Why are tech companies trying to put data centers in space?

Why are the world's biggest tech companies launching data centers into space and whether the idea is genius or just very expensive science fiction.

In today's episode of the Epicenter, we look at why the world's biggest tech companies are launching data centers into orbit and whether the idea is genius or just very expensive science fiction. But before we dive in, if you’re someone who loves keeping tabs on the world of startups and technology, hit subscribe if you haven’t already. And if you’re already a subscriber, thank you! Maybe forward this to someone who’d enjoy this story but hasn’t discovered it yet. Now onto today’s story….

---

In late 2025, a US startup called Starcloud sent a satellite into orbit carrying an Nvidia H100 chip — the most powerful processor ever deployed in space. The launch cost millions. The satellite weighs a few hundred kilograms. And the question it raises is obvious: why on earth would you do that?

Every time you ask ChatGPT a question, a data center somewhere lights up. ChatGPT has over a billion weekly users. Each request burns compute, which burns electricity, which generates heat, which needs water to cool. The machines that power AI need land, power grids, and water at a scale that is starting to outrun what Earth can quietly provide.

Mark Zuckerberg put it plainly: “Energy constraints have become the largest bottleneck to building out AI data centers.”

So what happens when the most powerful technology in human history runs out of room to grow? Some people think the answer is 400 kilometres above your head.

The obvious answer is to just build more. More data centers, more power plants, more cooling infrastructure. Except it’s not that simple.

No company has yet built a single AI data center that runs on one gigawatt of power — the equivalent of a nuclear plant. And the trouble is that AI needs hundreds of them. In the US, permitting new power infrastructure can take a decade. The grid wasn’t designed for this load. And in water-stressed regions, residents and local governments are already pushing back against data centers that drink hundreds of millions of litres a year just to stay cool.

Which brings us back to the Starcloud satellite.

Space has two things Earth is rapidly running out of: room and energy. There is no zoning board in orbit. The land constraint, which is one of the most contentious problems in the data center industry today, simply disappears.

And then there’s energy. Satellites in orbit get continuous, uninterrupted solar power. Enormous solar arrays unfurled in orbit can capture it at a scale that would be impossible to replicate on the Earth.

The vision, in concept, is straightforward. Launch satellites carrying AI chips. Power them with solar arrays that stretch kilometres wide. When you send a prompt from your phone, it gets beamed up via laser, processed in space, and the answer comes back down — the same way your data travels through undersea fiber optic cables today, except the cable is a laser and it runs through the vacuum of space at the speed of light.

Sundar Pichai has already sketched out the plan publicly. Google intends to send tiny racks of machines into orbit by 2027, starting small and scaling from there. Their custom TPU chips are headed to space. A Singapore-based startup called Transcelestial is building the laser communication backbone for exactly this kind of network. And Starcloud, the company that started this story, is already valued at over a billion dollars on the back of one satellite launch and a bold idea.

But here’s where it gets complicated.

Space looks cold but there’s no atmosphere in orbit, which means heat can’t escape through airflow or liquid exchange the way it does in a terrestrial data center. The only way to shed heat in a vacuum is radiation, and radiation is slow. To dissipate the heat generated by a serious AI workload, you’d need radiator arrays so large they’d make the International Space Station look modest. Engineers sometimes describe it this way: the radiators required for a single space data center would be equivalent to the radiators of a hundred ISS units.

Then there’s the problem of radiation. The chips inside modern AI data centers are not designed to operate in an environment where high-energy cosmic particles are constantly flying through them. Errors accumulate and hardware degrades. And if something fails, there’s no engineer you can send up to fix it. You either launch a replacement at thousands of dollars per kilogram or you write it off.

Orbital debris is a growing concern too. Low Earth orbit is already crowded with thousands of satellites. Adding more infrastructure to a congested environment introduces long-term operational risk that nobody has fully priced in yet.

And underneath all of it is the hardest problem: launch costs. Right now, getting mass into orbit costs somewhere between $1,000 and $10,000 per kilogram depending on the rocket. A serious data center would weigh thousands of tonnes. The economics, at current prices, simply don’t work.

So why are serious people spending serious money on this?

Because the constraints on the other side are getting worse faster than the constraints in space are.

SpaceX’s reusable rockets have already cut launch costs by 50 to 100 times compared to a decade ago. Starship — SpaceX’s next-generation rocket still being tested — promises to push them down further. The cost that makes orbital computing unthinkable today could look very different in ten years.

So companies are placing early bets. SpaceX is planning on building data centers in space. Google is sending TPUs to orbit. Amazon and Blue Origin have similar ambitions. More than a dozen startups are chasing the idea. Whoever figures out the economics first will be sitting on infrastructure that the entire AI industry depends on.

The history of computing is a history of moving computation to wherever the constraints are lowest. From mainframes in government labs, to servers in corporate basements, to cloud data centers in Virginia and Oregon. Each transition seemed impractical until it didn’t. The jump to orbit is an order of magnitude harder than any of those. But then again, so was the jump to cloud.

What happens when the most powerful computers on Earth simply can’t get enough power on Earth?

Does the answer lie in space?

---

If you liked this story on SpaceX, feel free to share this with your friends, family or even strangers.

---