. Designers specify the chips. AI infrastructure operators deploy the resulting systems at scale. ASML sits in the first layer. TSMC, Samsung and Intel operate foundries. Nvidia, AMD and Broadcom are designers. The largest cloud and AI operators — hyperscalers — depend on all three lower layers.

Two kinds of dependency run between these layers, and they must not be confused. A technical dependency exists when a capability exists nowhere else: no foundry can produce the most advanced nodes without EUV tools, and no equipment maker other than ASML produces them. A commercial relationship exists when two companies choose to do business: contracts, orders, partnerships. The first is structural; the second can change. Public discussion often conflates them.

The diagram below is constructed by DSC from publicly disclosed roles and company statements. It does not assert undisclosed contracts, and it should be read as a map of publicly known positions — not a record of private agreements.

Access to EUV and advanced DUV equipment is governed less by markets than by licensing. ASML is a Dutch company; the Dutch government licenses its most advanced exports. The United States, through the Bureau of Industry and Security, controls a broader category of advanced-computing technology — and because key subcomponents are American, its rules reach tools assembled in Europe. The two regimes have interacted since at least 2022, when the United States imposed sweeping export controls on advanced computing.

The record assembled below distinguishes the decisions that are governmental, the company statements that respond to them, and the coordination that is reported but not fully documented in public. Where a development rests on press reporting rather than a primary filing, it is labeled accordingly.

The practical effect is that access to the machines is a matter of state policy — exercised through license grants, refusals, and revocations — long before it is a matter of price or capacity.

China's domestic semiconductor program is the largest state-coordinated effort to substitute for controlled foreign technology. Its lithography ambitions are explicit; its published achievements are harder to verify. Domestic lithography development — led by SMEE and others — has demonstrated DUV-class capability in principle, but claims of advanced-node production using domestic tools rest on limited public evidence and should be treated with care.

Independent teardowns and industry reporting have described advanced logic production at Chinese foundries using varying tool mixes; the provenance of specific equipment remains contested. This investigation distinguishes three categories: verified technical developments, ambitious announcements, and unconfirmed claims. At present, the most consequential verified fact is negative: no EUV system has been shipped to China under license.

Export controls produce at least three simultaneous effects, and honest analysis must hold all three: they constrain the controlled party's capabilities; they accelerate domestic substitution programs by guaranteeing demand for alternatives; and they carry geopolitical consequences that extend beyond the semiconductor industry. A policy that succeeds on the first measure may still lose ground on the other two.

TSMC's share of the world's most advanced logic manufacturing has been widely estimated at around ninety percent — a figure that originates in industry analysis rather than company disclosure, and is labeled accordingly. Whatever the precise number, the structural fact is not contested: the leading edge of chip production is concentrated in facilities that took decades and tens of billions of dollars to build.

The major foundries have begun to diversify — fabrication capacity in Arizona, Japan and Germany is under construction or in production, with government subsidy programs supporting each. But leading-edge capacity remains concentrated, and the equipment that defines it is concentrated one layer deeper, in a single supplier. Diversification of fabs does not, by itself, diversify the tools inside them.

This investigation does not present speculative conflict scenarios as inevitable, and it treats claims about specific contingencies as outside its evidentiary scope. The dependency itself — geographic concentration at the manufacturing layer — is the documented condition; planning for it is a matter of industrial resilience and continuity, which is why capacity diversification, maintenance access, and stocks of tool parts appear in national security planning discussions.

Semiconductor capital expenditure runs to the scale of national infrastructure budgets, and the equipment layer captures a structural share of it. ASML's own financial record shows the scale: total net sales of 32.7 billion euros in 2025, with third-quarter 2026 sales of 11.7 billion euros and full-year 2026 guidance between 44 and 45 billion euros.

Governments have responded to concentration with subsidy. The United States CHIPS and Science Act, the European Chips Act, and comparable programs in Japan, South Korea and elsewhere direct public money toward fabrication capacity — in effect, toward the equipment that fabrication capacity buys. Where public money flows to private concentration, accountability questions follow it.

A lithography tool is not a quarterly purchase. Systems are configured, installed and serviced over years, and production lines depend on installed equipment for a decade or more. Financial commitments made today therefore constrain technological options for years after the signatories have moved on — a dependency structure that DSC examines as part of its broader analysis of infrastructure lock-in.

Advanced semiconductors are dual-use by construction. The accelerators that serve commercial AI also serve high-performance computing for weapons design, intelligence processing, cybersecurity operations, and autonomous systems. Export-control regimes exist precisely because this boundary cannot be drawn cleanly — the controls documented in section 05 are, at bottom, an acknowledgment that commercial products feed military capability.

The accountability question is symmetrical and largely unexamined: national-security exceptions to normal commercial and oversight rules are invoked in the name of protecting advanced technology, yet those exceptions themselves concentrate power in institutions that operate with limited public visibility. DSC's separate investigations of military computational procurement examine the demand side; this investigation establishes the supply side on which it rests.

Documented applications and hypothetical risks must be kept distinct. The documented record concerns what advanced chips demonstrably enable and what controls demonstrably govern. Hypothetical future capabilities — autonomous weapons at scale, AI-directed systems — are treated in this investigation only as questions, not as established facts.

