01 Executive Summary
Can we build a hyperscale AI datacenter in orbit? Physics says maybe. Economics says not yet.
This report presents a first-principles feasibility analysis of deploying a 100 MW-class AI training facility in low Earth orbit (LEO). We model every link in the engineering chain — thermal rejection, orbital mechanics, launch logistics, radiation hardening, and economic viability — to answer a single question: can orbital compute compete with terrestrial hyperscale datacenters?
The analysis covers a 650 km Sun-synchronous orbit hosting 200,000 NVIDIA H100 GPUs, cooled entirely by radiative heat rejection. We compare 10-year total cost of ownership (TCO) against an equivalent terrestrial facility and identify the technology thresholds that would need to be crossed for space-based compute to become cost-competitive.
02 Thermal Physics
The T⁴ law governs everything. No convection. No conduction. Only radiation.
On Earth, datacenters reject waste heat through air cooling, liquid cooling, or evaporative towers — all relying on convection and conduction. In the vacuum of space, neither mechanism exists. The only available path is thermal radiation, governed by the Stefan-Boltzmann law. This creates a fundamental scaling problem: radiative power scales as T⁴, meaning that low-temperature radiators are extraordinarily inefficient. A radiator at 350 K (typical silicon junction limit) emits only ~826 W/m² per side, requiring vast surface areas to reject 100 MW of waste heat.
Use the sliders below to explore how surface temperature, emissivity, and power level affect the required radiator size. Note how dramatically the area drops at higher temperatures — this is the core argument for GaN/SiC electronics that can operate at 600 K.
03 Radiator Systems
Liquid Droplet Radiators could cut mass 5–10× — if we can contain the droplets.
Conventional heat-pipe radiator panels (TRL 9) are heavy: 5–10 kg/m² of aluminum honeycomb with embedded ammonia or water heat pipes. For a 100 MW facility at 350 K, this translates to over 300,000 m² and 650+ tonnes of radiator mass alone — nearly half the total system mass. Liquid Droplet Radiators (LDR) offer a radical alternative: spray a sheet of hot liquid droplets into vacuum, let them cool radiatively during flight, then recollect them with an electromagnetic or electrostatic collector. Because the droplet sheet is mostly empty space, LDR specific mass drops to 1–3 kg/kW — a potential 5–10× improvement.
The critical design parameter is the droplet flight time: how long a droplet must travel in vacuum to cool from T_hot to T_cold. Smaller droplets cool faster (higher surface-to-volume ratio) but are harder to collect and more susceptible to solar radiation pressure drift. The equation below gives the exact integrated cooling time for a spherical droplet.
Radiator Technology Comparison
| Technology | Specific Mass [kg/kW] | Temperature Range | TRL | Key Advantage |
|---|---|---|---|---|
| Heat Pipe Panels | 5–10 | 250–400 K | 9 | Flight heritage |
| Liquid Droplet (LDR) | 1–3 | 300–520 K | 4–5 | 5–10× lighter |
| Liquid Sheet (LSR) | 1–2 | 300–500 K | 3–4 | No droplet loss |
| Deployable Panels | 3–6 | 250–400 K | 6–7 | Compact stowage |
04 Orbital Energy & Drag
Sun-synchronous orbit gives 95%+ solar exposure — but drag demands constant propellant.
A Sun-synchronous orbit (SSO) at 650 km altitude provides near-continuous solar illumination with a capacity factor of 90–99%, compared to 20–30% for terrestrial solar. This eliminates the need for massive battery energy storage systems (BESS) and provides roughly 3.3× more energy per square meter of solar array. The orbital velocity at this altitude is ~7,535 m/s.
However, even at 650 km, residual atmospheric drag is a showstopper for large structures. A 100 MW facility requires ~2,700 m² of solar arrays, and the drag force scales with cross-sectional area. This creates a relentless propellant requirement: electric thrusters (Hall-effect or ion, Isp ~3,000 s) must fire continuously to maintain altitude. For a 1 GW facility, drag consumes ~49 tonnes of propellant per day — requiring constant resupply launches. This is the second-largest cost driver after hardware.
