Elon Wants to Put
AI in Space.
Space Says No.

The Pitch

Elon Musk wants to solve AI's power problem by moving it off the planet

AI is eating the power grid. Building a single large data center now requires as much electricity as a small city. Water usage is enormous. Communities are pushing back. So Elon Musk looked at that problem and said: what if we just put it in space?

In January 2026, SpaceX filed with the FCCThe Federal Communications Commission — the U.S. government body that regulates radio frequencies, satellites, and communications technology. Any company that wants to launch a satellite constellation has to get FCC approval first. SpaceX filed on January 28, 2026 for permission to launch up to one million satellites to operate as orbital data centers. The FCC accepted the filing on February 2nd and opened a public comment period. for permission to launch up to one million computing satellites into orbit. Then in February, SpaceX merged with Musk's AI company xAI to make it official. Musk predicted orbital data centers would be more cost-effective than Earth-based ones "within two to three years."

The logic sounds reasonable on the surface. Space is always sunny — no clouds, no night if you pick the right orbit. Solar panels in orbitSolar panels on Earth only work when the sun is shining and the sky is clear. In orbit, depending on the satellite's path, panels can collect sunlight for the vast majority of their orbit with no weather interference. That's a real advantage — Google estimates this could mean roughly 8x more solar energy collection per panel compared to Earth. The problem is getting that energy to useful computing hardware and then getting the waste heat out. collect far more energy per panel than anything on the ground. No land permits. No water rights. No angry neighbors at a zoning board meeting.

There's just one problem. A problem that engineers say nobody in the press is talking about loudly enough. And it has nothing to do with rockets, satellites, or money.

It's heat.

The Problem

Space is -270°C. So why can't it cool a data center?

This is the part that breaks people's brains. Space is almost unimaginably cold. So how could cooling possibly be the problem?

Here's the thing: cold temperature and good cooling are not the same thing. On Earth, a data center pumps heat into air or water — both of which are incredibly good at picking up heat and carrying it away. In space, there is no air. There is no water. There is nothing to carry heat away from your equipment at all.

This means every orbital data center needs radiator panelsLarge flat surfaces — think giant metal wings — that emit waste heat as infrared radiation into space. The bigger the radiator, the more heat you can shed. The smaller it is, the hotter your computers get until they overheat and fail. These aren't optional extras. For an orbital data center, the radiator system IS the cooling system, and it has to be sized to match every watt of power the computers consume. — giant heat-dumping wings that radiate thermal energy into the void. And here's where the math gets brutal.

A two-sided radiator held at around 20°C — the temperature range electronics need to stay reliable — emits only about 633 watts per square meter. That means a single 1 megawatt facilityOne megawatt is 1,000,000 watts — roughly the power consumption of 1,000 average American homes. A modern large-scale AI data center can consume anywhere from 100 megawatts to over a gigawatt (1,000 megawatts). So the 1 megawatt example here is tiny by industry standards — a proof of concept, not a real hyperscale facility. needs roughly 1,500 to 1,600 square meters of radiator area before you account for real-world inefficiencies. That's about the size of four basketball courts. For one megawatt. A real AI data center runs hundreds of megawatts or more.

The Scale Problem

Musk's numbers require radiators the size of cities

Musk's publicly stated targets range from 100–200 gigawattsA gigawatt is 1,000 megawatts, or 1,000,000,000 watts. For reference, a large nuclear power plant produces about 1 gigawatt. The entire United States uses roughly 450 gigawatts of electricity at peak demand. Musk's orbital AI targets of 100-200 gigawatts annually represent an extraordinary fraction of total global power generation capacity — all of it needing to be cooled in a vacuum. of annual AI compute capacity, with longer-term statements reaching toward one terawatt. Let's run the radiator math on that.

At 100 gigawatts, even using optimistic radiator assumptions, you're looking at radiator surface areas measured in tens to hundreds of square kilometers. That's not a satellite. That's a structure the size of a major city — floating in orbit, constantly managing thermal loads, exposed to micrometeoroid strikesTiny particles of rock and debris traveling at orbital velocities — up to 17,500 mph. Even a speck of dust at that speed hits with the energy of a bullet. Large radiator arrays present enormous surface areas for these impacts, each strike potentially puncturing a cooling line or damaging a panel. The more radiator area you deploy, the more exposure you have, and unlike a data center on Earth, you can't just send a repair technician., and requiring a launch cadence nobody has ever come close to achieving.

And the radiators don't just need to be big. They add mass. More mass means more launch cost. More launch cost means you need even more satellites to justify the investment. The engineering analyses suggest that at scale, the radiator system would outweigh the computing hardware it's cooling by significant factors — meaning you're spending most of your rocket capacity launching cooling infrastructure, not actual computers.

What's Actually Real

Small orbital compute is happening. Hyperscale is a different conversation.

Here's where it's important to separate the hype from what's actually being built — because there's genuine activity at the smaller end of this spectrum.

Companies are already launching edge compute satellitesSmall satellites with onboard computing designed to process data right where it's collected, rather than beaming everything back to Earth. Useful for Earth observation, defense sensing, and AI-powered image analysis where the data is already in orbit. These are typically 10-100 kilowatts — tiny compared to a hyperscale data center, but solvable as an engineering problem. for specific use cases: processing satellite imagery onboard rather than downlinking raw data, AI inference for defense applications, low-latency handling of data that originates in space. These are 10–100 kilowatt systems, not gigawatt behemoths. The thermal management is hard but solvable.

China launched the first dozen satellites of its Three-Body Computing ConstellationChina's planned 2,800-satellite space supercomputer, announced in 2025. Each satellite is linked by laser communications. The full constellation is designed to achieve 1,000 peta-operations per second of computing power. That sounds enormous — but a single large hyperscale AI data center on Earth with tens of thousands of GPUs can far exceed that. It's a real technical achievement, but still orders of magnitude below what runs Earth's AI infrastructure today. in May 2025 — a 2,800-satellite space supercomputer linked by laser communications. Impressive engineering. Still orders of magnitude below a single hyperscale AI data center on Earth.

The honest engineering summary: putting a few smart servers in orbit is hard but doable this decade. Putting a hyperscale AI cloud in orbit is an unsolved thermodynamics and logistics problem that no current proposal has credibly resolved.

The Real Point

Musk isn't entirely wrong — he's just selling the wrong solution

Space does offer real advantages — unlimited solar power above the clouds, no water requirements, no land constraints. Those aren't imaginary benefits. The gap is between what's theoretically attractive and what's physically buildable at the scales being discussed.

The thermal physics don't care about the press release. A radiator in orbit emits 633 watts per square meter. That number doesn't change because Elon Musk is optimistic. At 100 gigawatts, the radiator field required would be visible from the ground.

Quick Rundown

The whole story in five lines

Sources

Where this comes from