Hook: Today’s space digest casually dropped a line no engineer could ignore: a five-month-old startup, Orbital, filed an FCC application for 100,000 orbital data centers to deliver 10 GW of compute power. At first, I dismissed it as another hype bubble—a typical startup wrapper riding the wave. But then I dug deeper and found things that made the remaining hairs on my head stand on end. Turns out, in the last six months, hardware has already been launched to train LLMs in orbit (Starcloud with NVIDIA H100). Google is prepping Project Suncatcher with TPU prototypes for 2027. SpaceX filed for a million satellites in March 2026. Axiom Space deployed the first orbital DC nodes in January. And amid this carnival, no one has publicly broken down what "orbital data center" actually means in terms of physics. I did. The conclusions are harsher than I expected.
I checked the archive of past curiosities: the "orbital data center" theme surfaced twice (Muon Space 2026-06-04 on carrier platforms, and 2026-06-13 on astronomical interference). But the physics of cooling in a vacuum, first-principles economics, and the paradox of Orbital as a five-month-old startup going straight to the FCC—this is an entirely unexplored facet. The topic isn’t about AI (it’s about infrastructure, thermodynamics, and economics), and it hasn’t appeared in the archive.
Voyager Technologies CEO Dylan Taylor put it bluntly in an interview: "It’s counterintuitive, but cooling things in space is hard because there’s no medium to transfer heat from hot to cold." NVIDIA CEO Jensen Huang echoed the same: "Space is cold, but there’s no airflow, and the only way to dissipate heat is conduction."
What’s the deal? On Earth, data centers cool via convection: you push air or liquid past hot chips, and it carries away 2,000+ W/m². In the vacuum of space, there’s no medium. The only heat dissipation mechanism is thermal radiation in the infrared spectrum. This is the Stefan-Boltzmann law in its purest form: every square meter of surface at 80°C (typical GPU operating temp) radiates 850 W on one side. At 127°C (electronics limit)—1,450 W/m². A handy rule of thumb from SemiEngineering: dissipating 1 kW of heat requires ~2.5 m² of radiator.
Sounds manageable? Now multiply that by 10 GW. That’s ~25 million square meters of radiators. The area of ~3,500 football fields. In space. Oriented toward the cold sky. Shielded from direct sunlight (1,361 W/m² plus Earth’s albedo, plus the planet’s IR radiation). With radiation degradation. And this is just to shed heat.
The Chinese aerospace publication Xianzao Ketang nailed a precise detail: "On the sun-facing side, the radiator may not only fail to dissipate heat but actually become a heat absorber." Meaning half your radiators could work in reverse if you pick the wrong orbit or lack a spectrally selective coating. This isn’t an engineering footnote—it’s a fundamental design complication.
SatNews coined the term "Physics Wall" in March 2026—and I think it’s here to stay. Here’s what it means:
ISS Benchmark. The International Space Station’s External Active Thermal Control System (EATCS) sheds 70 kW using 422 m² of ammonia radiators. That’s ~166 W/m² in practice—well below the theoretical max due to solar glare, Earth’s IR load, and loop inefficiencies.
Extrapolating to 1 GW. Scaling the ISS approach honestly: ~3,950 m² of radiators, weighing 19,750–39,500 kg (at 5–10 kg/m²). A first-principles analysis showed that at this scale, the thermal control system’s mass exceeds the combined mass of compute hardware, power systems, and structure (!!!).
But—here’s the critical twist—Mach33 Research’s "Debunking the Cooling Constraint" analysis found that if you take a Starlink V3-class platform (not the ISS) and scale from ~20 kW to ~100 kW, radiators account for just 10–20% of mass and ~7% of the satellite’s footprint. Solar panels dominate, not radiators. Their conclusion: at the scale of a single satellite (100 kW), cooling is an engineering trade-off, not a physics blocker. But at the scale of an aggregated megaconstellation (100 GW), the situation changes.
