GPUs in Space are dumb, right?

In February 2026 Elon Musk told Dwarkesh Patel that within 30 to 36 months “the most economically compelling place to put AI will be space.”¹ The whole question reduces to one number. Let’s run the math, shall we?

Solar

A solar panel in sun-synchronous orbit always gets perfect sunlight, no atmosphere, no night, no clouds, no seasons to plan around. On the ground, even if you put them down in a desert, you only get a fraction of that figure once you account for the day-night cycle, suboptimal weather, atmospheric losses, etc. Just doing that math, orbit wins by about 5x, and by roughly 8x against a mid-latitude site.

Storage

Running GPUs from solar on the ground 24/7 means you gotta produce the power needed at night during the day and, more important for cost, store it until then. A desert site delivers its day’s energy in a few peak-equivalent hours, so you have to oversize the panels and store ~12 to 18 hours of load in batteries. Even a normal low orbit spends only ~35 of each ~90 minutes in shadow, and the dawn-dusk orbit nearly eliminates that too. Here’s what that looks like per kilowatt of continuous load:

	Ground (desert)	Terminator orbit
Solar oversizing	~4x	~1x
Storage	~15 kWh	~0.5 kWh
Battery cost (~$300/kWh)	~$4,500	~$150

That yields an ~$4,000/kW advantage for orbital, which is actually quite decent.

Launch cost

To get our datacenter into space, we need to pay for the launch. Per kilowatt, that cost is just the system’s specific mass (kg/kW) times the launch price ($/kg), and it swamps everything else. Every orbital LCOE figure floating around, from $2/MWh to ~$900/MWh, is that same product with a different launch price plugged in.⁴ None of them is an independent result. The only question that matters is how cheap launch has to get to beat the ground.

Ground 24/7 solar-plus-storage is ~$40 to $50/MWh, about $5,000/kW of lifetime cost over 15 years. Subtract ~$1,500/kW for orbital hardware, leaving ~$3,500/kW for launch. Divide by system mass:

System mass	Break-even launch price
7 kg/kW (cells only)	~$500/kg
15 kg/kW (optimistic full system)	~$235/kg
30 kg/kW (realistic integrated)	~$115/kg
60 kg/kW (conservative)	~$60/kg

Current Starship estimates sit at $500 to $1,500/kg. Musk’s 2028 target is $100/kg. So the whole outcome hinges on two numbers, system mass and launch price. Get near 15 kg/kW and a few hundred dollars per kilo, and orbit becomes competitive.

Cooling

System mass is set mostly by cooling. The 7 kg/kW “cells only” figure doesn’t count any way to shed heat, and in space heat only leaves by radiation. A radiator near room temperature rejects ~245 W/m² (εσT⁴ at 20°C), and every watt of compute has to exit through it somehow:

Radiator	Mass per kW of heat
Thin-film (aspirational)	~1 kg/kW
Aluminium (conventional)	~8 kg/kW

That 7 kg/kW gap is roughly the difference between the 15 and 30 kg/kW rows above, or between a ~$235/kg and a ~$115/kg break-even.

Shipping the GPUs is cheap: an H100 is only ~3 kg/kW of silicon, a few percent of its purchase price even at $1,000/kg. Radiation, long assumed fatal, turns out to be a solved problem in practice. Google’s Trillium TPU survived 15 krad(Si) with no hard failures.³ Starcloud’s orbital H100 ran inference and training without crashing.² ECC overhead comes in around ~5 to 15%.

Throughput

Even if launch cost pans out the way we need it to, throughput still won’t: Musk’s 100 GW/yr at 30 kg/kW works out to 3 billion kg headed to orbit every year. Divide that by a 100-ton Starship-class launch and you need 30,000 launches a year, roughly 80 a day. Compare that to SpaceX’s entire 2025 cadence of 165 launches, total, on the comparatively tiny Falcon 9.⁵

Conclusion

The physics works. Solar and storage favour orbit by a real margin, radiation is solved, and break-even launch prices near $100 to $235/kg aren’t absurd. The workload isn’t an obstacle either: training and batch inference are latency-tolerant, and LEO round-trip latency (~20 to 40 ms) is fine even for interactive serving.

So whether orbit wins comes down to one race. Does launch price fall faster than ground storage does? Orbit’s main advantage is the battery it doesn’t need, and cheap ground batteries erode that from the other side, so it’s a question of which cost curve drops faster. Ground has the head start, and the larger industrial base.

Say launch wins that race anyway. Even then, the cadence ceiling of 80 launches a day for 100 GW/yr caps how fast the installed base can grow. What the math actually rules out is the timeline, not the scale. A large fraction of new AI compute built in orbit in the 2040s is plausible. “Ten times cheaper in thirty months,” and “more compute in orbit than on Earth in five years,” are not.