New Delhi: Elon Musk has never been shy about big claims, but his latest one, posted on X, is a stretch even by his standards. “Future super intelligence compute will obviously be overwhelmingly in space,” he wrote, replying to Sequoia partner Shaun Maguire and another user. Then he spent the following hours arguing about it.
The argument is mostly about heat. A critic, @broudsov, questioned whether cooling chips in orbit is as easy as Musk makes it sound. Musk’s reply to that account was a one-line insult. His reply to Maguire was longer and more useful: cooling is “much easier” in space, he said, because an aluminum radiator pointed away from the Sun mostly just sees the cosmic background, about three degrees above absolute zero. He put the figure at roughly 1,300 watts per square meter. Solar panels, he noted, produce about 250.
That comparison is the heart of his case. Chips need power and then need to get rid of it as heat. In orbit, sunlight is constant and strong, and the radiators have an enormous cold sink to dump heat into. No cooling towers, no water, no grid fights with the neighbors.
But the physics is less generous than the tweet suggests. A radiator’s output depends heavily on how hot it runs. At temperatures that keep chips happy, you get a few hundred watts per square meter, not thirteen hundred. Hitting Musk’s number means running radiators much hotter, or counting both faces, or both. Maguire has reportedly described passive ammonia-loop systems that carry heat to radiators, borrowing from the Starlink satellites and the International Space Station. Those are real precedents. They’re also systems that handle kilowatts, not the hundreds of megawatts a frontier AI cluster eats.
So scale is the catch. A serious training facility would need radiator area measured in acres. Every square meter has to be launched, unfolded, and kept intact in a vacuum where there’s no technician to swap a bad pump. Cooling is only half the problem, and arguably the easier half.
Which brings it back to the question Musk didn’t answer in his posts: who pays? The economic case hangs on Starship getting launch costs low enough that orbital capacity beats a warehouse in Texas. That hasn’t happened yet. Musk has floated a timeline in which space compute overtakes Earth-based capacity by late 2028. I’d treat that date with some skepticism. A little over two years is a short window for anything that has never been built.
Still, it’s worth asking why he’s saying it at all. Earth’s data centers are running into real walls: grid connections that take years, local opposition, water use, and power prices. If AI demand keeps climbing, those limits get tighter. Musk may simply be betting that the pressure becomes unbearable, and that he’ll own the only escape route.
That’s the part that makes SpaceX’s position unusual. Other AI companies would have to rent rockets, probably from him. He controls the launch vehicle, the satellite network, and an AI company, xAI. No competitor has that stack. If orbital compute ever pencils out, he’s the one with a head start, and it’s hard to separate the engineering argument from the sales pitch.
There’s an environmental trade-off too, and it gets less attention than it should. Space-based power is clean at the point of use, but rocket launches aren’t. Frequent Starship flights mean emissions in the upper atmosphere, and thousands of large radiator-covered satellites add to orbital debris and sky-brightness concerns astronomers already raise about Starlink. Moving the problem off Earth doesn’t make it disappear. It changes who has to deal with it.
None of this means the idea is silly. Cooling in vacuum is a solvable engineering problem, and solar power in orbit really is abundant. But “solvable” and “obviously” are very different words. Musk used the second one.
The honest summary is that space compute is plausible, expensive, and unproven. The thermal argument is real but oversold. The economics depend on a rocket still being perfected. And the people best placed to profit are also the loudest advocates. That doesn’t make them wrong. It does mean the rest of us can ask for the math before we move the data centers.