Solar Energy Advantage
| Metric | Terrestrial Solar | Orbital SSO | Advantage |
|---|---|---|---|
| Irradiance [W/m²] | ~1,000 (peak) | 1,361 (constant) | +36% |
| Capacity Factor | 20–30% | 90–99% | 3.3× |
| BESS Requirement | Critical ($100/kWh) | Minimal | −95% mass |
| Amortized Cost [$/kWh] | $0.05–0.15 | $0.001–0.01 | −97% |
05 Launch Economics
Everything depends on Starship at $200/kg. At Falcon 9 prices, it's dead on arrival.
The total mass of a 100 MW orbital datacenter is approximately 1,605 tonnes: 400t in GPUs, 272t in solar arrays, 653t in radiators, 200t in structural elements, and 80t in power management. At current Falcon 9 prices (~$2,720/kg), launching this mass would cost over $4.3 billion — before any operational expenses. SpaceX's Starship promises to reduce this to ~$200/kg through full reusability, bringing launch costs down to ~$321 million.
The mass budget reveals the dominance of thermal management: radiators account for 41% of total mass at 350 K. This is why higher-temperature operation (Section 10, GaN Revolution scenario) has such dramatic economic impact — it doesn't just shrink the radiator, it slashes the most expensive single line item in the launch manifest. Use the slider below to see how launch cost per kilogram affects the total facility deployment cost across different vehicles.
06 Radiation & Reliability
200,000 GPUs in LEO. Bits flip. Chips latch. Redundancy is survival.
Commercial-off-the-shelf (COTS) GPUs were never designed for the space radiation environment. In LEO, chips are exposed to galactic cosmic rays (GCR), trapped protons in the South Atlantic Anomaly, and solar particle events. These cause single-event upsets (SEU — bit flips in memory), single-event latchups (SEL — destructive current runaway), and cumulative total ionizing dose (TID) degradation. With 200,000 GPUs, even a small per-chip failure probability translates to hundreds of expected failures per day without mitigation.
The chart below compares five mitigation strategies. "No Mitigation" shows raw fleet attrition from SEL events. "Passive Shielding" (aluminum/polyethylene) reduces proton flux but is largely ineffective against heavy-ion GCR. "COTS-Capsule" encapsulates each GPU in a latchup-current-limiting housing, reducing destructive SEL by ~100×. "TMR + Scrubbing" uses triple modular redundancy (3 copies vote on every result) at the cost of 3× mass and power. "Full Stack" combines all strategies for near-zero attrition. Adjust the sliders to explore different radiation environments and fleet sizes.
Mitigation Strategy Comparison
| Strategy | Mass × | Power × | Latency Penalty | Effectiveness |
|---|---|---|---|---|
| Passive Shielding | 1.0× | 1.0× | 0% | Reduces TID; limited vs GCR |
| Triple Modular Redundancy | 3.0× | 3.0× | 0% | Prevents unrecovered SEU |
| Memory Scrubbing | 1.0× | 1.05× | 5–10% | Fixes SEUs in SRAM |
| COTS-Capsule | 1.1× | 1.05× | 2% | Prevents destructive SEL |
| Checkpoint/Restart | 1.0× | 1.03× | 2–5% | Recovers via state restore |
07 Formation Flight
Hill-Clohessy-Wiltshire equations govern relative motion in the LVLH frame.
A single monolithic 100 MW station would be impractical to launch, assemble, and maintain. Instead, we model a distributed formation of smaller modules (compute pods, radiator arrays, solar power stations) flying in close proximity. The Hill-Clohessy-Wiltshire (HCW) equations describe the linearized relative motion of a deputy satellite around a chief satellite in the Local Vertical Local Horizontal (LVLH) frame. These equations capture the key physics: radial-along-track coupling from orbital mechanics, natural drift ellipses, and the bounded/unbounded motion conditions that determine stationkeeping fuel costs.
The interactive plot below shows the relative trajectory of a deputy module initialized with a radial and cross-track offset. The characteristic "kidney bean" shape is the natural drift ellipse — deputies follow this path passively without thrust. By choosing the right initial conditions, modules can maintain safe separations (50–500 m) with minimal propellant. Adjust the offsets and number of orbits to explore different formation geometries.
08 TEA Dashboard
Multi-parameter economic model. All metrics update live.
