The Government Isn't Deregulating AI. It's Buying It.

In February, the Pentagon declared Anthropic a supply-chain risk. The president called the company a "RADICAL LEFT, WOKE COMPANY," and a $200 million contract was cancelled ^[1]. Then the military kept using the company's Claude model for intelligence and targeting in Iran and Venezuela anyway ^[1]. The government punished the firm in public and depended on it in private — and that contradiction is the clearest single view of what a year of AI policy has actually been building. Call it deregulation and you miss the mechanism. The administration has spent the past twelve months not loosening its grip on the industry but changing the kind of grip it has. Where the previous approach tried to contain AI with export bans and safety rules, this one brings the companies inside — through contracts that double as loyalty tests, through direct financial stakes, and through handing private firms a piece of the state's own offensive power. The loyalty test is the most explicit. Defense Secretary Pete Hegseth put the standard in one sentence.

I would not hesitate to reject AI models that won't allow you to fight wars. — Pete Hegseth

Anthropic failed it: CEO Dario Amodei refused to strip safeguards that kept Claude out of mass domestic surveillance and fully autonomous weapons, and the company was blacklisted ^[1]. The response was not to do without AI but to fund the firms that would comply — seven companies signed on, and smaller startups like Smack Technologies and EdgeRunner AI saw their contracts accelerated to fill the gap ^[2].

We want more, we want demos, let's talk about how we can move faster — Andrew Markoff

Then there's the money. The government is no longer just a customer; it's becoming a shareholder and a revenue partner. Trump confirmed a 9.99 percent equity stake in Intel, funded through CHIPS Act money ^[3].

I really think whoever's number one is going to really win the race. And right now, we're substantially number one. — Donald Trump

The administration lifted the China chip ban on Nvidia and AMD in exchange for a 15 percent cut of the sales ^[4]. And OpenAI, Anthropic, and Google offered federal agencies their frontier models for a dollar — or 47 cents — per agency ^[5].

Our offer provides straightforward and predictable pricing, with no hidden fees and no complex licensing tiers, so agencies can budget with confidence. — Google Public Sector

None of this restricts the industry. It ties the industry to the state's balance sheet. And then there is the power the state is handing back out. In March the administration's cybersecurity strategy shifted from defense to offense, with a promise to "unleash the private sector" ^[6].

We will unleash the private sector by creating incentives to identify and disrupt adversary networks and scale our national capabilities. — Donald Trump

This month Trump signed a memorandum formally authorizing companies to conduct offensive cyber operations against foreign criminal networks ^[7]. A private firm can now do what was once reserved for an intelligence agency — under a license the state can grant or withhold. The same reflex shows up far from AI. Columbia paid $200 million to the Treasury; nine law firms provided $940 million in legal services; activist networks were hit with federal subpoenas; ICE hired private contractors to build dossiers on social media critics ^[8][9][10]. Punish resistance, reward compliance, take a stake. The tool changes; the motion doesn't. The honest objection is that this overstates the coherence. Trump's own rhetoric runs the other way — he told CNBC you want "as little as possible" in the way of guardrails ^[11].

You need some guard rails, but you want to do as little as possible. — Donald Trump

Some AI firms have asked for federal oversight themselves ^[12]. And the courts are split: Judge Rita Lin called the Anthropic blacklisting likely First Amendment retaliation, while an appeals court upheld it on national-security grounds ^[13].

These broad measures do not appear to be directed at the government’s stated national security interests. — Rita F. Lin

a resounding victory for military readiness. — Todd Blanche

But that's the point. This is not a plan with an architect. It's a reflex — and a reflex is harder to reverse than a policy, because there's no single lever to pull and no single person to fire. The blacklist, the equity stake, the dollar-a-year access, the hack-back license: each looked, in its week, like a discrete decision. Lined up, they're a direction. And because no single person designed it, no single person can be made to undo it.