Sponsored by
A Falcon Heavy climbs to orbit. The entire orbital-data-center thesis rests on one number — the cost of lofting a kilogram to space. Photo: SpaceX (public domain).
Here is a sentence that would have sounded absurd two years ago and is now a funded business plan: the cheapest place to run an AI data center might be space. The logic is seductive. The terrestrial grid is hitting a power wall, permits take years, and water for cooling is scarce. In orbit the sun never sets, there is no land to zone, and the vacuum is free. So a wave of companies has started launching computers.
In November 2025, a startup called Starcloud put the first Nvidia H100 GPU into orbit on a sixty-kilogram satellite and trained a small model in space. Days later Google announced Project Suncatcher, a plan to fly its TPUs on solar-powered satellites. SpaceX unveiled an orbital compute satellite of its own, Jeff Bezos started talking about gigawatt data centers in space, and China launched the first twelve of a planned 2,800-satellite computing constellation. The pitch is real and the names are serious. The question is whether the economics survive contact with physics.
I. The whole ballgame is the rocket
Strip away the rhetoric and orbital compute is a bet on one number: the cost to launch a kilogram into low Earth orbit. Today a Falcon 9 runs about $2,720/kg and a Falcon Heavy about $1,450/kg. The entire orbital data-center thesis rests on SpaceX's Starship dragging that down toward $100 to 200/kg, a roughly twenty-five-fold drop. Google, in its own Suncatcher paper, names the threshold explicitly: launch has to reach about $200/kg before orbital compute amortizes to terrestrial energy costs, something it projects for the mid-2030s, not now.
Launch cost to low Earth orbit. Everything about orbital data centers depends on Starship hitting a price that does not yet exist. The dashed line is Google's stated parity threshold. Source: SpaceX, Google. Chart: Silicon & Steel.
II. What space actually fixes
To be fair to the idea, the energy advantage is genuine and large. In a dawn-dusk orbit a solar panel sees the sun almost continuously, and space-grade cells are more efficient than the terrestrial kind, so each panel can deliver something like 5 to 8 times the annual energy of one on the ground. Starcloud's own model claims an effective electricity cost near $0.003 per kilowatt-hour against roughly $0.05 for US industrial power. The sunlight, in other words, really is nearly free.
Most AI newsletters summarize headlines. MavSource is different.
We aggregate updates from all major AI newsletters, podcasts, company news, AI labs, public and private company activity, GitHub projects, funding rounds, earnings calls, and investor letters — hundreds of sources every day. Then we summarize what matters, analyze emerging trends, and add our own founder commentary so you understand why a development may matter — not just what happened.
One 5-minute email, every morning. Built for investors, founders, and operators who want to understand AI as a business, technology, and market trend — not just another news cycle.
The daily email is free. It's also the entry point to a deeper intelligence product covering watchlists, public-company read-throughs, startup trackers, and investor-letter tracking.
The one line where orbit wins decisively: energy. The catch is that capturing this free sunlight requires launching everything else. Source: Starcloud, EIA. Chart: Silicon & Steel.
And it is not only the energy bill. In orbit there is no land to permit, no community to object, no water to evaporate for cooling, and no multi-year interconnection queue, the exact frictions strangling terrestrial AI buildouts right now. If the only thing that mattered were energy and permits, space would win in a walk.
III. Who is actually up there
The reality check is in the scale. Today's orbital compute is single GPUs and storage demos, not data centers. Starcloud-1 carried one H100 and trained a Shakespeare-sized model. Google's Project Suncatcher is two prototype satellites with Planet Labs by early 2027. SpaceX's AI1 satellite is specced at 120 kilowatts, with first units targeted for late 2027. China's Three-Body Computing Constellation has 12 of 2,800 planned satellites flying. Lonestar put a storage box near the Moon. Every megawatt-scale claim, Starcloud's 5 gigawatt cluster, Bezos's gigawatt data centers, is rhetoric about the 2030s.
IV. Who this favors, and who should stay sharp
V. Glossary
Low Earth orbit (LEO): the band roughly 160 to 2,000 km up where these satellites fly, typically around 600 km. Closer means cheaper to reach and lower latency, but more atmospheric drag.
Dawn-dusk sun-synchronous orbit: a path along the day-night line that keeps a satellite in near-continuous sunlight, the trick behind the 24/7 solar claim.
Radiative cooling: in vacuum there is no air or water to carry heat away, so the only way to cool a chip is to radiate the heat as infrared. This is the binding physical constraint.
$/kg to orbit: the launch price per kilogram of payload, the single dominant cost lever for everything in space.
PUE (power usage effectiveness): total facility power divided by useful compute power; orbital advocates claim near 1.0 because there are no chillers, against about 1.1 to 1.2 for good terrestrial sites.
Total ionizing dose (TID): the cumulative radiation damage chips take in orbit; high-bandwidth memory is the most sensitive part and a key limit on hardware life.
VI. Further reading
So the story for free readers is simple: the energy advantage is real, the launch and cooling problems are bigger, and parity is a next-decade bet on Starship. Below the line, the part that took the most work, is the actual model: what a megawatt of orbital compute costs against a megawatt on the ground, line by line, and where exactly the math breaks.
For the next 30 days, all premium content is free.
If you know someone who wants to understand what's really happening in AI and the supply chains powering it, send them this newsletter.