The techno-economic analysis (TEA) integrates all physical subsystems into a unified financial model. CAPEX includes GPU procurement ($5B for 200k H100s), solar arrays, radiators, structural mass, launch costs, and a 15% assembly/integration premium. Annual OPEX covers propellant resupply, GPU replacements (5%/yr attrition), insurance (2.8% of hardware value), ground operations ($50M/yr), and data relay ($30M/yr). Revenue is modeled at $2.00/PFLOP-hr with 85% utilization.
The dashboard below links all parameters: changing the radiator temperature affects radiator mass, which changes launch cost, which shifts CAPEX, which cascades through NPV and LCOC. The Levelized Cost of Compute (LCOC) normalizes total discounted costs over total discounted FLOP-hours, providing a single $/PFLOP-hr metric comparable across deployment options. The TCO chart shows cumulative spending over the facility lifetime, with a terrestrial baseline for reference.
09 Monte Carlo Uncertainty Analysis
100,000 trials across 8 uncertain parameters. Run via Web Workers.
Deterministic models give a single answer, but the real world is uncertain. Launch costs could be $100/kg or $1,000/kg. GPU prices fluctuate. Discount rates depend on market conditions. The Monte Carlo simulation draws random samples from probability distributions for each uncertain parameter (see Appendix A for distributions), computes the full TEA for each sample, and builds a statistical picture of likely outcomes. Log-normal distributions are used for launch cost (right-skewed: more likely to exceed targets) while normal distributions capture symmetric uncertainties like GPU pricing.
The simulation runs entirely in your browser using Web Workers for parallel execution. After completion, the absolute NPV histogram shows the distribution of project returns (including revenue). The tornado chart identifies which parameters have the largest impact on outcomes. The Comparative Analysis section below strips out revenue (identical for both options) and shows the pure cost-structure difference between orbital and terrestrial deployment — the metric that actually matters for the build-or-don't-build decision.
Absolute NPV includes revenue ($2/PFLOP-hr at 85% utilization). Both orbital and terrestrial earn the same revenue — see Comparative Analysis below for the cost-structure difference.
Comparative Analysis: Orbital vs Terrestrial
10 What-If Scenarios
Three futures. Only one makes orbital compute viable.
The baseline analysis shows orbital compute is 1.5–3× more expensive than terrestrial. But the future is not static. We model three divergent scenarios to stress-test the conclusion. Scenario A (GaN Revolution) asks: what if wide-bandgap semiconductors enable compute at 600 K? Radiative power density scales as T⁴, so doubling the temperature from 350 K to 600 K yields ~8.6× more power per square meter — slashing radiator mass by ~87%. Scenario B (Launch Stagnation) asks: what if Starship fails and launch costs stay at $1,500/kg? This makes orbital compute 3–5× more expensive than terrestrial. Scenario C (Grid Collapse) models a regulatory moratorium on new terrestrial datacenter construction, forcing AI demand into orbit regardless of cost.
The feasibility contour map below plots the TCO ratio (orbital/terrestrial) across a 2D space of radiator temperature and launch cost. The white contour line marks parity (ratio = 1.0). Everything above and to the right is where orbital compute becomes cost-competitive — requiring both cheap launches (<$50/kg) and high-temperature electronics (>500 K).
11 Zero-Gravity Applications
What should we actually do in space? Hint: not compute.
If orbital compute struggles economically, what should we do in microgravity? The answer lies in value density: revenue generated per kilogram launched. Orbital compute achieves roughly $5/kg/year — because GPUs are heavy and compute is a commodity. By contrast, zero-gravity manufacturing produces materials whose value-per-kilogram justifies launch costs by orders of magnitude. ZBLAN fluoride glass fiber, drawn without convective defects, is worth $100,000–$1,000,000/kg. Protein crystals grown without sedimentation enable drug structures worth $1M–$100M/kg in pharmaceutical R&D value.
