Data centers in space are going to happen

In today’s Diary:

• Sun never sets on a satellite

• The 17,000-mph brain

• Soaking up a century of radiation

• Space’s first killer app

• Goodbye team Bezos

Hey Rational Optimist,

SpaceX went public on Friday in what was the largest IPO ever.

You know I’m a SpaceX superfan. It’s a once-in-a-century company. Our great great grandkids will study how SpaceX made getting to space cheap and reliable, opening up new potential for humanity while minting millionaires out of some 4,000+ employees.

In 2025 SpaceX flew more rockets than the entire rest of the world combined. Its launch business deserves every penny of the $75 billion it raised at its IPO.

But its $1.75 trillion valuation is really a bet on data centers leaving the planet.

In its IPO filing SpaceX says 85%+ of the company’s future market will come from AI data centers in orbit. SpaceX is an AI company!

“AI data centers… in space. Stephen, you’ve really lost it now.”

I get it sounds ridiculous at first. But orbital data centers are going to happen. As rational optimists, we should cheer it on for two huge reasons.

There’s not enough power on Earth.

A data center on Earth is just a building full of computer chips that turns electricity into intelligence. Trouble is the biggest data centers now gulp as much power as a mid-sized US city. And we’re trying to build hundreds of them at once.

If you want to plug in a big new data center today the utility company tells you to get in line until 2028. Finished data centers are sitting dark because they can’t get a wire to the grid.

Going to space solves the power problem.

By having solar panels ride the line around Earth that divides day and night… the sun never sets on them. A panel in orbit harvests 5X to 8X the energy of the same panel in Arizona. No expensive batteries needed.

Having no electricity bill is the holy grail for a system that turns electricity into thought.

By the way, when you picture data centers in space, are you picturing giant warehouses floating up there? That’s not what they’ll look like.

In space, orbital data centers will be sleek and minimal. Picture a Starlink satellite with a GPU, solar panel, and radiator attached. No building. No power lines. Just the chip, the thing that feeds it sunlight, and the thing that keeps it cool.

Then there’s the other P: politics.

Local governments rejected two data centers in 2023. In 2025 they rejected 50. They’ve already rejected more in the first three months of 2026 than all of last year. Maine just became the first state to ban large data centers.

We’ve seen this movie before and we know how it ends. A small group of activists can strangle a world-changing technology in its crib.

They did it to nuclear power and supersonic flight.

But luckily for AI…

There are no NIMBYs in space.

Last November a SpaceX rocket blasted off carrying a satellite the size of a fridge.

Bolted inside was a single Nvidia H100, the same kind of chip that trained ChatGPT. Built by 3-year-old startup Starcloud, the satellite ran Google’s Gemini AI while screaming around the planet at 17,000 miles an hour.

This was the first time a state-of-the-art AI chip operated in space.

Soon after, Google announced its own AI chips were bound for orbit by 2027. Nvidia built a version of its newest chip specifically for space too. See where this is going?

Sometimes readers ask me to go heavier on the “Rational” part of rational optimist. So let’s steelman the skeptics and address the four main critiques of orbital data centers.

Problem 1: Cooling

Today’s silicon chips are delicate. Push one above the temperature of a fresh cup of coffee and it cooks itself. Space is cold, so keeping the chips chilled must be easy, right?

Actually, no. Space is a vacuum, which makes it harder to dump heat.

On the ground we cool hot chips by blowing air over them or pumping water past them. In space the only way to dump heat is a radiator. The International Space Station hangs eight billboard-sized radiators to get rid of the same amount of heat as two AI racks.

This is a real problem, but genius innovators are working to solve it as we speak.

Companies like SpaceX, NVIDIA and Google are building chips out of tougher materials, designed to run cherry-red hot without melting.

Starcloud, the startup behind that fridge-sized satellite, is testing a radiator roughly 100X cheaper than the Space Station’s.

Problem 2: Radiation

Everything in space gets blasted with radiation. Over time this can chew a chip apart. It’s why spacecrafts use ruggedized processors.

Google took its newest AI chip to a nuclear lab and cranked the radiation way up. It kept blasting its chip until it swallowed the equivalent of roughly a century in space… and the processor still worked perfectly!

Problem 3: Maintenance.

When Meta trained one of its AI models on a cluster of 16,000 GPUs, the whole system crashed every three hours on average. That’s OK on Earth where a technician can swap the dead chip in minutes. In space you can’t just call a guy to go replace the melted chip.

