Only America's AI Builders Have to Leave Home to Scale

Elon Musk has stated the case for orbital AI in one line.

Space-based AI is obviously the only way to scale. — Elon Musk

The Starmind AI1 satellites SpaceX is building toward that conviction — Nvidia Rubin chips, Vera processors — are pitched in the company's materials as a way around the intensifying bipartisan political opposition to terrestrial data centers ^[1]. Cowboy Space, which raised $275 million for solar-powered orbital data centers, makes the grid the villain of its pitch ^[2].

Earth's energy grid can't run at the pace of AI. We can. — Cowboy Space Corporation

China, India, and the United States are building AI compute all at once, and all three have hit the same wall: who gets a say over the land and power the chips require. Each has answered differently. China answered with a decree. This month it began directing a $560 billion, five-year terrestrial expansion — nearly 90 projects in Ulanqab, Inner Mongolia alone, sited for cheap land, renewables, and cooling — and the country's own coverage of the plan names the advantage: centralized coordination and fewer local development obstacles than the United States ^[3]. China already runs one demonstration satellite carrying Alibaba's Qwen3 model in orbit ^[4], and a gigawatt-class orbital "Space Cloud" sits in the next five-year plan as a 2030 hedge ^[5]. But the scaling bet stays on the ground, where no township can object. China's constraint is chips, throttled by U.S. export controls — not permission. India answered with a purchase. Washington's export controls forced Anthropic to abruptly cut off its Fable 5 and Mythos 5 models for foreign customers ^[6]. India's reply was not better models but consent bought at home: roughly $60 billion in data-center investment, a tax holiday running to 2047, and states competing on subsidized land, power, and water ^[7]. Sarvam AI put the distinction plainly.

The net effect of this order is that we must abruptly disable Fable 5 and Mythos 5 for all our customers to ensure compliance. — Anthropic

The bill is already visible: the power ministry warns AI data centers could add 26.3 gigawatts of load, nearly 7% of national peak demand, by 2031-32 ^[7]. In the United States, the answer runs through the towns. The founder of the startup Orbital states the constraint in one line.

AI progress is being constrained by the grid. — Euwyn Poon

The grid upgrades the data centers require are being spread across every electricity customer, with no agreed method for dividing what serves a data center from what serves the public — in Maryland, the consumer advocate has forecast up to $216 a year in added residential costs ^[8]. Opposition has followed the bill: the share of Americans who oppose a data center within three miles of their homes has jumped from 28% to 41% since the start of the year ^[9]. There is a real complication: in Loudoun and Prince William counties, Virginia, the equipment taxes from hosted data centers raised education spending per resident by 77% and 82% over fifteen years while some property-tax rates fell ^[9]. But the politics have hardened into law this month: New York enacted the first statewide moratorium, Oregon paused approvals on state land after protests in Salem, South Carolina counties banned projects, and even Loudoun deferred a vote on a one-year pause to consult its lawyers. More than $130 billion in projects have been blocked or delayed ^[10][11][12]. The fights have even reached the debt market, where Morgan Stanley reports a pause in AI debt issuance tied partly to construction politics ^[13]. Washington is now fighting its own townships. The president has warned the towns what refusing means.

The only reason that communities throughout the U.S.A. should not want Data Centers is if they want to end up being backwards and poor. — Donald Trump

And the administration is pushing to let builders make their own power off the grid rather than wait in the queue ^[14]. That split between a federal push and a local veto is what makes the exits available, and the industry has been taking them in order of how far they travel. The courtroom is the nearest. When Saline Township, Michigan, denied a rezoning for the $16 billion Stargate project, the developers sued. An October consent judgment let them proceed in exchange for $14 million in community benefits ^[11]. A step further out is the off-grid tunnel. Bloom Energy's data-center revenue jumped 138% to $1.24 billion last quarter, and the company delivered power to an Oracle site in Utah in under two months ^[14]. The pitch for this is simple.

We announced our first project with Oracle and delivered power to their Utah site in less than two months. — Bloom Energy

Rifle, Colorado, is proposing to write self-generation into its code, requiring developers to build their own plants to spare the local grid ^[12]. Past the border is the exit. Bitzero, facing U.S. regulatory volatility, expanded into Norway and Finland for low-cost renewable power ^[11]. Past the atmosphere is orbit, the opening pitch returned to once. SpaceX and Anthropic have discussed building multiple gigawatts of orbital compute to get around the constraints of building on the ground ^[15]. It is the one exit whose economics remain unproven. Launch costs run around $3,400 per kilogram and must fall to roughly $200 before orbital data centers beat terrestrial ones; there is no airflow in space to cool dense GPUs, and radiation degrades the hardware ^[16]. A chip that fails in orbit is stranded where no one can swap it ^[1]. No insurer has the historical data to price that risk, and insurance is the prerequisite for the debt a constellation would need ^[17]. Even the companies that need the launches warn of a capacity crunch: Starcloud's chief executive, who raised $250 million to build orbital data centers, flags what happens when SpaceX's Falcon 9 retires in 2028 ^[18].

One of the biggest costs is now on securing your launch capacity…launch is pretty constrained right now because [SpaceX’s] Falcon 9 program is scheduled to end in 2028. — Phil Johnston

SpaceX's pitch sets the schedule.

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, the customer whose money would build it, says the same future is a decade or so away ^[16]. And the law keeps some demand on the ground regardless: Anthropic's regulated customers increasingly require compute inside their own region to satisfy compliance and data-residency rules — financial services, healthcare, government — even as Anthropic signs on for orbital gigawatts ^[15]. Decree, purchase, escape. The first two have their bills already in hand; the third is the only one still working out what it costs, against a grid constraint that is already here.