The bubble chart below plots each zero-gravity application by Technology Readiness Level (TRL) and value per kilogram, with bubble size proportional to addressable market. Note that orbital compute (gray, bottom-left) is dwarfed by every manufacturing application. Companies like Varda Space Industries (re-entry capsule manufacturing) and Flawless Photonics (ZBLAN fiber) are already pursuing these higher-value opportunities. The table below provides details on each application.
| Application | TRL | Market [$B/yr] | Value [$/kg] | Physics Basis |
|---|---|---|---|---|
| ZBLAN Fiber Optics | 5.5 | 1–10 | $100k–1M | No convection crystallization |
| Protein Crystallography | 6.5 | 5–50 | $1M–100M | No sedimentation; larger crystals |
| 3D Organ Bioprinting | 4.5 | 10–30 | $100k–10M | No gravity collapse |
| Semiconductor Crystals | 5.5 | 2–5 | $10k–100k | Diffusion-only transport |
| Metal Alloy Formation | 4.5 | 0.5–2 | $5k–50k | Containerless processing |
| Perfect Microspheres | 6.5 | 0.1–0.5 | $50k–500k | Surface tension dominates |
12 Risk Assessment
A structured view of what could go wrong — and what already has.
Deploying a hyperscale compute facility in LEO introduces risks far beyond those faced by terrestrial datacenters. The risk matrix below plots 15 identified risks by likelihood (1–5) and impact (1–5). Red markers indicate critical risks (score ≥ 16) that could individually render the project non-viable: radiator scaling, launch cost uncertainty, and space debris impact are the three highest-scoring items. Orange markers are significant risks requiring active mitigation. Green markers are manageable risks with known solutions.
Notable risks include the Kessler cascade scenario (a runaway chain of debris collisions that could render LEO unusable), panel deployment failure (no thermal rejection if radiators don't unfold), and the ASAT/sabotage threat (orbital infrastructure is a strategic target). Click on any risk in the chart to see mitigation strategies and current status.
Click on a risk item in the chart above for details.
A Assumptions & Parameters
All key assumptions underpinning this analysis, organized by domain.
Every model is only as good as its inputs. The tables below document every assumption, constant, and distribution used in this analysis. Where possible, we use established standards (CODATA physical constants, NASA solar flux, WGS-84 geodetic parameters). Engineering assumptions (radiator specific mass, GPU packaging mass, assembly cost fractions) are drawn from published literature and vendor specifications. The Key Simplifications section at the bottom lists modeling limitations that could affect results if relaxed.
Physical Constants
| Parameter | Value | Source |
|---|---|---|
| Stefan-Boltzmann constant (σ) | 5.670374419 × 10⁻⁸ W/m²/K⁴ | CODATA 2018 |
| Solar constant (G_sc) | 1,361 W/m² | NASA standard |
| CMB temperature (T∞) | 2.725 K | COBE/FIRAS |
| Gravitational parameter (μ⊕) | 3.986 × 10¹⁴ m³/s² | WGS-84 |
| Earth radius (R⊕) | 6,371 km | WGS-84 mean |
Orbital Environment
| Assumption | Value | Notes |
|---|---|---|
| Reference orbit | 650 km SSO (Sun-Synchronous) | 98° inclination, dawn-dusk terminator |
| Atmospheric density @650 km | ~1 × 10⁻¹³ kg/m³ | NRLMSISE-00 model, solar moderate |
| Drag coefficient (C_D) | 2.2 | Flat plate approximation for solar arrays |
| Solar capacity factor | 90–99% | Dawn-dusk SSO; eclipse fraction < 5% |
| Specific impulse (I_sp) | 3,000 s | Hall-effect or ion thruster (Xe propellant) |
Compute Hardware
| Assumption | Value | Notes |
|---|---|---|
| GPU model | NVIDIA H100 SXM | Reference accelerator for 2024-era analysis |
| TDP per GPU | 700 W | SXM5 module |
| FP8 throughput | 1,979 TFLOPS | Per NVIDIA spec sheet |
| GPU unit cost | $25,000 | Wholesale pricing, volume discount assumed |
| GPU mass (with packaging) | 2.0 kg | Includes module, heatsink, mounting |
| Number of GPUs (100 MW) | 200,000 | = 100 MW / (700 W × 0.72 overhead) |