Participation in advanced computing is gated by stacked barriers. Industrial capital: a leading-edge fab costs tens of billions of dollars. Equipment access: the machines are sold under license, not on open markets. Talent: decades of accumulated process engineering cannot be hired into existence. Connectivity: the computational economy assumes broadband and power infrastructure that much of the world lacks. And investment: private capital for frontier projects concentrates in a small set of firms and geographies.

These barriers compound, but they are not identical in cause, and they do not have identical remedies. Some are matters of cost; others of policy; others of infrastructure investment that no single actor controls. Treating them as one problem produces one-note answers. DSC's access analysis — developed across the WHO CONTROLS WHOM? series — examines each layer separately before asking who bears responsibility for the combined effect.

The question this investigation adds is the layer beneath them all: a country can have capital, talent and ambition, and still be excluded — by license, by coordination between states, or by the physics-dependent economics of the machines themselves. Exclusion at that layer is not a market outcome. It is a decision, made somewhere, by someone.

DSC's Scarcity Regime Triangle describes three coupled scarcity regimes: energy, hardware, and software. Semiconductor fabrication capacity is the hardware regime's physical core — and it couples tightly to the other two. Fabs consume electricity at industrial scale and require stable water supplies; their output determines the compute available for training and deploying AI systems; and the software stack — the models, tools and platforms built on that compute — shapes demand back onto the hardware layer.

The coupling runs in both directions. Constraints in energy supply limit fabrication expansion and datacenter operation alike. Constraints in fabrication capacity limit compute, and therefore limit what software can be built. Demand for AI software drives investment in hardware, which drives demand for energy. A shortage or decision in one regime propagates through the others — which is why DSC analyzes them as one system rather than three markets.

DSC's fuller treatment of the Scarcity Regime Triangle, including its relationship to computational governance, appears in Constitution Before Capability. This investigation adds the lithography layer to that framework: the machines examined here are the point where the hardware regime is narrowest.

DSC's Singularity of Control thesis describes a condition in which concentration at successive layers of the technological stack becomes mutually reinforcing: control of compute compounds with control of the infrastructure beneath it and the platforms above it. The thesis is DSC's analytical framework. It is not an established scientific law, and this investigation does not treat it as one.

Applied to semiconductors, the framework distinguishes three things that must be kept apart. Technical concentration: one supplier of EUV equipment, one dominant center of leading-edge fabrication. Commercial market power: pricing, allocation, and the terms on which access is sold. Political control: the licensing regimes of section 05, which decide who may obtain what is sold. Each layer of concentration is documented. Whether they combine into a self-reinforcing whole is precisely the question DSC's framework poses — and the question this investigation is designed to inform, not to answer by assumption.

The distinction matters because the remedies differ. Technical concentration is addressed by engineering investment and supplier diversification. Commercial concentration by competition policy. Political control by democratic oversight. Conflating them — treating every form of concentration as every other — produces remedies that miss.

If access to the machines that make advanced AI possible is decided by governments, then the first accountability question is which governments, on what public record, with what review. Export-control decisions of the kind documented in section 05 are exercised by executive bodies; their rationales are only partially public; and the affected parties — companies, countries, and ultimately populations — have limited formal recourse.

Public subsidy raises the mirror image of the same question. When public money builds private fabrication capacity, the public acquires an interest in what that capacity produces, for whom, and on what terms. Transparency, competition, public investment with public conditions, and international cooperation are the instruments available — and each is currently exercised unevenly.

This investigation rejects the framing that national-security oversight and civil liberties are mutually exclusive, and it rejects the opposite error as well: that oversight and openness can be discussed without acknowledging genuine security constraints. The question is not whether constraints should exist. It is who decides them, on what record, with what accountability.

Digital Sovereignty Coalition's response to the condition this investigation documents is not national autarky and not private monopoly, but plurality: multiple capable actors at every layer of the stack — equipment, fabrication, design, deployment — with resilient supply chains connecting them, international oversight of the export regimes that govern access, public-interest computing capacity that does not depend on any single owner, and meaningful access to AI infrastructure for communities currently outside it.

Civic AI — computation in service of the public interest, governed by public accountability — is the principle DSC proposes for the software layer of this plurality. For the equipment layer examined here, the corresponding principle is simpler: the industrial base of the digital age should not be controllable by any single state or single corporation, because everyone's access to the future depends on it.

DSC's Civic Machina is referenced here only as a proposed civic AI architecture — a design direction, not an existing operational semiconductor or computing platform. The distinction between proposals and operating reality is maintained throughout DSC's work, and readers should hold this investigation to the same standard.

This investigation has established a documented condition: one supplier of the most advanced lithography systems; a supply network in which critical knowledge is distributed but irreplaceable; licensing regimes that determine access by state decision; and a dependency structure that persists for decades once built. Much remains unverified — pricing, market shares, and the full scope of international coordination are flagged in the source registry rather than asserted.

The closing question of this investigation is not rhetorical. It is offered for critical examination — in classrooms, newsrooms, legislatures, and boardrooms — without sensationalism, and with the same evidence discipline DSC applies across its work.