Elon Musk has proposed launching AI datacenters into orbit, powered by solar arrays. He claims space will become the cheapest location for AI compute within 30-36 months.12 The physics of orbital solar are sound—roughly 5× the energy capture of terrestrial panels12 with no weather, no night cycle (in sun-synchronous orbits), and radiative cooling is at least plausible.
But if you do a complete thermodynamic accounting, a structural constraint emerges: the plan requires an enormous amount of Earth's electrical grid just to get off the ground. Manufacturing the panels alone consumes grid electricity on the same order of magnitude as the orbital array's annual output, and none of that energy ever comes back.
This page works through the total energy investment to get one Starship payload of solar panels into orbit, expressed as a fraction of what the array generates per year. It doesn't include any grid capacity for terrestrial AI datacenters (the gigawatts xAI and others are already racing to build on the ground). That demand is additive to everything below.
Every number below is derived from the assumptions panel. Change them and the entire analysis recalculates.
SpaceX's Starship delivers 100 tonnes to low Earth orbit. At 100 W/kg, one launch places 10 MW of solar generating capacity into orbit.
At a 55% capacity factor (LEO), this array produces 48,180 MWh in its first year, equivalent to running a 5.5 MW plant 24/7.
Getting those panels into orbit costs energy in three ways:
Producing solar panels (mining quartz, purifying polysilicon at >2,000°C, crystal growth, wafer sawing, cell fabrication) consumes 4.5 kWh for every watt of capacity. Nearly all of this is grid electricity.
The full Starship/Super Heavy stack carries 4,900 tonnes of propellant: 3,870 t of LOX and 1,030 t of liquid methane.7 Burning that methane releases 14,343 MWh of chemical energy. This is a fixed cost per launch regardless of payload, and it's a directly extracted resource. It doesn't show up in the grid bill unless propellant has to be synthesized.
| Component | MWh | Type | % of Year 1 |
|---|---|---|---|
| Panel manufacturing | 45,000 | Grid electricity | 93.8% |
| Propellant chemistry | 14,343 | Fossil fuel | 29.8% |
| Ground operations | 2,304 | Grid electricity | 4.8% |
| Total | 61,647 | 127.9% | |
| Annual orbital output | 48,180 | 100% |
What matters here is the grid EPBT: how long the orbital array has to run to generate energy equal to what Earth's grid put in. We're excluding the chemical energy of propellant (that comes from fossil methane reserves, not the grid) and just counting what the grid actually supplies: manufacturing energy plus ground operations.
For comparison, modern terrestrial solar panels achieve EPBT of 0.5–1.2 years10. At your current settings, orbital solar's grid EPBT is 0.98 years, comparable to terrestrial solar.
If the panels last 10 years before replacement, Earth must re-invest the full energy budget each cycle. Over 30 years the grid bears this cost 3×, with zero energy return.
Musk states both Tesla and SpaceX have a mandate to reach 100 GW/year of solar cell production.6 He projects eventually deploying hundreds of GW per year into orbit, reaching a terawatt per year.12 At your target of 100 GW/year:
Each launch delivers 10 MW, so 100 GW/year requires 10,000 launches per year, or 1.14 launches per hour, continuously.
Building 100 GW of panels per year at 4.5 kWh/W requires 51.4 GW of continuous grid power just for manufacturing.
For context, average US electricity generation is ~480 GW.11 Manufacturing alone would consume:
Bar represents fraction of US average grid generation (~480 GW)
10,000 launches × 1,030 t methane = 10,300 thousand tonnes per year, or roughly 500 Bcf of natural gas, or 1.4% of US production. Surprisingly plausible given the scope of the project.
Adding ground operations, the total continuous grid draw is 53.0 GW, or 11.0% of US generation capacity. That's about the entire electricity output of a mid-sized industrial nation, dedicated to building hardware for space, with no energy return to Earth.
Musk's thesis is that terrestrial grid capacity is the bottleneck for AI scaling. His solution requires massive amounts of that same bottlenecked grid to manufacture and launch orbital panels.
Every GW/year of orbital deployment requires 0.53 GW of dedicated terrestrial grid power. To deploy at Musk's target of 100 GW/year, you need 53.0 GW of grid capacity that does nothing but build space hardware.
If the US could allocate 50 GW of surplus grid capacity to this effort (a generous estimate, more than the entire output of New York State), maximum orbital deployment would be 94 GW/year. At a more realistic 10 GW surplus: 19 GW/year.
Despite all this, the lifetime energy math works out surprisingly well. Over 10 years, that orbital array generates far more energy than the grid invested to create it:
For every MWh the grid invests, the orbital array delivers 10.2 MWh of compute capacity in orbit over its lifetime. The point isn't dollar savings. The grid spends 47,304 MWh and gets 481,800 MWh of orbital compute in return.
If grid capacity really is what's limiting AI scaling, a 10× multiplier on grid electricity (even one that never sends energy back to Earth) changes the calculus.
Musk's best argument isn't about physics, it's about bureaucracy: no land permitting, no NIMBY opposition, no utility interconnection queues. He's probably right about that, at least in the near term. But orbital compute is routing around Earth's political constraints while remaining entirely dependent on Earth's grid for manufacturing and propellant.
The 10.2× lifetime multiplier is real: the grid trades 47,304 MWh per launch for 481,800 MWh of orbital compute over 10 years. That's a real advantage, but the capacity arrives over a decade, and in a scaling race where the next 2-3 years matter most, energy delivered in year 8 may not matter.
But every GW of orbital compute still requires 0.53 GW of dedicated terrestrial grid to manufacture and launch. At 100 GW/year, that's 53.0 GW, or 11.0% of US generation—building space hardware. Manufacturing alone is % of the energy bill. The conversation about launch cost per kilogram mostly misses the point.
And this is before counting a single watt of terrestrial AI datacenters. The grid capacity above is only for manufacturing and launching space hardware. It doesn't include the gigawatts xAI and others are already building on the ground. That demand is additive.
Here's the uncomfortable question: if you've figured out how to build 53.0 GW of new grid capacity for panel manufacturing, why not just plug datacenters into it? The whole premise is that terrestrial grid is the bottleneck. But the orbital program requires solving that bottleneck anyway, at massive scale, just to feed the factories. If you can do that, the case for launching the panels into space instead of powering compute directly gets a lot weaker.
Orbital solar doesn't solve Musk's energy problem. It's a bet that the 10× multiplier and the permitting bypass are worth the overhead of rockets. Maybe. But the path to space compute still runs through the same terrestrial energy buildout that Musk says is too slow.
The energy budget table in section 2 includes propellant chemistry for completeness, but the grid payback calculation in section 3 strips it out. Chemical propellant comes from geological methane reserves. It's cheap, it's abundant, and it doesn't compete for grid capacity. Grid electricity is the constrained resource in this analysis, so the grid EPBT only counts what the grid has to supply: manufacturing and ground operations.
The "Methane source" toggle above switches to synthetic propellant via the Sabatier process (CO₂ + 4H₂ → CH₄ + 2H₂O). It works, but it converts a non-grid cost into a grid cost: ground operations jump from ~2,300 MWh to ~34,000 MWh per launch. Since the whole argument for orbital solar is that grid capacity is scarce, adding a major new grid consumer for propellant synthesis makes the bootstrapping problem worse.
Yes, in principle. Enough nuclear capacity would solve the manufacturing energy problem, the synthetic propellant problem, and the terrestrial datacenter problem all at once. In practice, new nuclear in the US is constrained by the same permitting dysfunction that motivates the orbital approach in the first place. Musk's 30-36 month timeline doesn't align with nuclear construction timelines.