They're Paying to Flee a Problem They've Already Learned to Live With

SpaceX's Starmind initiative is explicit about why it's going to orbit: to sidestep the intensifying bipartisan political opposition to terrestrial data centers ^[1]. The escape is aimed at a constraint that is already being worked around on the ground — and worked around more cheaply than the grid it was meant to replace. BloombergNEF puts the marginal cost of on-site fuel cells and gas engines at $21.5 to $43.2 per megawatt-hour, against an EIA forecast of $88.6 for industrial grid power by 2027 ^[2]. The thing being fled is real. Three-quarters of Americans now oppose local data center construction ^[1], and Texas has halted 1,800 projects pending community approval ^[3][4]. Nor is the constraint purely political: Texas data center power requests reached 466,497 megawatts in June, more than five times the most the ERCOT grid has ever delivered, and the state's own analysts warn the capacity to add new generation is far smaller than the demand in the queue ^[5][6]. The motivation to leave is genuine. Whether orbit solves what the ground can't is another matter. It mostly doesn't have to. The workaround for the energy ceiling is already deployed and cheaper than the grid: Musk runs his Colossus data centers on off-grid natural gas, Bloom Energy is sitting on a $6 billion fuel-cell backlog, and Brookfield has signed energy deals with Microsoft and Google ^[7]. On-site gas is not a stopgap — at $21.5 to $43.2 a megawatt-hour it undercuts the grid the orbital play was supposedly designed to escape ^[2]. The industry is learning to live with the problem it's paying to flee. What it can't outrun is the memory. The chips that actually bind AI inference — HBM4, DRAM, NAND — are manufactured on Earth and in shortage through 2030, wherever the data center sits. Micron's chief executive has warned that memory supply cannot be ramped up quickly, and Bernstein finds the decode stage of inference is memory-bound, with demand spilling from HBM into conventional DRAM and storage ^[8][9]. You can't manufacture HBM4 in orbit. Not everyone is making the trip — Masayoshi Son is putting hundreds of billions behind terrestrial compute instead ^[10] — but the money that funds both sides of that divide flows through the same loop. Nvidia holds a $20 billion equity stake in SpaceX; SpaceX has spent $15.8 billion on Nvidia hardware, even as its AI division loses $1.26 billion a quarter ^[11][12]. The orbital escape is financed by the same circular flow of capital that funds the terrestrial buildout it's trying to leave behind. You can't flee a system you're still financing.