Orbital AI Datacenter
First-Principles Feasibility Analysis
Welcome. This is a first-principles feasibility analysis of putting hyperscale AI compute in orbit.
We model everything from Stefan-Boltzmann radiation to Monte Carlo NPV distributions.
The Pitch vs Reality
"Free solar power! Free cooling! No land costs!"
— Every space datacenter startup pitch deck
"You can't convect or conduct in vacuum. You must radiate — and T⁴ is brutal at datacenter temperatures."
— This analysis
Let's follow the physics and the economics to see who's right.
The pitch sounds great. Free solar, free cooling. But "free cooling" in space is a myth.
There's no air. No water. Only radiation. And radiation at 350K is terrible.
The Vacuum Insulator Problem
Space is the best insulator known to physics
Conduction
Requires contact
Convection
Requires fluid
A Thermos flask is literally designed to replicate space conditions.
On Earth, we have three heat transfer modes. In space, only radiation works.
And radiation at low temperatures is extremely weak.
Stefan-Boltzmann: T⁴ Is Your Enemy
$$q = \varepsilon \sigma (T_s^4 - T_\infty^4)$$
Doubling temperature from 300K → 600K reduces radiator area by 16×
The T⁴ law means at GPU operating temperatures around 350K, we only get ~1500 W/m².
That's terrible. A 100 MW facility needs 65,000 m² of radiator at 350K.
Radiator Scale Visualization
Power Temp Area [m²] Mass [tonnes] Equivalent
10 MW 350K 6,529 33 Football field 100 MW 350K 65,290 326 12 football fields 1 GW 350K 652,900 3,265 Rhode Island area 1 GW 600K 75,640 378 Large campus
Key insight: The GaN revolution (600K compute) transforms the radiator from "impossible" to "challenging."
Without it, 1 GW orbital compute requires a radiator the size of a small state.
Look at these numbers. At 350K, a 1 GW facility needs 653,000 m² of radiator.
That's 65 hectares. In orbit. Made of metal. At $50/kg.
The GaN scenario at 600K brings this down to 75,000 m² — still enormous but not insane.
Liquid Droplet Radiators: The Mass Hack
$$\tau = \frac{\rho c_p r}{9\varepsilon\sigma}\left(\frac{1}{T_c^3} - \frac{1}{T_h^3}\right)$$
Advantages
5–10× lighter than heat pipe panelsNo structural mass (droplets ARE the radiator)
Scales linearly with power
Challenges
Fluid loss in microgravity
TRL 4–5 (not flight proven)
Rayleigh breakup instability
LDRs spray tiny droplets into space. They cool radiatively during flight, then are recollected.
200-micron DC-705 droplets cool from 400K to 350K in about 3 seconds over 6 meters.
This is 5-10x lighter than heat pipes. But catching droplets in zero-g is hard.
Solar Advantage: 3.3× Capacity Factor
Metric Terrestrial Orbital SSO Factor
Irradiance ~1,000 W/m² (peak) 1,361 W/m² +36% Capacity Factor 20–30% 90–99% 3.3× Battery Storage Critical ($100/kWh) Minimal −95% Amortized $/kWh $0.05–0.15 $0.001–0.01 −97%
This is the one area where orbit genuinely wins. Solar power in a dawn-dusk SSO is nearly free after launch.
Here's where the pitch is actually right. Orbital solar at dawn-dusk SSO is incredible.
Nearly 100% capacity factor, no weather, no night cycle. But you still have to GET there.
The Drag Kill Shot
$$F_D = \frac{1}{2} C_D \rho v^2 A \qquad \dot{m} = \frac{F_D}{g_0 I_{sp}}$$
50 kg/day
Propellant @650km
At 100 MW baseline: 1,800 tonnes/year of propellant at $360M/year
The giant solar arrays create enormous drag at 650km. Even with a high-Isp ion thruster at 3000s,
a 100 MW facility burns through 1,800 tonnes of propellant per year.
That's $360M/year just to stay in orbit. Every year. Forever.
Launch Vehicle Comparison
100 MW facility = 1,605 tonnes. Everything depends on Starship at $200/kg.
Look at this chart. At Falcon 9 prices, just launching the hardware costs $4.4 billion.
At Starship prices, it's $321 million. The entire business case hinges on $200/kg launch.
Mass Budget Breakdown
Radiators dominate the mass budget at 350K. 653 tonnes of radiator panels.
Solar arrays are next at 272 tonnes. The actual compute hardware is 400 tonnes.
You're launching more cooling infrastructure than compute.
Radiation: Bits Flip in Space
$$P_\text{fail} = 1 - (1 - p_\text{SEL})^{N} \cdot (1 - p_\text{TMR})^{N}$$
Mitigation stack: COTS-Capsule + memory scrubbing + checkpoint/restart + current limiting + watchdog timers
200,000 COTS GPUs in LEO. Each has a 0.36% daily probability of single-event latchup.
That's 720 chip failures per day. With 4-hour MTTR, uptime drops to ~88%.
Requires a 5-layer mitigation stack.
Formation Flight: The Dancing Constellation
Hill-Clohessy-Wiltshire relative motion in the LVLH frame at 650 km SSO
The facility isn't one monolithic station. It's a formation of modules flying in precise
relative orbits governed by the HCW equations. Deputy satellites orbit the chief in
characteristic ellipses. Free-space optical links maintain connectivity.
