The Wall Street Journal reported today that Google is in talks with SpaceX to build data centers in orbit. Neither company is confirming anything, and the talks are early. But Google has been researching space-based data centers since at least November of last year, when it confirmed a project called Suncatcher and announced plans to launch two prototype satellites in partnership with Planet Labs by early 2027. So the talks are not random.

The pitch for orbital data centers is, on reflection, pretty sensible. Data centers on Earth have two big problems: power and cooling. Power is increasingly hard to get. Grid interconnection queues run years long in most markets, and at least six U.S. states are considering moratoriums on new data center construction. Cooling is expensive and water-intensive, which makes the problem worse in hot and dry places.

In low Earth orbit, both problems largely disappear. Solar irradiance at 550km runs around 1,360 watts per square meter, continuously, with no clouds and no night cycle. Cooling in vacuum requires no water and no grid power: just radiators, which take up significant real estate on a satellite but don’t draw from the electricity supply. The Sun and the vacuum of space, it turns out, are better infrastructure than a suburban utility district in Virginia.

The catch is getting there. Starcloud, the orbital compute startup that raised $170 million in March, established that you can run AI inference in orbit: its first satellite trained a model on an H100 while traveling at 17,000 miles per hour. The economics, though, only work at around $500 per kilogram to LEO. Falcon 9 currently runs about $3,600. The bridge to cost-competitiveness runs through Starship, SpaceX’s next-generation rocket, which is not yet commercially operational.

Everybody Wants SpaceX

Starship is why all roads lead to SpaceX. Google can design the satellites. Planet Labs can build them. But if you want to launch enough mass into orbit for meaningful AI compute at competitive cost, you need a rocket that only one company is building.

Anthropic got there first, signing a deal with SpaceX and xAI earlier this month that covers the Colossus 1 data center in Memphis, plus a more speculative interest in developing orbital compute capacity together. Google is now, apparently, in similar conversations.

This puts SpaceX in a notable position. It is the launch provider that anyone serious about orbital compute has to work with. It is also building its own orbital compute constellation: SpaceX has filed with the FCC to launch up to one million satellites for distributed AI inference, prioritizing xAI’s Grok workloads. The company whose rocket you need is also the company most aggressively building for its own purposes. Starcloud flagged this as a structural risk when it raised its Series A, and that was before Google showed up to join the queue.

SpaceX is also preparing for an IPO at a reported $1.75 trillion valuation. Having Google and Anthropic as orbital compute customers, or even as orbital compute conversation partners, is worth something in a roadshow. Google’s Project Suncatcher predates the IPO by months and involves real satellite development, so the interest is genuine. The incentive to make these conversations public, though, is not nothing.

The Jurisdiction Problem

There is a question nobody has answered about orbital AI compute: what law governs it?

Terrestrial data centers are heavily regulated. Data residency rules, GDPR, export control frameworks, law enforcement access obligations, physical security requirements: all predicated on the data center being in a country, which has a government, which has laws. A satellite at 550km is subject to the Outer Space Treaty, FCC licensing, and ITU coordination procedures. None of those were designed with AI workloads in mind.

The frameworks building around frontier AI infrastructure: who reviews models, what governments can access what compute, how export controls apply, were written for terrestrial data centers. Moving the compute off the Earth does not make those policy questions go away. It just makes them considerably harder to answer.

–
By the Control Plane Editorial Team