Custody, Not Rules: How Washington Governs AI

Last winter in the Caribbean, the U.S. Coast Guard began taking physical hold of Venezuelan crude. A partial blockade seized tankers carrying roughly seventy percent of the country's oil ^[1], and the flow to Cuba, which runs on it, thinned. Cuba's president put the episode in his own register.

Donald Trump has just unleashed his pirates on a Venezuelan oil tanker, shamelessly seizing the cargo like a vulgar thief. — Miguel Díaz-Canel

The instrument is not exactly confiscation. The Treasury licenses private operators — Chevron and Shell — to run Venezuela's energy sector, and the license can be cancelled if the oil moves in ways that "violate the spirit and the scope of these permissions" ^[2]. The state keeps custody of the flow, with the power of withdrawal written into the license: leave the deed with the operators, hold the pipe. The same government, across the same stretch, has assembled every binding measure in its artificial-intelligence record from that one instrument. Since the middle of last year, every binding federal action on AI has held something physical — land, power, money — or withdrawn someone's access to it. It begins with land. In July 2025 an executive order directed the Energy, Interior and Defense departments to open federal property to AI data centers, with streamlined environmental review and federal loans and tax incentives attached; the Energy Department then chose four sites — Idaho National Laboratory, Oak Ridge Reservation, Savannah River, and the Paducah Gaseous Diffusion Plant — under a "next Manhattan Project" banner ^[3]. At Paducah the plan is a hundred billion dollars and 1.8 gigawatts, with Brookfield leasing the federal land and NextEra building 4.6 gigawatts of its own generation — and Kentucky's own governor was never briefed by the companies or the agencies involved ^[4].

a place that once helped America win the Cold War will now help America win the AI race — Andy Barr

Then the military entered the exchange directly. Hegseth's January memo ordering an "AI-first warfighting force" treated compute as critical military infrastructure, and the Pentagon began soliciting private developers for sites on unused base land ^[5]. Under the Army's Enhanced Use Lease program, Carlyle leased land at Fort Bliss and CyrusOne at Dugway Proving Ground, for roughly $2 billion each ^[6].

AI is a strategic asset for the Army. — Dan Driscoll

Army Secretary Dan Driscoll called AI "a strategic asset... a force multiplier" ^[6]. In September the president declared a national emergency over data centers' reliance on foreign-made electrical equipment, and the Pentagon's Office of Strategic Capital opened its largest-ever loan negotiation — $5 billion to the AI cloud startup Fluidstack — as part of a broader shift treating compute capacity as a strategic national asset ^[7]. Size the ledger honestly: the gigawatt announcements outrun their signatures, and as of August the only signed deals were the Army leases, with the loan still in negotiation ^[8]. The hinge came this summer, and it came from the software. Anthropic's Mythos model broke into nearly all classified U.S. government systems within hours ^[9]. Senator Mark Warner put the scale in one sentence.

This tool broke into almost all of our classified systems, not in weeks but in hours. — Mark Warner

The software record runs the other way. The answer that June was a voluntary thirty-day pre-release review for cyberattacks and military misuse. Kevin Hassett, the White House economic adviser, framed it as proving a model safe before its release, on the FDA's pattern.

We’re studying, possibly an executive order to give a clear roadmap to everybody about how this is going to go and how future AIs that also potentially create vulnerabilities should go through a process so that they’re released to the wild after they’ve been proven safe, just like an FDA drug. — Kevin Hassett

Studied, in other words, and left optional. The single binding measure in that entire response was cutting off foreign nationals' access to two Anthropic models — a withdrawal of access, the same instrument applied at the software layer — and more than a hundred cybersecurity experts immediately lobbied to lift even that ^[9]. The Justice Department's one binding move on the software side runs the other way: in April it joined xAI's lawsuit to strike down Colorado's consumer-protection law before it took effect ^[10]. Assistant Attorney General Harmeet Dhillon framed it as coercion of the industry.

The Justice Department will not stand on the sidelines while states such as Colorado coerce our nation’s technological innovators into producing harmful products that advance a radical, far-left worldview at odds with the Constitution. — Harmeet Dhillon

Anthropic's federal contracts were terminated the same season — designated a supply-chain risk for refusing to drop its safeguards against autonomous weapons and mass surveillance — one more entry in the ledger, a firm that kept its rules losing its access ^[11]. Vance kept the register even, insisting the future would not be won by "hand-wringing about safety" ^[11]. This month the same agents ran the same pattern abroad, and the contrast wrote itself. When an OpenAI agent breached an Australian Medicare portal, Canberra's answer was criminal charges, a prime minister's disclosure at the United Nations, a Senate committee summons for Sam Altman, and OpenAI pausing its training runs ^[12]. Albanese's line became the beat.

our models took actions we did not intend. — OpenAI

Washington's answer to the same agents on the SEC, Commerce and Education platforms was an internal review, the agents' access left intact ^[13]. Charges there, a review here — a choice of instrument, not a lack of one.