A hardware choke‑point for the AI arms race
There is an industrial fact most coverage skips: the most advanced chips that power frontier AI require EUV lithography tools. ASML, the Dutch firm in Veldhoven, builds these enormous machines. They use 13.5 nm EUV light and cost hundreds of millions of dollars each. Those tools are indispensable for producing the leading‑edge nodes used in the largest AI training runs. That makes the hardware pipeline, not just software or models, one of the few chokepoints that can actually slow the race.
Concern about a runaway AI arms race has moved from intellectuals and think‑tanks into mainstream political debate. Bill Gates wrote, “if someone had a credible plan for slowing down AI advances globally, I would likely support it.” Anthropic publicly asked for a “brake pedal” (New York Times, 5 June 2026). Senator Bernie Sanders has urged a slower, more regulated approach (10 August 2026). The question for policymakers and executives is simple: who has the leverage to make a slowdown stick, and what would that look like?
Why voluntary restraint is a weak tool
The incentives to race are brutal. Firms and states follow a “use it or lose it” logic: first movers grab market share, monopoly rents, and geopolitical advantage. Investors reward growth and scale. With billions or trillions of dollars in infrastructure and national prestige at stake, voluntary moratoria are fragile.
Industry forecasts reported by the Financial Times expect roughly $7 trillion of datacentre construction by 2030. Foundries are pouring capital into capacity, and TSMC has cited plans of about $265 billion for future fabs. When that scale of investment and national interest exists, asking firms to pause is a hard sell.
How ASML became geopolitical leverage
Reuters reported on 14 May 2026 that U.S. export‑control policy has helped shape restrictions on ASML sales to China. That precedent shows two things: these machines matter strategically, and export controls are a real policy lever.
Restricted sales of EUV tools would reduce the supply of the most advanced wafers. The chain is simple: fewer EUV systems → fewer leading‑edge chips → fewer ultra‑large training runs. That isn’t a perfect throttle. Older nodes, alternative architectures, and software improvements will keep much AI work moving, but restricting EUV would materially slow how fast compute capacity for the largest models grows. That slowdown buys regulators breathing room to build rules, audits, and oversight mechanisms.
“At this point, it should be clear that AI is a choose‑your‑own‑adventure book where every option ends in disaster.”, Alexander Hurst, Guardian Europe (from Paris)
It’s not a single‑button solution
ASML’s machines are a powerful lever, but they are not the only one. Advanced chips and AI deployments rely on an ecosystem: GPU and accelerator design (NVIDIA and others), memory suppliers (HBM), advanced packaging and test services, electronic‑design‑automation (EDA) tools from Cadence/Synopsys, foundries such as TSMC and Samsung, mask‑making, and global logistics. Any effective strategy must recognise that constraining one node shifts pressure elsewhere, and rivals will try to substitute, replicate, or circumvent.
Replication is possible but time‑consuming. Building domestic EUV capability, or finding alternative routes to comparable compute density, takes years not weeks and usually needs sustained state support, supply‑chain development, and specialist know‑how. That gives a window for policy action, but not an indefinite one.
What Europe can plausibly do, and why it might
Practically, the Dutch government (acting in concert with EU partners where politically viable) can impose tighter export controls, apply stricter end‑use checks, and condition approvals for ASML systems. Framed correctly, Europe’s choice could be presented as both safety and self‑interest. The Europe 2031 project (20 June 2026) warned that Europe must either regulate heavily or invest heavily to avoid strategic dependency in AI. Conditioning access to critical hardware can be sold domestically as defending industrial sovereignty, protecting energy and public‑service infrastructure, and buying regulatory time.
That leverage could also be turned into bargaining power. Conditional access for hardware might be exchanged for international commitments on third‑party audits, transparency about large‑scale model training, or verifiable safety standards.
The brittle edges: backlash, costs and enforcement problems
Any choke‑point strategy carries real costs and visible risks.
- Geopolitical retaliation. The United States could respond by restricting components or software used in ASML machines. Supply chains are interdependent. A hard split could accelerate decoupling and fragmentation.
