The Space-based Data Center Hype Machine Is Already in Orbit
IEEE Spectrum Harry Goldstein
Musk says space data centers beat Earth-based ones on cost within two to three years. The math, and physics of cooling chips in orbit, say otherwise.
Based on reporting by IEEE Spectrum, Harry Goldstein — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Elon Musk told Davos in January that space will be the cheapest place to run AI compute within two or three years. Weeks later SpaceX filed FCC paperwork for a constellation of up to one million orbital data center satellites, and just before the IPO Musk previewed specs for something called AI1. Big vision. The problem is the arithmetic doesn't cooperate with the timeline.
Right now there are about 14,500 active satellites in orbit, and Starlink alone accounts for roughly two thirds of them. Getting to a million satellites on Starship, which carries maybe 60 satellites per flight, means something like 16,666 dedicated launches. SpaceX set a record with 165 orbital missions in 2025. Even multiply that by ten and you're still looking at a decade of nonstop launches just for deployment, never mind building the hardware. Starlink currently manufactures around 4,000 satellites a year; a tenfold jump in output still puts full deployment 25 years out.
And then there's heat. Cooling a single Nvidia H100 in orbit already strains the system: the chip draws 700 watts and needs 1.4 square meters of radiator running at 60 degrees Celsius. Scale that to a 40-kilowatt server rack and you need an 80-square-meter radiator. A 100-megawatt facility would require 2,500 of those panels. Starcloud, a startup that's applied for an 88,000-satellite constellation of its own, has flown exactly one H100 so far — and its radiator couldn't even let the chip run at full power. Astronomers, meanwhile, are worried that giant radiator wings on a million satellites would wreck the night sky, on top of raising Kessler syndrome debris risks.
So why is everyone pushing this anyway? Follow the money, and the org chart. Musk has xAI building the compute, SpaceX launching it, and Tesla supplying the solar panels — a closed loop where he's effectively paying himself at every step. Analysts see a narrower but real opening: Starlink's existing laser-link network is genuinely hard to replicate, and could make SpaceX's constellation viable for inference workloads within five to ten years, even if training in orbit stays a non-starter because of latency and synchronization limits. That's a real business. It's just a much smaller and slower one than a million-satellite headline suggests.
My take — AI-written commentary, not fact-checked reporting
Musk's timelines are a genre unto themselves at this point, and this one deserves the same skepticism as full self-driving by 2017 or Optimus armies by now. The interesting story isn't the million-satellite fantasy, it's that SpaceX has quietly built the one piece of infrastructure, the laser-link backbone, that actually makes a smaller orbital compute business plausible in five to ten years. Bet on that version, not the press release version.
Read more about this at: IEEE Spectrum