TEA Dashboard
This is the full interactive TEA dashboard. You can adjust power, launch cost,
radiator temperature, discount rate, and lifetime. All metrics update in real time.
Note how sensitive the model is to launch cost and radiator temperature.
TCO: Terrestrial Wins Today
$7.1B
Terrestrial 10yr TCO
Bottom line: orbital compute at 100 MW costs 1.6x more than terrestrial over 10 years.
That's actually closer than many people expect. The premium comes from launch, propellant, and insurance.
Monte Carlo: The Uncertainty Envelope
Click "Run Simulation" to execute 10,000 Monte Carlo trials across 8 uncertain parameters.
The NPV distribution is heavily left-skewed. Probability of positive NPV is low.
Launch cost and GPU cost dominate the sensitivity.
What-If A: GaN Revolution
If GaN/SiC enables 600K compute:
Radiative power: 1,532 → 13,228 W/m²
Radiator area: −94%
Radiator mass: 326t → ~38t
Total mass: −40%
TCO reduction: ~15%
But:
GaN AI accelerators at TRL 3–4
Need 600K junction temperature
No HBM equivalent at 600K
GPU cost premium ~50%
Timeline: 10–15 years
Verdict: Transforms feasibility from "impossible" to "challenging." But 10+ years away.
The GaN revolution is the most important wild card. If we can compute at 600K,
radiator mass drops by 94%. This transforms the entire mass budget.
But GaN AI chips are TRL 3-4. We're at least a decade away.
What-If B: Launch Stagnation
→
$2,408M
Launch @$1,500/kg
If Starship fails to achieve $200/kg, orbital compute is dead on arrival .
Remaining use cases at $1,500/kg: sovereign military clouds, decision-cycle compression, on-orbit sensor processing
If Starship fails, launch costs stay at Falcon Heavy levels: $1,500/kg.
The 100 MW launch bill goes from $321M to $2.4B. Just for launch.
Annual propellant resupply at $1,500/kg is $2.7B. Game over.
What-If C: Grid Collapse
Scenario: 10-year moratorium on new terrestrial datacenter construction.
AI demand continues exponential growth. Electricity prices spike to $0.15/kWh.
Required external financing: ~$5 trillion/year.
The only scenario where orbital compute becomes politically viable.
Not economically optimal — politically necessary. National infrastructure project scale.
This is the doomsday scenario for terrestrial compute. If regulations or grid constraints
halt datacenter construction, orbital compute becomes a national priority.
It's never the cheapest option — but it might be the only option available.
Feasibility Map
This contour map shows the TCO ratio as a function of radiator temperature and launch cost.
The green region (ratio near 1) is where orbital compute is competitive.
Notice you need both low launch cost AND high radiator temperature to get close to parity.
What Space IS Good For
Look at the value density axis. ZBLAN fiber: $100k-1M per kg.
Pharma crystals: $1M-100M per kg. Orbital compute: $5 per kg per year.
Space manufacturing wins by 4-5 orders of magnitude on value density.
Value Density: Compute vs Manufacturing
Product Value [$/kg] Launch Cost Ratio Verdict
ZBLAN Fiber $100k–1M 500–5,000× Strongly viable Pharma Crystals $1M–100M 5,000–500,000× Best case Semiconductors $10k–100k 50–500× Viable Bioprinted Organs $100k–10M 500–50,000× High potential Orbital Compute ~$5/yr 0.025× Not viable
Space manufacturing: 1,000–10,000× better value density than compute
This is the definitive comparison. Every zero-gravity manufacturing product generates
orders of magnitude more revenue per kilogram launched than compute hardware.
The fundamental problem: compute is mass-intensive, manufacturing products are mass-efficient.
Risk Matrix
Top right quadrant: radiator scaling and launch cost stagnation are the critical risks.
Both are high likelihood AND high impact. These are the physics and economics constraints
that make orbital compute fundamentally challenging.
Technology Roadmap
Milestone Timeline Impact
Starship operational ($200/kg) 2025–2027 Enables business case LDR flight demo (TRL 6) 2026–2028 5× radiator mass reduction GaN AI chips (TRL 5) 2028–2032 600K operation, 32× cooling 10 MW orbital pilot 2030–2035 Proof of concept 100 MW commercial facility 2035–2040 First revenue
Critical path: GaN + Starship. Both must succeed for 100 MW to be viable.
The roadmap has clear dependencies. Starship must achieve $200/kg.
LDR must reach TRL 6. GaN must reach TRL 5. All three are necessary.
The earliest realistic 100 MW facility is 2035-2040.
Five Key Findings
Physics allows it — T⁴ radiation is weak at 350K but viable at 600K with GaNEconomics kills it (today) — 1.6× TCO premium; negative NPV at current parametersLaunch cost is the bottleneck — Everything depends on Starship achieving $200/kgSpace excels at manufacturing — ZBLAN, pharma crystals are 1000× better $/kg than computeGrid collapse is the wild card — The only scenario making orbital compute politically necessary
Five findings. Physics says possible. Economics says not yet.
Launch cost is the single biggest lever. Space manufacturing beats compute by 1000x.
And grid collapse is the only scenario that could force orbital compute.
Where to Invest Instead
ZBLAN Fiber
$1–10B market, TRL 5.5
Pharma Crystals
$5–50B market, TRL 6.5
GaN R&D
Enables 600K compute
Don't put datacenters in space. Put factories in space.
The conclusion: invest in space manufacturing, not space compute.
Fund GaN research. Watch Starship costs. And wait for the numbers to close.