The battle for AI sovereignty

In 2025, it appeared as if the defining contest of the 21st century would pit national governments against global technology companies, with the most likely outcome being an uneasy coexistence in which each side largely remained within its own sphere. Now, that equilibrium has broken, and not only in the United States. Across major democracies, the same pattern is emerging: When artificial intelligence touches the core interests of the state – military power, geopolitical position, national security – governments will not accept private autonomy.

The new model is not regulation. It is capture, but of an unfamiliar kind. “Regulatory capture” traditionally describes firms that capture the state, bending public institutions to private interests. What is emerging here is the reverse: The state captures firms whose technology has become critical.

Enforcing a national AI agenda

On February 27, 2026, U.S. Secretary of War Pete Hegseth announced that he was directing the Department of War to designate Anthropic, one of the world’s leading AI companies, a “supply chain risk to national security.” It was the first time that the designation – previously reserved for firms with ties to adversarial governments, most notably Chinese telecommunications companies – was applied to an American company. Its trigger was not a data breach, a foreign ownership structure or evidence of espionage. It was a contract negotiation.

Anthropic had refused to allow its AI model, Claude, to be used for two specific purposes: autonomous lethal weapons systems and mass surveillance of American citizens. The Pentagon’s position was unambiguous. Mr. Hegseth insisted that a private company could not dictate the policies of the U.S. government. “America’s warfighters will never be held hostage by the ideological whims of Big Tech. This decision is final,” he declared. The company’s competitor, OpenAI, signed a competing agreement with the Pentagon soon after Anthropic’s blacklisting, accepting use for any “lawful purpose.”

The message to the AI industry was clear: You may develop frontier models, raise venture capital, compete for market share – but when the state requires your technology, the terms are set by the government.

Anthropic has since challenged the designation in two federal lawsuits backed by nearly 150 retired judges. How the courts will rule is unsettled: As of late May 2026, the Washington appeals court looks likely to uphold the designation, even as a California court temporarily blocked the ban on using Claude. But the precedent of attempted coercion has already done its work, and the courts may end up endorsing it. The industry has seen what compliance looks like, and what resistance costs.

The military proof of concept

One day after the decision to blacklist Anthropic, the U.S. and Israel launched Operation Epic Fury – a coordinated strike campaign against Iran’s nuclear infrastructure and military leadership. CENTCOM Commander Admiral Brad Cooper confirmed publicly that AI tools “help us sift through vast amounts of data in seconds” to enable faster targeting decisions. The Wall Street Journal reported that Anthropic’s Claude was used in intelligence assessments and target identification during active operation even as the company simultaneously sued the government that was using it.

Analysts now describe what is emerging as the era of “precise mass” – AI-assisted targeting combined with swarms of cheap autonomous systems that invert the economics of conflict entirely. A drone built from commercial parts costs roughly 1 percent the cost of a single interceptor missile.

Israel’s contribution illustrates the operational frontier. The Financial Times reported that Israeli forces synthesized traffic camera footage and billions of data points to track and kill Supreme Leader Ali Khamenei. This feat of AI-enabled intelligence fusion was built on years of integration across operations in Gaza and Lebanon.

Every major military establishment watching this conflict is drawing the same conclusion: AI-mature forces represent a qualitatively different order of capability, and no government will accept having those capacities constrained by the ethical policies of a private company’s board.

But the conflict also exposes a deeper structural problem that extends well beyond the U.S. The same class of general-purpose models that can power a hospital’s diagnostic system and a law firm’s contract review is identical to the one informing targeting decisions. The distinction between civilian and military AI exists as a policy category, not a technical one.

Anthropic’s red lines were an attempt to impose that distinction contractually, precisely because it cannot be enforced at the level of the technology itself. The Pentagon’s rejection of those limits is therefore not merely a procurement dispute – it is a declaration that the civilian-military boundary in AI will not be honored once a model becomes strategically consequential. Other governments are reaching the same conclusion by different routes.

The geopolitical architecture of AI

If the Anthropic case represents the enforcement mechanism and Iran demonstrates the effect, Pax Silica is the institutional architecture of AI power. Signed in Washington in December 2025 by nine nations – the U.S., the United Kingdom, Japan, South Korea, Singapore, the Netherlands, Israel, the United Arab Emirates and Australia – the framework formalizes what had previously been implicit: Access to AI infrastructure is conditional on political alignment. Chips, computing power and frontier models are strategic assets managed through alliance structures rather than open markets. Sweden joined in March 2026; India in February. The European Union is conspicuously absent.

Pax Silica is meaningfully different from the export control regimes it superficially resembles. Export controls are defensive: They restrict what adversaries can acquire. Pax Silica is more constitutive: It builds a parallel supply chain that member states depend on, creating structural leverage over insiders as well as exclusion of outsiders. This is closer in logic to a monetary system than a trade restriction. Membership shapes a country’s entire technological trajectory, from its computing capacity to its AI development pathway and defense capabilities. The framework extends strategic capture beyond the domestic relationship between governments and firms into the international order itself. Alliance membership becomes the price of access to critical technology.

The Chinese AI model

In China, corporate law requires Communist Party committees in major firms, licensing structures constrain what companies can build or publish, and the state directs research priorities by decree. In Western democracies, companies can sue and parliaments can enact laws. Those channels do not exist in China in the same independent, adversarial form.

What is emerging in the West is therefore not the Chinese model, but something that produces a functionally similar outcome under conditions of strategic stress. When national security is invoked, the systems converge on the same result: private companies subordinated to state requirements. The mechanism differs – Washington uses procurement leverage and security designations where Beijing uses law and ownership – but the function is the same. Call it functional convergence under stress. In ordinary times, the AI industry operates with genuine autonomy. When the state decides a technology is sovereign-critical, that autonomy ends.

The European experience illustrates the same dynamic through a different mechanism. The EU AI Act, passed in March 2024 with overwhelming cross-party support, is being progressively softened under competitive pressure and U.S. lobbying – its high-risk provisions delayed, its liability frameworks stripped back. France, the UK and Germany have each pivoted from regulatory ambition to industrial policy, prioritizing AI investment and national champions over citizen protection frameworks. The direction of travel is the same as in Washington: states asserting primacy over private AI governance, whether through coercion or competitive capitulation.

AI enters politics

Strategic capture does not occur in a vacuum. Across Western democracies, AI governance is becoming an electoral issue for the first time. In the U.S., the 2026 midterms are already being shaped by AI-related spending at unprecedented scale. Political action groups backed by major AI companies have raised over $125 million, while pro-regulation groups fund opposing candidates. In Europe, the rollback of the AI Act has generated significant civil society backlash, and AI-generated disinformation featured prominently in the 2025 German election campaign.

Voter sentiment is running well ahead of political action on both sides of the Atlantic. Gallup found 80 percent of Americans want safety rules even if they slow development, and Eurobarometer data shows comparable European majorities favoring strict AI oversight. The Anthropic confrontation has crystallized cross-party discomfort. Free-market conservatives are uneasy with governments designating domestic companies as security threats for maintaining their own safety policies, while civil liberties advocates across the spectrum are alarmed by the surveillance dimension. Whether this politicization produces democratic accountability or simply a more contested form of capture will be one of the defining political questions of the next several years.