# Could AI data centers really move to space?

Date: 2026-08-17 · Author: Robert Scales · https://guidetoartificialintelligence.com/en/learn/data-centers-in-space/

**Answer:** The idea is serious enough that Google, SpaceX and Blue Origin are working on or filing for it, because solar panels in the right orbit produce far more than on Earth and need no land or local permits. But no orbital data center exists at useful scale, and the economics depend on launch prices that have not been achieved yet.

On Earth, AI data centers run into slow grid connections, water disputes and local votes. A growing number of companies have noticed that none of those exist in orbit. That idea went from science fiction to regulatory filings within about a year.

What makes this worth a beginner's attention is how much of it is checkable. Google published its calculations, so you can see which assumptions carry the argument. The honest summary is that the physics may be workable and the economics are not there yet.

## Who is doing what

| Company | Plan | Status |
|---|---|---|
| Google, with Planet | Project Suncatcher: satellites carrying its TPU chips, linked by laser | Research paper in November 2025; two prototype satellites slated by early 2027 |
| SpaceX | Application for up to one million satellites | Filed with the US Federal Communications Commission (FCC) in January 2026 |
| Blue Origin | Project Sunrise: 51,600 satellites | FCC filing in March 2026 |
| Starcloud | 60,000 spacecraft proposed | Flew an Nvidia H100 in orbit aboard a test satellite |

A filing with the FCC asks for permission to use radio spectrum and orbits. It does not show that money or hardware is ready. [Experts told TechCrunch](https://www.techcrunch.com/2026/03/20/jeff-bezos-blue-origin-enters-the-space-data-center-game/) that such projects are unlikely before the 2030s. CNBC notes that Google owns 6.1% of SpaceX and that SpaceX is a potential launch partner, which is worth remembering when these companies discuss the idea.

## Why anyone thinks this could work

**Sunlight.** Google says a well-placed orbit can make a solar panel up to eight times as productive as on Earth, with nearly continuous sun. It proposes a dawn-dusk orbit, which rides the line between day and night, so that satellites rarely enter shadow and need fewer heavy batteries.

**No local fights.** Blue Origin's filing presents the idea as a way to shift energy and water use away from communities on Earth.

## The four hard problems

**Links between satellites.** AI work needs chips that exchange data constantly. Google says matching a data center needs tens of terabits per second between satellites, which requires very close formation flying. Its example is a cluster of 81 satellites at 650 kilometers altitude, within a radius of one kilometer, with neighbors only 100 to 200 meters apart.

**Radiation.** Space radiation can corrupt or damage electronics. Google tested its Trillium TPU in a particle beam. The chips' memory, the most sensitive part, showed irregularities after 2 kilorads, nearly three times the 750 rads expected over five years in orbit behind shielding. That is encouraging, but it is a lab test of one chip, not years in space.

**Heat.** A vacuum has no air or water to carry heat away, so it must radiate from large surfaces, using heat pipes and radiators. Radiators add mass, and mass is what launches charge for. Google's paper lists thermal management among the issues that still need to be proved in orbit.

**Launch cost.** This one decides the rest.

## A worked example: the launch price

Google's calculation compares launch cost, spread over a satellite's lifetime, with what a data center on Earth pays for power.

| Item | Figure from Google's paper |
|---|---|
| Launch price needed | about $200 per kilogram, projected by the mid-2030s |
| Launched power cost at that price, Starlink v2-style satellites | about $810 per kilowatt per year |
| Same, across a wider range of satellite designs | about $810 to $7,500 per kilowatt per year |
| US data centers' power spend today | about $570 to $3,000 per kilowatt per year |

Take 1 gigawatt, or 1 million kilowatts, at $810: that comes to about $810 million a year for launch alone. The comparison sounds favorable until you read what it leaves out. It covers launch, not the cost of building satellites, replacing them when they wear out or communicating with the ground. Google itself describes this as a rough point of comparison, and $200 per kilogram is its projection, assuming launch costs keep falling as more mass is flown.

*The multiplication here is our own illustration from Google's figures.*

Others are more skeptical. In CNBC's June 2026 report, Harvard economist Mark Weinzierl said he had not seen any analysis showing these systems are cost-competitive now, and called the long-term case a bet. Elon Musk claims costs will favor space within two to three years, while CNBC notes his record of missed schedules. Jeff Bezos called the concept very realistic but said two to three years is probably a little ambitious.

## What it would and would not solve

Even if it works, space would not remove the problems entirely. Orbital computers would still need chips, which are in short supply, and rocket launches of their own. A million satellites would raise questions about crowded orbits that this article cannot settle.

For Canadian readers, the near-term question remains on the ground: provinces are [rationing power to data centers](/en/learn/ai-data-center-power-grid/) now, and nothing in these plans changes that soon.

> **Try it:** Change one assumption. Using the table, work out what happens to the launched power cost if launch stays at twice $200 per kilogram, assuming cost scales directly with price. Then compare that with the range for power on Earth.

## When would we know whether orbital data centers work?

The first real data should come from small demonstrations, not from the large filings. Google's two prototypes with Planet, slated by early 2027, are meant to test how its chips and models run in space and to validate the laser links. They will not be a working data center. Starship's progress will also be a signal, since many of these plans assume its launch prices.

Three things would show the idea becoming real: a satellite running a useful AI task continuously for months, a launch price near the numbers in these papers and a customer paying for orbital computing at a price competitive with the ground. Until then, treat the large satellite counts as requests for permission rather than plans under construction.

## Sources

- Exploring a space-based, scalable AI infrastructure system design: https://research.google/blog/exploring-a-space-based-scalable-ai-infrastructure-system-design/
- Towards a future space-based, highly scalable AI infrastructure system design: https://services.google.com/fh/files/misc/suncatcher_paper.pdf
- No one wants AI data centers on Earth. Do they make sense in space?: https://www.cnbc.com/2026/06/21/do-space-based-ai-data-centers-make-economic-sense.html

---

From the Guide to Artificial Intelligence, by Robert Scales, published by Holistic Hybrid Inc. Share and adapt it under CC BY-NC-SA 4.0 (https://creativecommons.org/licenses/by-nc-sa/4.0/): credit the guide, link to https://guidetoartificialintelligence.com/en/learn/data-centers-in-space/, no commercial use without permission, and keep the same licence.