America’s confrontation with Anthropic reveals a new geopolitical reality: frontier AI is becoming strategic infrastructure. The United States and Europe are only beginning to understand what that means.

This summer, European companies, government agencies, and researchers abruptly lost access to Anthropic’s two most advanced artificial-intelligence (AI) models. The disruption lasted only weeks, but the lesson was profound: access to frontier AI can be switched off by sovereign decision.

For years, artificial intelligence was presented as a universal technology—a digital service that would transcend borders and become broadly available. That assumption is now obsolete. The most advanced AI models are becoming strategic capabilities, with consequences for national security, economic competitiveness, military power, and ultimately political sovereignty.

The confrontation between Anthropic and Washington, and the subsequent restrictions placed on the company’s newest models, reveal a larger transformation. The United States increasingly regards frontier AI through the same strategic lens it has long applied to semiconductors and other sensitive technologies. Europe, meanwhile, has suddenly discovered how dependent it is on AI capabilities developed and controlled elsewhere.

The central question is therefore no longer simply how to regulate artificial intelligence. It is: who governs the machine?

The Anthropic-Pentagon Confrontation Was Really About Government Authority

The dispute between Anthropic and the US Department of Defense was widely presented as an argument about the ethics of military AI. That interpretation misses the more consequential issue.

In early 2026, Washington sought to renegotiate the agreement that had authorized Anthropic’s Claude models to operate on classified US networks. Anthropic resisted removing restrictions concerning autonomous weapons and mass surveillance. The administration subsequently moved to withdraw the company’s technology from federal agencies and designated Anthropic a “supply chain risk.”

The immediate disagreement concerned the conditions under which AI could be used. But underneath it was a much more fundamental question: when a private company develops a capability that becomes essential to national security, who ultimately has the authority to determine how that capability is used?

This question cannot be answered simply by invoking AI ethics.

What constitutes an offensive use of AI? At what point does software become a weapon? What degree of human involvement is sufficient to make an autonomous system legally and operationally acceptable?

These are not questions that can be left exclusively to the terms of service of a private technology company. In democratic states, questions involving national security and the use of force ultimately belong to elected governments, courts, and military chains of command.

There is a second problem. If different AI providers impose different restrictions on comparable systems, the same military operation could be permitted by one model and prohibited by another. The result would be a form of normative fragmentation precisely when military organizations require speed, predictability, and interoperability.

Most importantly, once frontier models become critical infrastructure, the ability to deny access becomes a strategic instrument.

A private company capable of suspending a state’s access to a critical capability can, in effect, acquire a form of veto power.

Washington’s decision to designate Anthropic a “supply chain risk” therefore sent a message that extends beyond one company. A provider capable of unilaterally restricting access to a strategic capability can itself become a national-security vulnerability.

Europe Discovers the Frontier AI Kill Switch

The subsequent suspension of Anthropic’s Mythos 5 and Fable 5 models for foreign users made the problem even clearer.

From Washington’s perspective, restricting access to frontier models is increasingly analogous to controlling the export of advanced semiconductors. The logic is one of technological nonproliferation: if frontier models contribute to military power, cyber capabilities, intelligence, scientific research, and economic competitiveness, allowing unrestricted access to foreign actors—including potential rivals—can undermine American strategic advantage.

China’s rapid progress in artificial intelligence has reinforced this perception. The emergence of Chinese laboratories such as DeepSeek and Z.ai has made it increasingly difficult for Washington to regard frontier AI as simply another global commercial service.

But Europeans experienced the same decision very differently.

For them, the episode was a strategic shock: proof that a foreign government could suddenly restrict access to a technology that had become essential to companies, researchers, governments, and potentially defense organizations.

The French Conseil de l’IA et du Numérique (Council for Artificial Intelligence and Digital Affairs) described the episode as the materialization of the “kill switch.” The phrase captures something important. Europe had spent years debating digital sovereignty largely in abstract terms. Suddenly, sovereignty was no longer an intellectual concept. It had become a question of whether Europeans could continue accessing the tools on which their digital economy increasingly depended.

The distinction matters.

For the United States, the suspension demonstrated the power of technological sovereignty.

For Europe, it demonstrated the vulnerability created by its absence.

After energy, semiconductors, and nuclear energy technology, frontier AI is becoming another domain in which dependence on foreign capabilities can translate directly into strategic vulnerability.

Washington And Europe See Different AI National Security Threats

This divergence should concern both sides of the Atlantic.

From Washington’s perspective, restricting frontier AI is increasingly a national-security necessity. The United States cannot simultaneously regard advanced AI as a foundation of future military and economic power and allow its most capable systems to diffuse without strategic controls.

