AI Is Running Out of Electricity on Earth. Engineers Are Seriously Building Server Farms in Orbit.

Somewhere in the American Midwest, a building the size of six football fields is humming so loud you can hear it from the highway. Inside: servers, chillers, transformers, and enough cooling water to fill an Olympic swimming pool — every single day.

There are hundreds of buildings like this. And they need more power than the grid can provide.

Microsoft signed a deal to restart the Three Mile Island nuclear plant just to keep its AI models running. Google broke its own net-zero climate commitments. The world is building artificial intelligence faster than it can build power stations to feed it.

And a small group of engineers have a proposal that sounds like science fiction until you think about it carefully:

Move the data centers to orbit.

Why Space Makes More Sense Than You Think

Here's the thing most people don't realize about data centers: they don't actually use most of their power to compute. They use it to cool down. Up to 40% of a data center's electricity bill goes not to running the chips — but to fighting the heat those chips produce.

On Earth, cooling is a war. Chillers. Water towers. Airflow systems. Enormous infrastructure. In Arizona or Singapore, where the air itself is scorching, the problem gets worse by the year.

In orbit, cooling works on entirely different physics.

There's no atmosphere in space — which means no air, no convection, no water cooling. That sounds like a catastrophic problem. But it also means the cold of deep space is just there, waiting. Heat radiates outward directly into a void that sits at −270°C. The International Space Station uses this exact principle: large flat panels called radiators glow infrared into the darkness, shedding thermal energy into nothing. No electricity required. No water consumed. The universe absorbs it for free.

And power? In orbit, the sun never fully sets for long. A satellite in low Earth orbit receives uninterrupted solar energy at 1,361 watts per square meter — roughly 40% more than reaches Earth's surface after the atmosphere absorbs its share. Solar panels in space are dramatically more productive than anything you can build on the ground.

The Hardest Part Isn't Getting There

The concept sounds elegant. The execution is brutal.

Engineers working on orbital computing hardware recently said something that stopped a room cold: "The ISS radiators are expensive and heavy. We're focused on making them cheap and light."

That single sentence captures the entire challenge. The International Space Station — humanity's most complex structure ever assembled — requires hundreds of square meters of radiator panels just to manage heat from its crew of seven and a handful of science experiments. A modern AI data center runs thousands of chips, each burning 300 to 700 watts. The math is unforgiving.

Launch costs have dropped by a staggering factor in the past decade. SpaceX Falcon 9 now delivers a kilogram to orbit for roughly $3,000 — compared to around $54,000 per kilogram during the Space Shuttle era. That shift is what's making these conversations possible for the first time. But building hardware that survives radiation, vacuum, extreme temperature swings, and years of microgravity — then operating it at a distance where the speed of light introduces real latency — is an entirely different problem from running a server rack in a climate-controlled room in Iowa.

The Companies Actually Building This

Several startups are now past the whiteboard stage, engineering actual orbital computing hardware. The pitch to investors isn't "replace Earth's data centers tomorrow" — it's "prove the concept works for specific workloads where geography doesn't matter and where power and cooling economics could favor orbit within a decade."

The calculation that makes serious investors pay attention: if power in orbit is effectively free (solar constant, no fuel required), and cooling is effectively free (radiation into deep space), then the only real cost is getting hardware up there — and that cost falls every year as launch competition intensifies.

The use case they're targeting first: AI model training. Training doesn't require low latency. You upload data, wait hours or days, and download results. Whether the training runs in a server room in Virginia or a satellite over the Pacific is completely irrelevant — as long as the answer arrives. That's the wedge. And it's a real one. You can learn more about orbital mechanics and why altitude determines so much about what's possible in space.

What Has to Be Solved Before Any of This Works

Engineers working on this problem don't sugarcoat the obstacles. Chips in orbit degrade faster from cosmic radiation than chips on the ground. You cannot walk in and replace a failed drive. Heat management requires radiators that are simultaneously large enough to work and light enough to afford launching. Power and data connections between modules must survive years of thermal cycling — swinging by hundreds of degrees every 92 minutes as a satellite alternates between direct sunlight and Earth's shadow.

None of these are theoretically unsolvable. Military satellites, scientific instruments, and the ISS have operated in this environment for decades. The question is whether it can be done cheaply enough — reliably enough — to make commercial sense at the scale AI demands.

The satellite investment numbers tell you which way the wind is blowing: $8.1 billion flowed into satellite companies in just the first half of 2026 — already surpassing every previous annual total on record. That's not all going to orbital computing, but it signals how seriously money is taking the idea that orbit is infrastructure, not just science.

You can see the current orbital picture — all 16,072 tracked objects — on the SkyLens live tracker right now. Most are communications satellites and debris. But that mix is changing faster than most people realize.

The Moment Two Curves Crossed

The strangest part of this story is the timing. The same decade that made large-scale AI possible also made reaching orbit affordable. Those two curves — AI's exploding power hunger and the falling cost of space access — are converging right now, in 2026, for the first time in history.

The people watching both trends from the same vantage point are starting to connect the dots.

Above you right now, 16,000 satellites are circling in near-perfect vacuum, bathed in unfiltered sunlight, radiating heat into the coldest environment in the known universe. For 60 years, we put telescopes and weather sensors and GPS transmitters up there.

Engineers are now asking: why not the servers? Explore more stories about what's happening in orbit right now.