The physics doesn't rule it out. The economics depend on launch prices falling to a few hundred dollars per kilogram, roughly 20 to 35 times cheaper than today's rideshare prices of about $6,000–7,000/kg. That is the range fully reusable rockets are aiming for. It's a plausible 2030s scenario, not a certainty, and launch price is only one of several costs that have to come down.
Why anyone wants to do this
A large AI training cluster now draws hundreds of megawatts. On the ground, the limiting factor is less often chips and more often grid connections, power plants and permits. A satellite in the right orbit has none of those problems.
In a dawn–dusk sun-synchronous orbit, a spacecraft rides the line between day and night and stays in sunlight almost all the time. There are no clouds, no night and no seasons. The sunlight is also stronger, because none of it is filtered by the atmosphere.
Put together, a solar panel in that orbit can produce several times more energy per year than the same panel on the ground. Google's Project Suncatcher team cites a figure of up to 8×. Space doesn't make energy free, but it does make it very abundant per kilogram of panel.
The catch: getting rid of heat
Every watt a GPU uses becomes a watt of heat. On Earth, air and water carry it away. In a vacuum there is no air and no water, so the only way to dump heat is to radiate it as infrared light. The physics is the Stefan–Boltzmann law:
P = ε · σ · T⁴ · A
A good radiator (emissivity ε ≈ 0.9) at 27 °C (300 K) sheds about 410 watts per square meter from each face. If it radiates from both faces, that's roughly 830 W/m². So every megawatt of compute needs on the order of 1,200 m² of radiator, about a sixth of a football pitch, before you count heat soaked up from Earth and the Sun.
Running hotter helps a lot because of the T⁴ term. At 77 °C, the same area sheds about 85% more. Chips don't like being that hot, though, so radiator size is a real engineering trade-off, not a dealbreaker.
Radiation
Above the atmosphere, chips are hit by high-energy protons and cosmic rays. These flip bits and slowly damage transistors. Space-rated processors are hardened against this, but they run years behind commercial AI chips.
The surprise of the last two years is that ordinary AI hardware may be tougher than expected. Google reported that its Trillium TPUs survived proton-beam doses well above what a five-year low-orbit mission would deliver, with the high-bandwidth memory being the most sensitive part. Starcloud flew an Nvidia H100 in November 2025 to test the same question in real orbit, and Google's first Suncatcher prototype, launched in October 2026, is now gathering in-orbit data on how its TPUs cope with radiation and temperature extremes. If those results hold up in orbit, error-correcting software and modest shielding may be able to handle much of the problem.
Moving the data
Training a frontier model means thousands of chips exchanging data constantly. In orbit, that has to happen over laser links between satellites flying in tight formation, with the data center spread across many spacecraft rather than housed in one building. Google's design flies clusters of satellites only hundreds of meters apart, and lab demonstrations have reached terabit-per-second links.
Latency to the ground is not the problem. Low Earth orbit is only a few milliseconds away. The harder issue is maintenance: nobody is going to swap a failed GPU at 650 km, so systems need to be designed to degrade gracefully as parts fail.
The number that decides everything
The calculator below uses 30 tonnes per megawatt as its starting mass. This is our own back-of-envelope estimate, not a published figure, built from three parts:
- Solar arrays, about 8 t. 1 MW at roughly 120 W/kg, typical of today's flexible roll-out arrays.
- Radiators, about 7 t. The ~1,200 m² worked out above at roughly 6 kg/m².
- Compute servers, about 13 t. An 8-GPU H100 server weighs about 130 kg and draws about 10 kW, or roughly 13 kg per kW.
That adds up to about 28 t, which we round to 30 t to allow for structure, wiring and networking. Purpose-built space hardware could be much lighter, so the calculator lets you change it. At today's rideshare prices of roughly $6,000–7,000 per kilogram, launching 30 t costs about $180–210 million per megawatt.
For the ground side, the model assumes about $10 million per megawatt to build the facility, excluding chips, plus electricity. Treat this as a model assumption, not an industry constant. Real costs vary widely with location, redundancy, cooling design and power infrastructure, so it's adjustable too.
Lower the launch price and the picture changes completely. Try it yourself: