Start neutral
Both columns start with the same hardware and electrical load. Equal useful output is a baseline assumption, not evidence that orbit performs equally well.
GW
0–100 GW of computing hardware, before cooling or spacecraft overhead.What changes the numbers?
IT capacity and the Advanced assumptions below. Workload-specific performance has not been measured.
Workload considerations — information only
These examples explain what to investigate. They do not change the estimates: there is no verified workload benchmark behind this model.
Independent batch inference: independent requests can be grouped. Communication losses still need measurement; no numeric advantage is assumed.
Interactive inference: response time and ground-link availability matter. Annual throughput does not predict latency or service-level compliance.
Closely coupled training: frequent accelerator exchanges make networking critical. A simple multiplier does not establish whether a training job can run at all.
For a specific workload, enter measured relative efficiency, availability and communication losses under Advanced assumptions. Without those measurements, treat the numbers as scenarios.
Advanced assumptions: useful work, energy and cost
All starting numbers are neutral or illustrative. There are no automatic workload-specific performance factors. For real predictions, replace them with measured values for the same model, quality target and hardware.
%
Useful work per IT kWh, before the separate losses below. Earth = 100%. Same-hardware neutral default, not a measured benchmark.%
Fraction of scheduled time delivering service. Earth reference = 100%. Count outages here, not again as communication loss.%
Throughput lost to networking while service is available. Separate from outages; 10 means retain 90%.%
Average draw / installed IT rating. Applied equally to both locations. Idle and failed work can still use power.×
1.2 means 20% more facility energy than IT energy. Assumed constant with load.$/kWh
Illustrative tariff, not a local electricity quote.$/kg
$200/kg is an optimistic future scenario, not an expected current price. Try $2,000/kg.kg/kW
Include computing, solar arrays, radiators, structure and propulsion. Illustrative default.years
Economic life, which may be shorter than physical survival. No financing included.In space
Installed IT electrical capacity
Useful-work equivalentHypothetical 100 MW Earth-output units.
Annual IT electricityIT only; excludes spacecraft overhead.
Annualized launch expense onlyLaunch only, not total cost.
On Earth
The same installed IT electrical capacity
Useful-work referenceSame workload and hardware; no orbital losses.
Annual facility electricityIT electricity plus the selected PUE overhead.
Annual electricity expense onlyElectricity only, not total cost.
No total-cost winner: launch and electricity omit different major costs. Default $200/kg launch is an optimistic future assumption.
Reference units model annual useful work, not real buildings or global compute share. Energy percentages use the IEA’s 950 TWh/year 2030 benchmark.
Explain the results · energy scale · cost limits · sources
What does the useful-work result mean?
One reference unit means the annual useful output of hypothetical 100 MW IT hardware on Earth, at full electrical load, 100% service availability, and the same workload and hardware. It is not the output of a named data center, a count of existing buildings, or a share of global AI computing.
This is normalized output, not tokens/second. Orbital efficiency × availability × communication retention is applied once. The same average electrical load applies to both alternatives. The model assumes useful output scales linearly with average electrical draw; real low-load processor behavior may differ. Output penalties do not automatically save electricity: idle, unavailable or failed work may still consume power.
Energy scale, separately from computing output
The IEA projects 950 TWh of annual data-center facility electricity in 2030. This fixed benchmark puts energy in context; it is not global compute capacity.
0%100% of the energy benchmark
Bars stop at 100%; numerical values remain uncapped.
Neither percentage predicts demand displaced on Earth. Orbital computing could serve new demand instead. The space numerator excludes spacecraft overhead; the Earth numerator includes facility overhead.
Read the cost comparison carefully
No total-cost winner is calculated.
Space costs omitted: spacecraft and chips, communications, ground stations, operations, financing, insurance, failures, spares and disposal. Earth costs omitted: construction, chips, land, grid connections, maintenance and financing. Manufacturing, launch and reentry also have environmental impacts.
Hardware scale and formulas
Let L = average electrical load / 100, and F = (orbital efficiency / 100) × (availability / 100) × (1 − communication loss / 100). Earth reference units = installed GW × 10 × L. Space equivalent units = Earth reference units × F. The full-load Earth IT power that matches this modeled orbital output = installed GW × L × F.
Annual IT energy (TWh) = installed GW × 8.76 × L. Annual Earth facility energy = IT energy × PUE. Launched mass = installed GW × 1,000,000 × kg/kW. Annualized launch expense = mass × $/kg / useful years. Annual Earth electricity expense = facility TWh × 1,000,000,000 × $/kWh. Straight-line annualization, no financing, inflation or cost of capital.
Availability loss describes unavailable service time; communication loss applies only during remaining available time. Do not count the same failure in both. Reliability and networking assumptions change useful output, not installed mass. At zero load there is no modeled useful output, but installed hardware still has a launch cost.
How to use this responsibly
Test a penalty
Try 90% service availability and 10% communication loss. Together they retain 81% of useful work, not 80%, because the second loss applies to the remaining work.
Use measured data next
For a real comparison, measure the same workload, model quality, hardware, power, recovery time and completed useful work. Peak chip performance alone is insufficient.
Space risks and architecture context
Google explores tightly linked TPU spacecraft; SpaceX’s AI1 concept emphasizes localized computing linked through Starlink. Networking demands depend on workload. Radiators, radiation-related errors, storm-driven drag, collision avoidance, safe disposal, cyberattacks and interrupted ground or launch access affect the outcome. No operational-fleet cost advantage is assumed.
Primary sources and limitations
- IEA, Key Questions on Energy and AI (2026): 950 TWh central projection for 2030.
- SpaceX Starmind / AI1: proposed 175 kW average compute-payload rating. Official indexed text checked; direct retrieval timed out. This is a design claim, not a measured fleet rating.
- Google Suncatcher preprint, June 17, 2026: architecture and conditional launch-price analysis, not full economics. $200/kg is a speculative future assumption.
- Google’s orbital test report, October 30: a technical milestone, not established commercial cost savings.
- NASA thermal guidance and ESA 2026 environment report.
Sources checked October 30, 2026. No workload-specific benchmark is supplied. All performance modifiers are user-selected assumptions. The reference Earth system is idealized; a measured Earth baseline may also have downtime and networking losses.