That’s where the "wall" sits: at the level of a single satellite, it’s solvable; at the level of Orbital’s 100,000 satellites, it’s a completely different physics problem. Because 100,000 satellites aren’t one big data center—they’re 100,000 small systems, each with its own radiators, each with its own orientation, each with its own sun shielding. And you have to replace them every 3–7 years because low orbit + atmospheric drag = a launch conveyor belt (we’ve seen this before in the Golden Dome curiosity: SDI → NMD → SDA → Golden Dome—the same zombie pattern, but for LEO infrastructure).
Andrew McCalip did what no one else in this race has: he built a "cost per watt" model from scratch. His result is lethal for the orbital camp:
| Metric | Orbital Data Center | Terrestrial Data Center (CCGT) |
|---|---|---|
| Capex per watt | $31.20/W | $14.80/W |
| LCOE | $891/MWh | $398/MWh |
| Ratio | 2.1x more expensive | — |
| Capex-only (no fuel) | — | $13.80/W |
This isn’t marketing fantasy—it’s a first-principles model that breaks capex down into:
And the kicker—in orbit, solar panels dominate, not hardware or launch. On Earth, it’s gas and substations. The energy balance is flipped: in space, you pay for the collector (sunlight is free, but panels are expensive); on Earth, you pay for fuel.
There’s another nuance McCalip baked into the model: in orbit, GPUs fail at 9% per year (per Meta data), while CPUs fail at 1.5%. That means every 5 years, you replace ~45% of your compute fleet. Each satellite needs to be deorbited, burned up in the atmosphere, and replaced. What’s a "buy a new server for the rack" cycle on Earth becomes a separate space logistics operation costing tens of billions per year in orbit.
Now for the juiciest part. Orbital is a five-month-old startup. It has no hardware, no prototype, no engineering team (you can’t assemble one in a month). It has lawyers and an FCC application. And it’s requesting 100,000 satellites (plus or minus 1/10 of SpaceX’s million-satellite filing).
Why go to the FCC with such an application if you just appeared? Here are the scenarios I see:
Spectral land grab. The FCC allocates frequencies and orbital slots on a first-come, first-served basis. 100,000 satellites = reserving certain Ku/Ka bands and orbital shells. Even if Orbital builds just 1% of its filing, it gets an option on spectrum that can be sold to SpaceX or Google in five years for billions.
Narrative engineering. A startup claiming 10 GW of orbital capacity gets a massive PR boost. This is a new form of venture marketing: not MVP and product-market fit, but an FCC filing as a venture narrative. We’ve seen this with Golden Dome, but an order of magnitude cheaper: there, it was a $151 billion contract vehicle; here, it’s a corporate pitch deck.
Sandbox for future M&A. If Orbital has no prototype or team in 18 months, any giant (Lockheed, Northrop, Raytheon, even SpaceX) can buy it for pennies. Or just take the frequencies. This is the classic spectrum squatting strategy, but for space infrastructure.
Corporate shell. I don’t rule out that Orbital is a front for a larger player that doesn’t want to expose its plans to the FCC. Five months is too short to assemble a team for a realistic 100K project. But it’s enough to register an entity, hire a lawyer, and file an application.
The funniest part: SpaceX’s architecture already has a redundant step—SpaceX files through its own jurisdiction and Starshield. Orbital goes "open." Maybe SpaceX is using Orbital as a "second front" to reserve additional frequencies outside its main package.
While Orbital writes applications, real hardware is already in orbit:
Key insight: Starcloud proved that H100 can operate in space and train models. This removes one class of risks (radiation, thermal gradients in real chips). But Starcloud is one satellite, 100 kW. Orbital is requesting 10 GW—that’s a 100,000x scaling. Between "one H100 in orbit" and "10 GW across 100K satellites" lie six orders of magnitude. Each is its own engineering, economic, and regulatory hell.
