Washington Couldn't Hold the Buildout Down. Now It's Signing the Way Up.

Intel has unveiled a chip built for nowhere on the ground. Starfire is a radiation-hardened AI inference chip off the company's Trusted Foundry line — the manufacturing program reserved for defense customers — and its stated purpose is to give the United States government the ability to run AI inference in orbit ^[1]. Radiation-hardening is for a place with no atmosphere to soften the sun: a chip designed to think where there is no zoning board, no water to draw, no neighbor to wake. Washington built its sovereignty program to hold the buildout on the ground. In April the administration designated AI data centers critical national-security infrastructure ^[2]; in August federal energy regulators ordered utilities to speed up grid connections for AI projects ^[3]. Both were aimed at land, water, and the electric grid — at making it easier to pour concrete where people live. The ground is refusing. Opposition to a data center within three miles of home rose from 28% to 41% in a year, and New York's governor signed the first statewide moratorium ^[4]. In Abilene, Texas, home of the flagship Stargate site, an 85% tax abatement bought the initial consent; a year later rents were up roughly half, more than a hundred students had arrived to find no housing, and shelter demand rose ^[5]. CoreWeave's chief executive read the moratoriums as a relocation problem, not a reduction — a local veto moves the concrete, it doesn't shrink the demand ^[6]. The turn is happening in the same buildings that wrote the ground rules. In April the Federal Communications Commission voted to bar China Mobile, China Telecom, and China Unicom from running data centers on U.S. soil ^[7]. The chairman's reasoning was plain.

to secure our networks from these bad actors, including limiting their interconnection ability. — Brendan Carr

Months later the same agency accepted the initial phase of a plan to orbit 100,000 AI-computing satellites ^[8]. The company behind it has asked permission for a million ^[9]. Opposition is already migrating to that single federal docket too — environmental groups petitioned the same agency to block a separate orbital project ^[10]. The second hand is the checkbook. Weeks before SpaceX's IPO roadshow, the Space Force signed a $2.29 billion fixed-price contract ^[11]. The program is communications — the backbone layer for the Golden Dome missile-defense system — which makes it anchor tenancy for the launch platform, not a purchase order for orbital compute ^[12]. The third is the customer. SpaceX is negotiating a multibillion-dollar agreement to manage the Pentagon's classified AI workloads ^[13]. Nothing on any of these pages says leave the ground; the terrestrial program is still running, still aimed at the ground. But the same national-security apparatus that could not win the ground's consent is, on three separate lines of paper, licensing the route, supplying the silicon, and paying the company that would leave it. None of this is a migration yet. There is no operating orbital compute. Launch costs sit roughly seventeen times above the level the economics need ^[14], the test chips can run only about fifteen minutes before they need to cool down ^[8], and insurers cannot price the radiation, debris, and launch risk — the orbital firms are courting Lloyd's of London just to get the debt that scaling would require ^[15][16]. SpaceX's own IPO filing conceded the risk in one sentence.

Our initiatives to develop orbital AI compute and in-orbit, lunar, and interplanetary industrialization are in early stages, involve significant technical complexity and unproven technologies, and may not achieve commercial viability. — Space Exploration Technologies Corp.

That is the company's disclosure, sitting beside Elon Musk's public claim that space would be the cheapest place for AI within two or three years ^[9] — and beside Sam Altman's flat dismissal of the economics ^[17]. The company offers two reasons for being up there, and both are on the record. One is political: the orbital program is, in the company's own telling, built to get around intensifying bipartisan opposition to data centers on the ground ^[18]. The other is physical.

the computing power needed to develop the next generation of AI is "outpacing what terrestrial power, land, and cooling can deliver on the timelines that matter," — Space Exploration Technologies Corp.

Google is doing both things at once, which makes it the cleanest exhibit. It is paying SpaceX $920 million a month for 110,000 GPUs in terrestrial buildings, the first tranche starting this month ^[19]. And today its own orbital prototype flew, built to get around what the company calls terrestrial bottlenecks — power-grid strain and water use ^[8]. This week the same company shipped its frontier model, Argon, to vetted governments first, while the public waits, under a voluntary White House accord ^[20]. China is running the same race by central plan — gigawatt-scale orbital hubs as a pillar of its five-year program ^[21]. Washington licenses the route, supplies the silicon, and pays a private firm to build the way up it could not clear on the ground. The question the program was built to answer — where American compute sits — is being answered by the ground itself, which keeps saying no. What the government is keeping for itself is narrower, and easier to hold: who builds the machines, who funds them, and who is on the list. Argon's first users are a list of governments. The rest of us are not on it yet.