Will data centers go to space?
Short answer: not this decade, and not at a price that beats the ground. The arithmetic for a single gigawatt takes about two minutes. The full underwriting is linked at the bottom if you want to argue with it.
Not before 2033 on the best day, and never at the price.
One gigawatt of orbital compute, with every technology hitting its target and a third of every Starship flight devoted to it, is 265 launches of hardware that adds up to 50 International Space Stations by mass. With hardware that has actually flown it is 470 launches. The same gigawatt on the ground went from dirt to full power in 24 months in Indiana this year.
Three things the coverage isn't telling you
- The solar panels alone are the whole launch industry. A gigawatt of load needs 1.35 GW of array at beginning of life, which is 3.3 square kilometers of cells. That is three times the solar area of the entire Starlink constellation and about 1,000 times the ISS. Space-qualified arrays have been stuck near 150 W/kg for a decade, so that is 9,000 to 13,500 tonnes and 105 to 160 Starship flights before a single chip goes up. Starlink's factory, the largest spacecraft line ever built, makes about 0.25 km² of array a year. This is 13 years of it.
- SpaceX's own number is below anything that has flown. The January FCC filing for one million satellites says one million tonnes per year yields 100 GW. That is 10 kilograms per kilowatt. My best case, with every subsystem at its research target, is 22.5. Hardware that exists is about 40. Ten kg/kW needs 500 W/kg arrays and 2 kg/m² radiators, plus a server rack at 2 kg/kW, all at once, and none of those has flown. It also needs 10,000 Starship flights a year, about 30 times every orbital launch on Earth in 2025.
- The ground isn't slow. The grid is. Amazon's New Carlisle campus passed 1 GW in March 2026, roughly 24 months after ground-break. The thing that takes 4 to 7 years is a utility interconnection in Northern Virginia or Phoenix, and the industry is already routing around it with on-site gas in 12 to 18 months. The best day in orbit is seven years out. The honest version of the space pitch is that Earth's constraints are political and orbit's are engineering. That is a real argument. It is not a 2030 argument.
What the claim asks you to believe
The best case isn't a forecast. It is a conjunction. Five separate industrial curves each have to move roughly an order of magnitude, in the same five-year window, before the first gigawatt exists. Their track records:
| Curve | Today | Best case needs | Move | History says |
|---|---|---|---|---|
| Starship cargo flights per year | 2 so far in 2026, 5 in 2025 against 25 approved | 300+ | 100× | Falcon 9 took six years from a reliable vehicle to 100 a year |
| Solar array W/kg, whole wing | 100 to 150 | 150 to 500 | 1.5 to 5× | One doubling since JPL's 1985 program; its 300 W/kg goal is still unmet |
| Radiator kg/m² at 60°C | 15 on the ISS, near 0°C | 3 | 5× | No pumped-loop radiator has flown at that mass or that temperature |
| Space array production, km² per year | 0.25 | 1 to 2 | 4 to 8× | Starlink's line is already the largest in history |
| Largest self-deploying spacecraft | iROSA, 20 kW | 3.8 MW per node, 265 in formation | 190× | Nothing hectare-scale has ever deployed itself |
Give each curve a generous 50% chance of landing on schedule and the joint probability is 3%. Give each the 20 to 30% that aerospace programs actually post, and it is under half a percent. Starship itself flew 5 times in 2025 against an FAA cap of 25. The lunar landing it is contracted for was due in 2024 and is now 2028. Every input to the best case has already slipped at least once.
SpaceX owns the only spacecraft mass-production line on the planet: about 2,500 Starlinks a year, 1,800 tonnes of satellites with arrays and thrusters already coming off it. A compute node derived from Starlink needs a bigger factory, not a new one. Launch cost has fallen 97% in two decades and the Starship internal target is under $10M a flight. Orbit has no interconnection queue and no electricity bill, which is worth about $0.6B a year against a terrestrial gigawatt. Google's own conditional projection puts orbit cheaper than the ground in 2035, and that date sits inside my range. The strategic case is real and it is sized on slide 22: it survives at megawatts, for sensor data processed where it is collected, and it does not rescue the gigawatt. I put a gigawatt-class cluster before 2033 at under one chance in ten, before 2040 at roughly one in three, and I put real weight on it never mattering, because terrestrial cost curves keep moving while orbit catches up.
Three dates that matter more than anything else
First attempt to put operational Starlink V3 satellites into orbit and catch the ship. Flight 13 deployed 20 of them on a suborbital arc. Until a Starship delivers a payload to orbit and comes back, every cadence curve above is a drawing.
Blackwell hardware in orbit with 100 times the power of the first satellite. Watch the thermal numbers, not the training demo. A radiator specification in kg/m² at operating temperature would be the first real datapoint anyone has published.
Twenty-five or more Block 3 cargo flights in 2027 puts Starship on the doubling curve and moves my dates in. Fewer than ten puts it on the Falcon 9 curve and 1 GW lands around 2040. This one number does more work than every announcement combined.
If you're going to bet on it anyway
- Don't count demonstrators as capacity. A GPU in orbit is a payload. The unit that matters is tonnes per kilowatt, and the first company to publish one under 15 kg/kW for a flown node has said something new.
- Watch the two numbers nobody quotes: the W/kg of a flown array wing above 200, and any radiator flown under 4 kg/m² at over 50°C. Those two move the launch count more than anything Starship does.
- Treat SpaceX's 10 kg/kW and Musk's "cheapest compute within two to three years" as targets, the way SpaceX's own S-1 does when it lists orbital data centers as something that may not prove viable.
- Size it the way the SpaceX position was sized: a venture bet you can lose entirely. The payoff shape is one good outcome in several, and the good outcome is late.
The full version
Everything above comes from a 25-slide underwriting: the mass ledger subsystem by subsystem, the launch and cadence model, why the best case is a conjunction, the delay reference class with numbers, the terrestrial baseline from Epoch's model and this year's builds, the cost and time comparison, the strategic case sized, the bull case argued properly, and the tripwires with dates.
Open the full deck →Best on a laptop. It's built as 1920×1080 slides. Arrow keys to navigate. Every hard number is sourced on the slide it appears on, and anything I estimated is labelled as an estimate.
Prepared 26 August 2026. Companion to Should you buy SpaceX?, where "orbital data centres becoming an actual market" sits in the bull case; this is the underwriting of that line. Same chips on both sides of every comparison. This is my own analysis shared between friends. It is not personalised investment advice and I am not a licensed investment adviser. You should make your own decision, ideally after arguing with mine.