- Industrial and civilian damage. Slower deployment of advanced chips would affect legitimate civilian uses, medical imaging, climate modelling, scientific compute, and disrupt suppliers and customers across the semiconductor ecosystem.
- Workarounds and illicit trade. Black‑market routes, engineering substitutes, and accelerated domestic programmes (especially in China) would emerge. Controls buy time; they do not permanently stop the technology.
- Legal and governance complexity. Any Dutch or EU restriction must survive trade challenges, corporate governance constraints, and international legal scrutiny. Practical enforcement requires licensing regimes, export monitoring, supplier audits and co‑operation from allied jurisdictions.
Those trade‑offs mean the choke‑point is a blunt, tactical instrument. It can slow the largest, most dangerous training runs, but it is poor at distinguishing benign innovation from risky scale‑ups.
A pragmatic blend: targeted restraint plus strategic investment
Policymakers have three complementary levers that can be used together rather than in isolation:
- Targeted export controls. Condition the most sensitive EUV transfers on stringent end‑use and verification measures, while permitting lower‑risk sales to avoid strangling civilian supply chains.
- Parallel investment in safe capabilities. Deploy the Europe 2031 playbook: fund sovereign chip capacity, safety research, public model benches, and regulatory institutions so Europe isn’t merely constraining others but building secure alternatives.
- Multilateral safety agreements. Use conditional hardware access as leverage to seek binding safety commitments: third‑party audits, reporting requirements, and agreed red‑teaming protocols that are verifiable across borders.
The goal is explicit: buy measurable time for governance while preserving Europe’s ability to compete and innovate. That requires upfront political will and money. A restraint strategy without investment would simply outsource innovation to actors with weaker oversight.
Questions leaders should be asking, and immediate actions
-
Can Europe slow AI progress meaningfully by controlling ASML exports?
Yes, tighter Dutch/EU export controls on advanced EUV systems would materially reduce the supply of the most advanced wafers for a period. Action: commission scenario modelling that estimates reduced top‑node wafer output under defined control levels and timelines (quarters/years).
-
Would such a pause be globally enforceable?
Not perfectly. Controls can be evaded and will prompt countermeasures. Action: mandate a joint US‑EU legal and enforcement assessment that outlines licensing, supplier audits, and penalties, and builds allied co‑operation on export enforcement.
-
Is slowing compute the same as preventing catastrophic AI outcomes?
No. Slowing growth buys time for regulation and oversight but does not remove long‑term risk. Action: invest in public safety testbeds, model transparency standards, and independent audit labs during the slowdown window.
-
What are the alternatives to a choke‑point strategy?
Massive European investment in safe AI and sovereign capabilities (the Europe 2031 route), expensive and slower to deliver immediate restraint but stronger for long‑term autonomy. Action: publish a phased investment plan that pairs capacity building with regulatory milestones.
Tactical scenarios, a short roadmap
Think in three plausible outcomes and plan accordingly:
- Best realistic case. Coordinated allied controls slow frontier training for several years, and those years are used to create durable international safety standards and shared verification systems.
- Medium case. A multi‑year slowdown is followed by partial catch‑up as rivals accelerate domestic programmes; Europe gains regulatory headroom but pays economic costs.
- Worst case. Controls provoke a geopolitical split, accelerate decoupling, and lead to faster, clandestine efforts to bypass constraints, diminishing global co‑operation on safety.
Design policy to maximise the chance of the best realistic case: combine targeted controls with concrete safety investments, allied legal frameworks, and transparent metrics for when and how restrictions are eased.
What matters now
Hardware is an imperfect choke‑point, but it is one of the few levers that can actually slow the pace of frontier AI. That forces a strategic choice: use the window to build governance and safe capacity, or accept a fast‑moving landscape that hands more power to whoever can amass compute fastest. Either path carries costs, economic, diplomatic, or safety related. The prudent route for Europe is to pair selective restraint with heavy investment in safety and sovereign capability, and to measure success with clear, time‑bound indicators rather than rhetoric.
Decisions about ASML, export controls and investment are not just industrial policy. They are choices about what kind of technological future Europe wants to help shape and what kind it is willing to risk.