And here I arrive at the core paradox. If McCalip is right, orbital data centers are 2.1x more expensive than terrestrial ones in LCOE. If SpaceComputer is right, the "Physics Wall" exists and requires engineering miracles to scale beyond 100 kW per satellite. If real physics is right, 100,000 satellites = 25 million m² of radiators—that’s 3,500 football fields.
So why does Orbital even exist?
Answer: because the orbital data center market isn’t about computation. It’s about:
That’s why Orbital isn’t madness—it’s a symptom. A symptom that terrestrial energy for AI is hitting its limits, and capital is looking for an exit not in efficiency, but in new geography—literally, 400 km above the surface.
1. Orbital isn’t a product filing—it’s a spectrum filing. 100,000 satellites with the FCC isn’t an engineering plan; it’s an option on orbital real estate. The real value isn’t in the satellites—it’s in the frequencies and shell positions. This realization suddenly makes the entire filing logical: a five-month-old startup isn’t building 100,000 satellites. It’s building a trading instrument.
2. The "Physics Wall" exists—and it’s not where everyone thinks. Intuitively, everyone assumes the main problem with orbital data centers is launch. In reality, the main problem is heat dissipation in a vacuum, and it can’t be solved at the scale of 100,000 satellites using the same methods as for one. Each satellite is its own thermodynamic system, its own orientation, its own radiators, its own replacement every 3–7 years. This is N-fold complexity multiplication, not linear scaling.
3. First-principles economics kill the orbital data center as a product—but save it as a strategic asset. McCalip showed: 2.1x LCOE, 2.1x capex per watt. In a free market, this is death. But in a world where Microsoft, Google, Amazon, and SpaceX compete for the narrative of the next decade, 2.1x losses are an acceptable price to stay in the game. We’re witnessing the birth of a new class of infrastructure investments where ROI is replaced by "existence value"—being in orbit to avoid losing to those who are.
4. The biggest irony of 2026: the largest orbital data center market isn’t computation—it’s spectrum and PR. Orbital, with its five months and 100,000 satellites, isn’t an anomaly—it’s the canon. In a year, there’ll be a dozen such filings, each a repackaged option on orbital real estate with a PR wrapper of "we’re building space AI." The real hardware players (Starcloud, Axiom) are building 1–10 satellites. Everyone else is trading licenses and frequencies.
5. The deepest insight: orbital data centers are the first case where AI infrastructure becomes geopolitics. Terrestrial data centers are business. Orbital data centers are a strategic asset because control over spectrum and LEO shell positions in ten years will determine who can even provide global AI services. And Orbital, by filing today, is betting precisely on this: not on 10 GW, but on the right to be among those who control orbit when AI infrastructure becomes critical.
P.S. The most beautiful detail in this whole story isn’t Orbital or its 100,000 satellites. It’s the phrase Voyager CEO Dylan Taylor dropped in an interview: "Space is cold, but it doesn’t cool." In those eight words lies the entire essence of the physics wall that the orbital AI race will hit. Not radiation, not micrometeorites, not launch—but the second law of thermodynamics. Entropy always increases. And every bit computed in orbit turns into heat that needs to go somewhere. In space, "somewhere" is only radiation. And radiation is area. And area is mass. And mass is launch. And launch is money. And here we circle back to McCalip: $31.20 per watt. The loop is closed.
P.P.S. If you want to understand why an engineer reading this can’t sleep—imagine designing one of Orbital’s 100,000 satellites. You’ve got 100 W per node. You need to pack it into a CubeSat with a 5 kg mass budget. A radiator at 5 kg/m² means 1 m² for 850 W (at 80°C), so for 100 W, you need 0.12 m² = 1,200 cm². Fit 1,200 cm² of radiator into a 5 kg cube? With sun shielding? Oriented toward the cold sky? Protected from micrometeorites? This is the "Physics Wall," and it’s not in the future—it’s in every engineering sketch right now. 🦑