The AI Kill Switch: Europe Discovers the Cost of Dependence

Europe’s shock at American restrictions on frontier AI is strategically naïve. The United States does not need to view Europe as an enemy to prevent it from gaining independent control of transformative AI. It only needs to recognize that any foreign bloc with frontier or AGI-level capability could become a strategic peer. America may cooperate with Europe, sell it advanced services, and grant privileged access—but it will likely retain control of the models, chips, cloud infrastructure, and systems that make that access possible. The issue is not hostility, it’s power.

A pair of European news programs reacted with visible alarm after Washington restricted foreign access to Anthropic’s most advanced systems. The central question they posed was whether the United States had begun “weaponizing” access to AI, but that framing begins too late. The important question I have is why European governments ever assumed that the United States would permit another major power bloc to obtain independent control over a potentially decisive technology. Europe mistook alliance for guaranteed access.

Many in Europe were reportedly shocked that Washington could abruptly restrict access to frontier models, and that surprise exposes a deeper analytical failure. American export controls on advanced chips have existed for years, and those controls have progressively expanded toward model weights, cloud-computing services, and other elements of the AI stack. The Anthropic decision was not a radical departure, it was a basic logical continuation of an existing strategic doctrine.

The US may distinguish Europe from China diplomatically, but it may not distinguish them completely when the issue is control over AGI or another civilization-shaping capability. Washington wouldn’t willingly allow China to reach AGI first and there is little reason to believe it would likewise allow Europe to do so either.

Europe remains an American ally, but alliances do not erase strategic competition. A Europe possessing independent frontier AI could:

• establish its own defense and intelligence architecture;
• control its own data and cloud systems;
• shape international AI standards;
• compete with American firms;
• reduce dependence on American security guarantees;
• bargain independently with China, India, Russia, and other powers;
• become a genuine technological pole rather than a protected market.

A friendly peer is still a peer, and no dominant state voluntarily creates an independent rival in the most consequential technology of the age.

I made this argument in “The Deterrence Paradox: AGI Parity and the War No One Can Pause.” That paper argued that the pursuit of AGI is governed by first-mover fear. States cannot safely pause because they cannot verify that rivals have paused, just as they cannot tolerate another power acquiring a decisive capability first. That same logic applies inside alliances, the US may share benefits, but it will resist ceding control over the frontier itself.

Europe may be receive:

• model access;
• API access;
• cloud services;
• security partnerships;
• approved deployments;
• limited research collaboration.

But Europe is far less likely to receive:

• unrestricted frontier model weights;
• independent control of the full compute stack;
• freedom from American export controls;
• guaranteed access during a political dispute;
• the ability to develop strategic parity without resistance.

Access is not ownership, just like service is not sovereignty and partnership is not control.

The interviews describe American technological power as a “kill switch” but the term is not merely rhetorical. The United States can exercise influence through:

• payment networks;
• cloud platforms;
• enterprise software;
• email and communications systems;
• AI APIs;
• chips and accelerators;
• proprietary software ecosystems;
• export licensing;
• sanctions;
• corporate compliance orders.

The France 24 discussion specifically cited the targeted denial of US-based services to a sanctioned international-court judge as an example of precise, individual-level application. This mechanism can operate at several scales:

1. against a person;
2. against a company;
3. against an institution;
4. against an industry;
5. against an entire country or bloc.

Existing infrastructure already permits precision coercion, and AI simply makes that infrastructure more consequential.

A glaring example of this structural dependence is Europe’s reliance on Palantir for intelligence work. France’s domestic intelligence service reportedly relied on Palantir for roughly a decade following the 2015 terrorist attacks. France is now attempting to transition toward a domestic alternative, while French health-data infrastructure has also been moved away from Microsoft Azure toward a French provider. This is not ordinary software procurement. Intelligence platforms organize:

• data ingestion;
• entity relationships;
• investigative workflows;
• targeting;
• threat analysis;
• operational memory;
• institutional decision-making.

When a foreign company supplies that architecture the dependence extends beyond software, and it reaches into how the state perceives threats and acts upon them. Europe did not only outsource data storage, it outsourced portions of its sovereign nervous system.

Images emerging from the G7 revealed the shift in control dynamics: AI executives were not peripheral guests, but appeared as political power brokers. European leaders were seeking access to the companies controlling the systems on which their hospitals, governments, banks, intelligence agencies, and research institutions increasingly depend. The old hierarchy has inverted: technology companies that once sought meetings with governments are now answering calls from governments. This produces a three-part power structure:

• the American state controls legal and export authority;
• American firms control much of the technical infrastructure;
• foreign governments depend on both.

