I’ll admit that I’m already shaking my head when I hear someone talking about the concept of data centers in space. You’re always going to hear the same three points trying to sell it. There is unlimited solar power, no land or water constraints, and cooling in space is free because it’s cold. Sounds good in theory but physics always wins. The costs associated with the drawbacks means that what you’re being sold is entirely different from what is being delivered.

Lots of companies are jumping in to get into space. Starcloud, Axiom, Sophia Space, Google Project Suncatcher, and even the crazy million satellite promise from SpaceX want a piece of the action. However, read through the marketing and you’ll see that these companies aren’t building data centers in space. They’re really building disposable compute modules that just happen to be in space because calling them “data centers” is where you get your funding.

The hints are there if you look closely. Coverage of India’s space aspirations talk about plans to deorbit failed modules and replace them, much like Starlink satellites or mobile phones. Other consultants have talked about servicing and unit refreshes as the real factor behind how this scales past a slick demo for investors.

Solar So Good

I will admit there is one area where this whole idea makes sense. Solar power in orbit is as abundant as can be. There’s no atmospheric attenuation. No need to worry about cloud days. And satellites don’t have to worry about night so they can get power constantly in the right orbit. Solar is also much less expensive than you might think, provided you don’t care about how long they last. Typical hardened triple junction cells that are used for things like the ISS cost about $50-$100 per watt. The cheaper commercial cell used by Starlink run about $.20-$.50 per watt. That’s a massive cost advantage. Those commercial cells are also more resistant to radiation damage than you might otherwise think. Real data from satellites show about 0.18% power loss per year due to degradation, which is less than 2% loss over a decade.

Power generation isn’t the big issue here. It’s probably the one thing that works in the favor of the people backing these ideas. But keep in mind that whole 2% power loss over a decade. We’re going to come back to it in a minute.

No Fuego

Heat dissipation is one of the biggest hurdles to clear. On Earth, data centers are cooled by liquid and/or air moving over the devices and convective cooling. You can cool something like 2,000 watts per square meter that way. It’s a very efficient way to cool things. In orbit there is no convection. There’s no air to move. You have to use radiative cooling and that’s very inefficient at the temperatures that most of these devices are going to be running at. If you don’t believe me then just look over at the vacuum tumbler that sitting on your desk. Do you know why it keeps your drinks cold? Because there’s a vacuum inside it. Vacuums don’t radiate heat well at all.

Let’s check the math though. If you want to keep your electronics running around 70 degrees F you’re going to need about 1,200 square meters of radiator surface to handle one megawatt (MW) of heat. That’s roughly four standard tennis courts. A typical data center averages between 5 and 10 MW of power consumption. While the math isn’t perfect for power usage versus heat generation it’s close enough to say that a typical data center is going to need at least one or two American football fields worth of radiative cooling to operate. And the power budgets for these things are only going up.

Companies like Sophia Space are trying to turn the entire chassis into a kind of heat exchanger with a solar cell on one side and a passive radiator on the other. They’re claiming that the typical HVAC overhead of 92% you see on Earth can be brought down to something like 8%. But when those nodes scale up your cooling needs become multiplicative with your compute scale too. No matter how you look at it you’re going to have some very big radiators in space. Half the things sticking off the ISS are radiators, not solar panels.

The Deorbit Cycle

Let’s set aside physics for a minute and talk about money. The current thinking around compute in orbit says that the satellite should be expected to last about 5-6 years before it degrades to the point where it is useless. Great timeline for writing off the cost of the unit. But the GPUs inside the satellite that are doing the hard work double in performance every two years or so. Yes, Moore’s Law is slowing down a little but every two years or so is still very much in the realm of possibility for huge performance gains.

That means a satellite operating for the full intended lifetime is going to be three generations of GPU behind in performance. And unlike a terrestrial data center where you can just go in and swap out some of the parts you’re stuck with what’s in the satellite for the life of the system, practically. So what you launch today is already behind the curve. If you want to beat the current model you have to be able to iterate faster than dirt-bound data centers, not slower.

Don’t even think about trying to repair these things in orbit. Remember trying to fix the Hubble Space Telescope? That was massively expensive. Even with modern efficiencies the current estimations for a repair are about $100 million. That means it is effectively cheaper to just deorbit a failed or outdated satellite instead of trying to repair or replace components. The math behind launch costs supports this disposable idea. Falcon 9 can deliver payloads to space today for about $1,225 per pound, which is like 90 percent cheaper than the old Space Shuttle. SpaceX says that the target for Starship is closer to $33 per pound, provided it can stop blowing up or turning into a boat when it lands. If the numbers can keep moving down and hit those lofty goals then it really does become cheaper to blow up the old stuff than try to fix it. And you thought we had a space junk problem before.

That’s not to say there aren’t challenges. Starship’s schedule has slipped three times already this year. The price of a Falcon 9 launch went up this year for the first time since 2022. Everyone is betting that the cost curve is going to keep going down and they’re making their bets on it, but what happens when the curve becomes a plateau?

Not everyone is building disposable data centers in space. Sophia Space is building their modules so they can be swapped out. Axiom is also building tiles that can be serviced. But when you compare those two approaches to SpaceX just wanting to launch a million satellites into orbit and hoping that most of them stay up there for a year or two you can see the industry hasn’t quite figured out their approach yet. But both approaches eventually lead to the same conclusion. If you want to service devices you have to do it before they become obsolete. And you have to make it cheap enough to do in an industry that thinks in quarters and not years. That’s why the deorbit model looks the most economical right now.

Tom’s Take

Space data centers aren’t impossible. They’re just really impractical right now. Cooling is a challenge that can be solved with money and money is the real root behind the serviceability issues. The way this is being pitched to investors right now doesn’t make the economics work for anything other than a market that looks more like an iPhone launch than a rocket launch. When the GPUs in the satellites are obsolete on the launch pad you know you’ve got some bigger issues to examine.

All of this assumes that the money in space data centers get you returns on a terrestrial data center timeline. You have to be willing to take some losses while the cost curves come down to a place where you make some money. How many investors do you know that are willing to take some losses for the next couple of quarters? And that’s only if the costs keep moving down at the current rate, which isn’t guaranteed.

The next time someone tells you they’re excited to build a data center in space, just tell them what they’re really building is a hardware subscription that might not make financial sense in the long run.