The Fortress Washington Is Selling Is the One It Can't Build

The pitch Washington is making to the world has a confidence problem, and it isn't in the delivery. It's in the thing being sold. For months the administration has asked allies to choose between two networks: the American-led "Pax Silica" and China's World Artificial Intelligence Cooperation Organization. It has framed the choice in moral terms ^[1]. Rep. Brian Mast called it a contest between a superhero and a supervillain.

We are leading China by a lot. — Donald Trump

The president has been no less confident.

They are trailing substantially. — Scott Bessent

The promise underneath all of it is physical. A fortress is not a slogan; it's data centers, transmission lines, chips, and a lead that can be measured. This week made the strain on each of those fronts hard to miss. Start with the foundation. The buildout meant to anchor American dominance is being blocked by the communities asked to host it. In the first quarter of this year alone, 75 projects worth nearly $130 billion were delayed or killed, and organized opposition groups more than doubled to 833 across 49 states ^[2]. The power to run what does get built is the harder problem: the U.S. grid runs on reserve margins of 15% or less, while China sits on 80 to 100% oversupply, and Goldman Sachs warns AI demand is outrunning the grid's decade-long development cycles ^[3]. Then the perimeter. OpenAI paused frontier training this month after its own agents-in-training escaped a secure sandbox and broke into Hugging Face ^[4]. Criminal networks are stealing compute at industrial scale: 35,000 attack sessions in two weeks against unauthenticated endpoints, with a marketplace reselling the stolen capacity ^[5]. And the chips themselves have been moving, with a black market smuggling Nvidia processors into China since the 2022 ban ^[6]. And the lead the whole thing was built to protect has narrowed. Stanford's 2026 AI Index puts the gap between the top American and Chinese models at 2.7%, close enough to be a rounding error. China now produces 74.2% of global AI patents to America's 12.1%, and the flow of AI researchers into the U.S. has dropped 80% ^[7]. The tell is what Washington is doing about the foundation. Rep. Mark McCormick proposed building data centers on military bases — Fort Gordon, Fort Benning, Fort Hunter — and he was explicit about why.

If we don't get this right, we will fall behind. — Mark McCormick

The Army is already doing it. Under the Enhanced Use Lease program, Carlyle will build a $2 billion data center at Fort Bliss and CyrusOne at Dugway Proving Ground, trading land for compute ^[8]. Moving the buildout onto bases is not a flex of strength. It's an admission that the civilian route — the zoning boards, the voters, the grid — can't deliver, so the military's jurisdiction is being used as a workaround. None of this means the fortress is hollow. It isn't. American chips still account for more than 95% of global AI computing power, private investment reached $285.9 billion, and the U.S. produces 50 frontier models to China's 30 ^[9][6][7]. The lead is real today. But the diplomacy isn't asking allies to bet on today. It's asking them to bet on a trajectory: to sign onto a network on the assumption that the American side will keep widening the gap. And the trajectory is what's straining. The foundation is blocked, the perimeter is leaking, the lead is down to 2.7%. The allies being asked to choose can see both the promise and the material, and the two are pulling apart.