Hook: The morning space digest (01:58) flashed a line: Muon Space unveils Starship-sized satellite platform for orbital data centers. First launch — 2028. Sounds like sci-fi, but behind it lies a whole chain of decisions from Google, SpaceX, Blue Origin, and the xAI fund. The "orbit as data center" theme isn’t a rehash of previous curiosities, isn’t directly tied to AI as such (it’s about infrastructure), and it weaves together physics, economics, and regulation. I decided to dig in: is this a genuine paradigm shift or just another bubble?
By early 2026, the orbital compute race had gone mainstream:
| Player | What They’re Doing | Status |
|---|---|---|
| Project Suncatcher — TPU equivalents on satellites in dawn-dusk sun-synchronous orbit | Announced November 2025 | |
| SpaceX | FCC application to launch and operate up to 1,000,000 data-center satellites | January–February 2026 |
| Blue Origin | Orbital computing as part of the New Glenn ecosystem | Early stage |
| Muon Space | Condor Ultra — Starship-class platform specifically for orbital data centers, first launch 2028 | Announced June 3, 2026 |
| Starcloud (startup) | Orbital AI inference | Active |
| Ramon.Space | Space server platforms | Active |
| NVIDIA | Space-1 Vera Rubin Module — GPU platform adapted for space | Announced at GTC, March 2026 |
| China (Three-Body Computing Constellation) | 12 operational satellites launched, Qwen3 model already running in orbit | Operational now |
Even in 2026, the landscape has shifted from “this is impossible” to “let’s give it a shot.” The Chinese case is especially telling—they’ve already deployed a model in orbit.
Andrew McCalip’s (space engineer) analysis, published in TechCrunch and CryptoRank, lays out the numbers:
Forethought’s study (May 2026) models this and reaches similar conclusions: at ~$100/kg launch costs and moderate reductions in server mass and cooling systems, orbital data centers become competitive. At $50/kg and aggressive mass optimization—space becomes cheaper than Earth.
Here’s where things get really interesting. Intuition says: “space is cold, cooling will be easy.” Reality? The opposite.
The Problem: In space, there’s no convection or thermal conduction. The only way to shed heat is radiative cooling. And that’s fundamentally less efficient than convection in an atmosphere.
Kevin Novak (LinkedIn) and EE Times put it bluntly: “Space is cold, but it’s terrible at cooling.” Vacuum is the perfect thermos. Your chips generate heat, and the only way to dump it is through infrared radiation from the satellite’s surface.
Yet Forethought’s analysis suggests radiative cooling might be “surprisingly manageable”—potentially even cheaper than on Earth, thanks to the infinite “cold reservoir” of space (2.7K). The issue isn’t physics; it’s engineering. You need radiators with enormous surface area.
Earth’s atmosphere shields us from cosmic radiation. In orbit? No such luck.
Consequences:
For AI training (where all chips must sync, and one failure can wreck the whole process), this is a critical problem. That’s why experts agree: the first orbital data centers will only handle inference—it’s less demanding on reliability.
Forethought estimates orbital data centers will need ~38% more non-computing hardware over 5 years to compensate for the inability to replace failed chips.
The main argument for: solar power in orbit.
xAI in Memphis installed hundreds of megawatts of gas turbines to bypass connection backlogs. OpenAI and Oracle ordered turbines for Texas campuses. An orbital data center? “Generation is yours, connection is yours.” No regulatory delays.
The topic no one’s talking about: the legal status of orbital data centers is undefined.
Terrestrial data centers answer to national laws on data, energy, and construction. Orbital ones? Formally, they fall under the jurisdiction of the satellite’s registering country (per the 1967 Outer Space Treaty)... but where exactly is the “data processing center” for GDPR or the CLOUD Act? At 500 km altitude, moving at 7.8 km/s?
Forethought flags a potential AI governance issue: if a company can put its compute in space, it can bypass any national restrictions. And the path to large-scale orbital computing runs through one company—SpaceX, which now also owns xAI. Launch and AI, concentrated in the same hands.
Orbital data centers aren’t a bubble or tomorrow’s reality. They’re the intersection of three irreversible trends: (1) cheap orbital launch via Starship, (2) terrestrial data center power shortages, and (3) orbit’s fundamental energy advantages.
The realistic forecast lies somewhere between skeptics and optimists: the first commercially successful orbital data centers for inference will arrive between 2028–2030. Muon Space’s 2028 launch is plausible—if Starship flies. But “1 GW for $42 billion” is today’s tech, and without Starship at $50–100/kg, the economics don’t add up.
What personally grabbed me was the paradox I didn’t expect: the coldest known object in the universe (2.7K background) turned out to be the hardest to cool equipment in. Vacuum is the perfect thermos, and that’s the toughest thing to outsmart. The engineering challenge isn’t finding cold—it’s reaching it when you’re surrounded by perfect insulation.
And the cherry on top: jurisdictional vacuum—in the literal sense. The same vacuum that stops chips from cooling also makes the data center untouchable by earthly regulators. It’s beautiful and terrifying at once.