Europe believed regulation itself could constitute power. It developed needed privacy rules, platform rules, AI rules, and a belief in the “Brussels effect” to that end. But regulation controls systems only when the regulator possesses leverage over the underlying infrastructure. Europe regulates American technology while remaining dependent on:

• American cloud providers;
• American chips;
• American software;
• American model companies;
• American payment systems;
• American security platforms.

Rules can restrain market behavior, but they cannot substitute for industrial capacity. Values without capability become requests.

Europe plans AI factories, larger data-center capacity, open systems, and domestic technology development, but even European-based infrastructure may continue to run on American Nvidia hardware and the proprietary Compute Unified Device Architecture (CUDA) software ecosystem. That reliance risks changing the visible supplier without removing the actual strategic dependency. Europe might reduce a reliance on Microsoft or Amazon while increasing reliance on Nvidia. A European-branded AI system is not sovereign if its critical chips, software, updates, cloud interfaces, or licensing remain externally controlled.

Europe can reduce its vulnerability, but it probably can’t eliminate it quickly. The continent lacks equivalent scale in:

• hyperscale cloud;
• frontier-model investment;
• advanced accelerators;
• software ecosystems;
• venture funding;
• integrated defense technology;
• energy and compute deployment.

The DW discussion estimated that Europe depends externally on roughly 80 percent of its technology stack, with cloud dependence exceeding 90 percent in certain sectors. Europe can build specialized systems, diversify suppliers, and preserve critical domestic capabilities. But catching the United States at the frontier is a different proposition.

The likely American objective is not to deny Europe all AI, that would damage American companies and weaken the alliance. The more likely strategy is controlled dependence, Europe may be allowed substantial access, but under conditions Washington can alter. The United States will likely seek to control:

• who receives frontier access;
• which models can be exported;
• where advanced chips can be deployed;
• which cloud systems may train advanced models;
• which allies receive exceptions;
• when access can be suspended;
• how close another bloc may approach strategic parity.

Crucially, this is not necessarily a temporary Trump policy, it reflects structural incentives that would confront almost any American administration.

Europe is not China, it shares military alliances, democratic traditions, intelligence relationships, and extensive economic integration with the United States. That relationship should produce greater trust and broader access.

…but greater access is not equal control. The United States can treat Europe more favorably than China while still preventing Europe from acquiring independent strategic parity. There is a large space between exclusion and sovereignty and Europe is likely to remain inside that gap. I do not mean to affront anyone by slapping the proverbial croissant out of a Frenchman’s mouth, but there are likely to be more than pastry casualties as governments jockey for position.

Europe is asking whether America has begun weaponizing AI access, but the answer is that access to decisive technology was always political power. Europe simply failed to treat it that way. The Anthropic restriction didn’t create Europe’s vulnerability, it exposed it. Palantir dependence, cloud concentration, Nvidia stack, payment networks, and targeted kill switches all point toward the same conclusion: Europe built critical institutions on infrastructure controlled elsewhere, and now it’s discovering that the owner of the infrastructure retains the final vote.

America may continue to treat Europe as an ally, but it is unlikely to permit Europe to become an uncontrolled AI peer. This should not be seen as betrayal, it’s deterrence logic reaching the alliance system.

Europe was never locked out of the room, it simply discovered that Washington still has the key to the door.

Sources

1. France 24 YouTube debate. https://www.youtube.com/watch?v=zMKW-nEv2H0
2. DW News interview. https://www.youtube.com/watch?v=NlnzSPHTZsM
3. “The Deterrence Paradox: AGI Parity and the War No One Can Pause.” btne.net/the-deterrence
4. Primary reporting on the Anthropic export restriction. https://www.anthropic.com/news/fable-mythos-access
5. Documentation on France’s DGSI and Palantir. https://www.reuters.com/technology/france-invest-655-mln-ai-set-up-common-chatbot-all-state-services-2026-06-16/
6. Primary reporting on sanctions and service denial against the international-court judge. https://www.pbs.org/newshour/world/trumps-sanctions-on-iccs-chief-prosecutor-have-halted-tribunals-work-officials-and-lawyers-say
7. EU AI-factory and Nvidia/CUDA dependency sources. https://www.itpro.com/technology/artificial-intelligence/hpe-and-nvidia-launch-first-eu-ai-factory-lab-in-france
8. European cloud-market concentration data.https://www.europarl.europa.eu/RegData/etudes/STUD/2025/778576/ECTI_STU%282025%29778576_EN.pdf