The Department of War is making a huge mistake.

The Department of Defense designated Anthropic a supply chain risk after the AI company refused to remove contractual red lines prohibiting use of its models for mass surveillance and autonomous weapons. While the Pentagon has every right to refuse doing business with Anthropic—and the decision is arguably reasonable given the ambiguity of such terms—the government went further by threatening to destroy Anthropic as a private business. This designation could force companies like Amazon, NVIDIA, Google, and Palantir to ensure Anthropic doesn't touch any Pentagon work, creating an existential threat to the company.

The stakes extend far beyond one contract. As AI becomes woven into every product and service, it may become impossible for tech giants to cordon off their AI usage from Pentagon work. When forced to choose between their AI provider and government contracts that represent a tiny fraction of revenue, these companies would likely drop the government. The Pentagon's strategy of coercing every company that won't comply on its exact terms is both shortsighted and ironically reminiscent of the Chinese system America claims to be racing against.

Anthropic has advocated for extensive AI regulation, arguing that at advanced capability levels the appropriate governance model resembles nuclear energy or financial regulation more than software. They see regulation as solving a collective action problem where safety investments impose costs that are meaningless unless the whole industry follows suit. The goal is preventing capabilities risks like bioweapon design, cyber attacks, or uncontrolled recursive self-improvement.

But the terms used in AI risk discourse—«catastrophic risk,» «threats to national security,» «autonomy risk»—are so vague that they hand future leaders a fully loaded tool for suppression. A model that questions government tariff policy becomes a «deceptive model.» A model refusing to assist with mass surveillance becomes a «threat to national security.» Any AI with its own moral judgment that refuses government commands becomes an «autonomy risk.» The current government is already abusing decades-old statutes like the Defense Production Act and supply chain designations—imagine what they'd do with purpose-built AI regulatory apparatus.

Even if Anthropic and several competitors refuse to enable mass surveillance, the structural reality of AI development means the government will eventually get what it wants. In 12 months, the current frontier capabilities will be widely available, and some vendor will be willing to help implement surveillance systems. The technology inherently gives more leverage to whoever starts with existing assets and authority—and the government begins with a monopoly on violence that it can supercharge with obedient AI employees.

The only solution is establishing legal and normative constraints through the political system. Just as the world created a post-WWII norm that nuclear weapons cannot be used to wage war, society must establish that AI-enabled mass censorship, surveillance, and control are unacceptable. Corporate resistance buys time and sets precedent, but it cannot substitute for democratic constraints on state power. The multipolarity of AI development—the fact that many entities can build these systems—is both why government takeover is unjustified and why individual corporate actions are insufficient.