| Compute utilization | 85% | Industry standard for cloud GPU |
Thermal Management
| Assumption | Value | Notes |
|---|---|---|
| Radiator temperature | 350 K (baseline) | Limited by silicon junction temps (~105°C) |
| Emissivity (ε) | 0.90 | Z-93P white paint or silverized Teflon |
| Double-sided radiation | Yes | Both faces radiate to space (2× area efficiency) |
| Radiator specific mass | 5 kg/m² | Aluminum honeycomb with embedded heat pipes |
| Thermal power = Electrical power | 1:1 | All electrical energy becomes waste heat |
| LDR droplet fluid | DC-705 silicone oil | ρ=1,070 kg/m³, c_p=1,600 J/kg/K, ε=0.85 |
| LDR droplet diameter | 200 μm (baseline) | Optimized for flight time ~3 s at ΔT=50 K |
Launch & Mass Budget
| Assumption | Value | Notes |
|---|---|---|
| Baseline launch cost | $200/kg to LEO | Starship fully-reusable target price |
| Starship payload | 150,000 kg/launch | Expendable mode to LEO |
| Solar array specific mass | 2,720 kg/MW | Roll-out arrays (ROSA-type), 300 W/m² |
| Structure mass fraction | 200 t per 100 MW | Trusses, docking ports, thermal buses |
| PMAD mass | 80 t per 100 MW | Power management and distribution |
| Propellant resupply | 1,800 t/yr per 100 MW | For drag makeup at 650 km SSO |
Economics & Financial
| Assumption | Value | Notes |
|---|---|---|
| Discount rate | 10% | Weighted average cost of capital (space venture) |
| Facility lifetime | 10 years | Technology refresh cycle for GPUs |
| Revenue model | $2.00/PFLOP-hr | Cloud GPU rental equivalent (H100 pricing) |
| Assembly & integration cost | 15% of (hardware + launch) | Orbital assembly, testing, commissioning |
| Insurance rate | 2.8% of hardware value/yr | Space asset insurance market rate |
| GPU replacement rate | 5% per year | Annual attrition from radiation, failures |
| Ground operations | $50M/yr per 100 MW | Mission control, network operations |
| Data relay | $30M/yr per 100 MW | TDRSS or commercial relay constellation |
Terrestrial Comparison
| Assumption | Value | Notes |
|---|---|---|
| Datacenter construction | $10M/MW | Hyperscale greenfield (Meta, Google benchmark) |
| Electricity rate | $0.06/kWh | US average industrial + PPA |
| PUE (Power Usage Effectiveness) | 1.2 | Modern hyperscale with liquid cooling |
| Land & permitting | $50M per 100 MW | US/EU average for datacenter sites |
| Staff & maintenance | $45M/yr per 100 MW | Operations, security, maintenance |
Monte Carlo Parameter Distributions
| Parameter | Distribution | Range | Rationale |
|---|---|---|---|
| Launch cost [$/kg] | Log-normal(ln(200), 0.5) | ~$74–$540 (P5–P95) | Right-skewed: more likely to be higher than target |
| GPU unit cost [$] | Normal(25000, 5000) | $15k–$35k | Market pricing uncertainty |
| Discount rate | Normal(0.10, 0.02) | 6%–14% | Capital market conditions |
| Facility lifetime [yr] | Discrete {5, 7, 10, 15} | Uniform selection | Technology obsolescence uncertainty |
| Propellant cost [$/kg] | Log-normal(ln(200), 0.5) | ~$74–$540 | Correlated with launch cost |
Key Simplifications
- Atmospheric density model uses piecewise exponential fit, not full NRLMSISE-00 with solar flux variation
- Radiator view factor to Earth/Sun not modeled (assumes free-space radiation to 2.7 K background)
- No orbital lifetime degradation for solar cells (assumed constant η=30% over mission)
- Station-keeping propellant uses continuous drag model, not impulse-based maneuver planning
- All waste heat treated as uniform — no hot/cold side thermal gradients
- GPU performance assumed constant (no clock throttling from radiation or thermal effects)
- Inter-satellite optical links assumed available with zero development cost
- Single-orbit constellation — no relay or ground station handover modeling
- Space debris flux model assumes current LEO density, not projected 2030+ Kessler growth
- Revenue model assumes 100% demand for orbital compute at $2/PFLOP-hr (no market risk discount)