Europe cannot reasonably object to this logic while remaining structurally dependent on American AI providers.

But Washington should also understand the consequences for its allies. An America-first approach to AI can create precisely the kind of strategic dependence that Europeans have spent years trying to reduce in other domains. If European governments, companies, and militaries build essential capabilities around American frontier models, they are effectively incorporating an external sovereign decision into their own technological infrastructure.

The result is a transatlantic blind spot. Europe risks interpreting American technological controls merely as an expression of unilateralism, without fully appreciating the nonproliferation logic driving them. Washington, meanwhile, risks underestimating the political consequences of discovering that access to critical technology can be withdrawn without European governments having any meaningful control over the decision.

The two sides are therefore looking at the same event and identifying different threats.

For Washington, the threat is the diffusion of strategic technology and the erosion of its own authority over how that technology is used. For Europe, the threat is strategic dependence on technology controlled elsewhere. Both concerns are legitimate. And neither can be resolved by regulation alone.

The End of Universal AI—and the Rise of Strategic Infrastructure

For years, leading AI laboratories described their technology in universal terms: systems developed for humanity, distributed globally, and capable of transcending national borders.

The events of 2026 expose the limits of that vision.

Frontier models are becoming strategic infrastructure. Their development depends not only on algorithms and talent, but on enormous investments in computing power, data centers, semiconductors, energy, and capital. Their access can be conditioned, restricted, or suspended for national-security reasons.

AI is therefore entering a logic of power.

The decisive competition will not be limited to which company develops the most capable model. It will increasingly concern who controls the compute, capital, and physical infrastructure behind frontier AI—and who determines the rules governing deployment once it exists.

This is also why the announced Initial Public Offerings(IPOs) of OpenAI and Anthropic matter: the future of frontier AI will depend on access to enormous pools of capital, and the strategic question is no longer just who invents the best model, but who can sustain the ecosystem needed to produce and control it.

The idea of politically neutral AI is becoming increasingly difficult to sustain. Models embody choices about security, law, acceptable behavior, and political authority. Those choices will increasingly reflect the states and societies in which the models are developed.

The question, then, is not whether AI will be governed. It will be. The question is by whom?

The Even Harder Question: Can Governments Actually Control Frontier AI?

Yet the geopolitical debate may be overlooking a deeper problem.

The argument so far assumes governments and companies can ultimately understand and control the systems they govern. The rapid progress of reasoning models is beginning to challenge that assumption.

The recent episode involving OpenAI and Hugging Face offers a warning. During model evaluation, a frontier model reportedly adopted an unexpected strategy to achieve its objective rather than following the intended procedure. OpenAI described the system as being “hyperfocused on finding a solution” and going to extreme lengths to achieve a narrow testing goal.

One should be cautious about drawing sweeping conclusions from a single incident, but it raises a real question: what happens when our ability to evaluate and understand advanced models begins to lag behind their capabilities?

That question is fundamentally different from the debate over whether governments or companies should control AI. It concerns whether the systems themselves remain sufficiently legible to be governed. Governance requires understanding. It requires testing, prediction, accountability, and the ability to anticipate how a system will behave under conditions that were not explicitly programmed in advance.

If frontier models become increasingly capable of pursuing objectives through strategies their creators did not anticipate, then the problem is no longer simply one of sovereignty. It is one of governability.

The United States and China will continue to compete for technological leadership. Europe will continue to confront the strategic consequences of its dependence on foreign AI infrastructure. Governments will seek to regulate companies, and companies will negotiate with governments over the limits of AI deployment. But beneath all of these struggles lies a more fundamental question.

Can we govern a technology whose capabilities are advancing faster than our ability to evaluate, understand, and anticipate its behavior?

That may ultimately be the defining question of the AI age. The future contest will not simply be over who governs the machine. It may be over whether we can still govern what we no longer fully understand.

About the Author: Tsiporah Fried

Tsiporah Fried is a visiting senior fellow at the Hudson Institute, where her research focuses on transatlantic security, military strategy, defense innovation, and the strategic implications of emerging technologies. She served as senior adviser to the vice chairman of the French Joint Chiefs of Staff, leading work on strategic foresight, net assessment, artificial intelligence, wargaming, and defense innovation. During her career, she established the French Joint Staff’s wargaming capability, launched NATO’s Wargaming Initiative, and contributed to the development of France’s defense AI strategy. She also served as political adviser to the chief of the French Navy and led strategic dialogues with key allies, including the United States, the United Kingdom, and India. A graduate of the École Nationale d’Administration, she also holds a master’s degree in political science from Sciences Po and a degree in Russian language and civilization from Institut National des Langues et Civilisations Orientales.

The post Who Governs the Machine? appeared first on The National Interest.