Here is the pitch, and you have to admit it is a good one. The planet is running out of cheap places to put the computers that run artificial intelligence. They eat too much power, drink too much water, and the towns next door keep saying no. Space has none of those problems. The sun never sets in the right orbit, there is no water bill, and nobody files a zoning complaint in low Earth orbit. So move the data centers up there.
That is the idea now drawing real money and serious engineers, and as Digital Journal’s Tim Sandle put it this week, it is moving “from speculative fiction into strategic planning,” Digital Journal reported. The question is whether it is the next frontier or an expensive detour. The honest answer is that it is a little of both, and which one depends almost entirely on a single number that has not come down yet.
Start with what is real, because this is no longer a slide deck. A startup called Starcloud put an Nvidia H100 GPU into orbit in November 2025, then trained a small language model in space the following month, the first time that had been done, and raised $170 million more in March, TechCrunch reported. It is not alone. Google is running an orbital data center effort called Project Suncatcher with the satellite maker Planet Labs, CNBC reported. Blue Origin filed with regulators to launch 51,600 compute satellites, CNBC reported. And SpaceX, the company that just minted Elon Musk the world’s first trillionaire on the strength of bets exactly like this one, asked the U.S. government for permission to build and operate up to one million satellites for computing in space, TechCrunch reported.
So the smart-money interest is genuine. Now the physics.
The first wall is heat. On Earth you cool a rack of screaming-hot chips by blowing air or running water past them. In space there is no air and no water to move, so the only way to dump heat is to radiate it away, slowly, off big panels, Digital Journal reported. The cruel irony writes itself. The thing that makes AI valuable is its appetite for dense, high-performance computing, and dense, high-performance computing is exactly what is hardest to keep cool in orbit. Starcloud’s answer is a radiator bigger than anything flown commercially, and its long-term renderings show solar and cooling panels spanning four kilometers, per Nvidia’s account. That is not a satellite. That is a structure.
The second wall is the neighborhood. Orbit is a shooting gallery of radiation, micrometeoroids and debris, and a single bad impact can kill a platform and create more debris for everyone else, Digital Journal reported. On Earth a technician swaps a dead server in an hour. In orbit you either build hardware to survive years untouched, invent routine robotic repair, or accept that you are flying disposable computers, Digital Journal reported. None of those is cheap, and chips go obsolete in space on the same brutal schedule they do on the ground.
Which brings us to the number that decides everything: the cost of getting a kilogram to orbit. The whole orbital case rests on launch getting radically cheaper. Google’s own Suncatcher paper says orbital running costs only start to rival a terrestrial data center’s energy bill if launch falls below about $200 per kilogram by the mid-2030s, CNBC reported. A Falcon 9 today runs around $2,700 per kilogram, Introl reported. So the bull case is not “space is cheaper.” It is “space will be cheaper, once a rocket that mostly does not exist yet flies routinely.” That rocket is SpaceX’s Starship, and Musk has a long record of being right about where the line goes and wrong about when it gets there.
The skeptics are not hand-wavers, either. Andrew McCalip, an engineer at Varda Space Industries, built a public calculator for this exact question, and the base case comes out to orbital compute costing roughly three times as much per watt as the equivalent on Earth, Introl reported. Even Jeff Bezos, who wants this future, called the aggressive timelines optimistic. Some people talk about two or three years, he told CNBC; that is “probably a little ambitious,” CNBC reported. One analyst pegs the real break-even somewhere between 2030 and 2035, Network World reported. In Wall Street terms, that is not a quarter away. That is a decade away.
Here is the part the bigger headlines tend to skip, and it is the most important part. The near-term use for a data center in space is not running your email or your chatbot. It is serving things that are already up there: processing Earth-observation imagery, crunching data from scientific and defense satellites, feeding compute to other spacecraft, Digital Journal reported. The first real customers for space computers are in space. That is a smaller, stranger, far more credible business than “we are moving the cloud to orbit,” and it is the framing Sandle rightly calls more honest than the market rhetoric.
So how should you hold all this? The way you hold any frontier bet. The long-run argument for orbital data centers is not that space suddenly got easy. It is that Earth keeps getting harder. Land, water and power for ground-based AI all face rising costs and rising public anger, while launch costs may fall, CNBC reported. Two lines on a chart, one drifting up and one maybe drifting down, and the entire trillion-dollar thesis lives at the point where they cross. Nobody knows the year. Some of the people selling you the future are quietly putting it in the next decade, not the next election cycle.
Which loops us back to that record-breaking IPO. When SpaceX investors paid up for “orbital data centers” as part of the story, this is what they bought: a real technology, a real race, a real set of physics problems that will take real money and years to beat, and a payoff date that keeps the word “eventually” attached. It might be the cheapest place in the universe to run AI someday. Someday is doing an awful lot of work in that sentence.