The fix is out of Starlink’s playbook. Starlink satellites are built to be replaced. SpaceX deorbits four to five of them every day and launches fresh ones in their place. It’ll be the same for orbital data centers. They’ll be constantly refreshed with newer, faster chips.

As launch costs keep falling this replacement cycle only gets cheaper.

Problem 4: Cost

Energy is only 10% to 15% of what a data center costs over its life. The chips make up roughly 70%. So space cuts the smallest cost while raising the biggest cost.

I’ve read every serious analysis of orbital data centers. They all estimate building compute in space currently costs 2X to 7X what it costs on the ground.

But “expensive” is a lot better than “impossible” when we need to move fast. Building on many parts of Earth is impossible. No amount of money can speed up the building of a transmission line that takes ten years to permit.

The choice isn’t “cheap compute on Earth” versus “expensive compute in space.”

In many cases it’s “expensive compute now in space” versus… waiting.

Besides, “expensive” isn’t a reason to give up. It’s an opportunity.

Space launch was expensive. SpaceX cut the cost by 99%. It’s now a $1.75 trillion company.

Does SpaceX have competitors in space data centers?

Nominally yes. But not really.

SpaceX owns the entire stack:

• The rocket: Starship.

• The satellites: Starlink.

• The power: SpaceX is spinning up its own solar panels.

• The chips: Musk is building a giant chip factory of his own called “Terafab.”

• The AI: SpaceX absorbed xAI and now owns Grok.

That’s hard to compete with.

Jeff Bezos is the closest competitor.

Bezos founded Amazon which builds its own satellites and is making its own AI chips. He also owns Blue Origin, which is flying very large, partially reusable rockets.

But given Blue Origin’s rocket still can’t fly reliably, team Bezos is a distant second.

Elon is betting SpaceX on AI.

It’s worth remembering how Elon kept SpaceX going for 2+ decades to reach this point.

SpaceX was its own customer, in the form of Starlink. It could afford to keep flying Falcon 9 rockets over and over because it was hauling Starlink satellites up there. Starlink is now SpaceX’s only profitable division.

Its next big project—Starship, the biggest rocket ever flown—is built to reach Mars. But that’s not happening for at least a decade.

Starship needs a job now. Orbital data centers are the cargo that keeps Starship flying often enough to get cheap. And to defray some of its gargantuan capex costs.

This is a flywheel only SpaceX can spin.

A decade from now I think SpaceX could have a near monopoly on the orbital data centers, same as it owns launch today.

This will take time—it’s a 2030s story. Orbital data centers will follow the same path as launch, but speedrun it.

For 40 years the tonnage humanity launched was basically flat, then whoosh!

AI is the cargo that makes the skyscraper-sized, Mars-bound rocket pencil out.

Intelligence will pay for the rockets that carry us to the stars.

I bet within a decade SpaceX’s orbital-compute business will be bigger than Starlink…

and SpaceX will be the most valuable company on Earth.

It’s 2035.

You’re in the backyard at dusk with your family. Your son points at a thin ribbon of light streaking across the sky and asks, “Dad, what’s that?”

You tell him it’s a data center. Flying through space, running on nothing but sunlight.

You tell him that’s where the patient AI tutor that taught him to read lives. It’s helping cure the disease that took his great-grandmother.

Then you tell him about the moment in the 2020s when the future almost stalled. But when we hit walls on Earth, we carried the whole thing up to the sky.

He’ll shrug. To him, a sky full of machines thinking for humanity is the way the world has always been.

Before you go… thank you!!

To those of you who’ve become Patrons. You pushed Rational Optimist Society to the #1 rising bestseller across all of Substack (!), which is no small feat.

We might be the first “bestseller” ever that doesn’t sell anything.

If you’re new here, we do not paywall any of our articles or research. Our mission is to attract as many members as possible to make the world more rationally optimistic for our kids.

So, we simply let our members contribute via Substack’s subscription functionality if they’d like to support our research and growth.

Clearly, there are a ton of rational optimists out there. I doubt we’ve reached even 1% of them yet.

Onwards and upwards!

To help us grow and help rational optimism flourish, become a Patron by clicking here. 

—Stephen McBride

P.S.: The ROS Podcast is back.

This week: Ouros founder Ethan Loosbrock on why America is handing the battery war to China on a silver platter… and how Ouros is fighting back.