Google and SpaceX are not just chasing a space story. They are testing whether AI's next infrastructure bottleneck can be moved off the power grid and into orbit.
The race to feed artificial intelligence has already taken over power markets, chip supply chains and local planning fights. Now it is moving into space. According to The Wall Street Journal, Google and SpaceX are in talks over a possible launch partnership for orbital data centers, a deal that would connect Google's AI hardware ambitions with the rocket company that already dominates low Earth orbit.
This is not a near-term replacement for the data centers now spreading across Virginia, Texas, Arizona and the Gulf Coast. It is still early research, with difficult physics standing between the pitch deck and commercial reality. But the talks matter because they show how far the AI infrastructure race has stretched. When companies begin asking whether compute should leave the planet, it usually means the constraints on Earth are starting to bite.
Google has already put a name to its part of the idea: Project Suncatcher. The company announced the research moonshot in November 2025, describing a future network of solar-powered satellites equipped with its Tensor Processing Units, the AI chips that sit behind much of Google's machine learning work. Planet is set to build and operate two prototype satellites for the project, with a target launch by early 2027.
The appeal is obvious. A satellite in the right orbit can see far more continuous sunlight than a solar farm on Earth. It also avoids land fights, water consumption complaints and the strain that massive data centers place on local utilities. For AI companies that increasingly measure growth in megawatts as much as model performance, that is a tempting thought.
The hard question is whether sunlight in orbit can beat concrete, substations and fiber on Earth. Google has said Project Suncatcher is a moonshot, and that language is doing important work. The prototype mission is meant to test whether TPUs can operate reliably in space, not whether Google can run Gemini from a fleet of satellites next year.
Launch cost is the first test. Space-based compute only begins to make sense if getting hardware into orbit becomes cheap enough that the energy advantage can offset manufacturing, deployment and replacement costs. Google's own research has pointed to the possibility that launch prices could fall sharply by the mid-2030s, but that depends heavily on reusable rockets becoming cheaper and more frequent than they are today.
Then there is heat. Space sounds cold, but data centers do not just need cold surroundings. They need a way to move heat away from chips constantly and predictably. On Earth, operators can use air, liquid cooling and large mechanical systems. In orbit, there is no air to carry heat away, so satellites need radiators and careful thermal design. That makes every watt of compute a design problem.
Radiation is another quiet but serious barrier. Google's early tests suggest its Trillium-generation TPUs survived simulated low Earth orbit radiation without damage, but a lab test is not the same thing as years of commercial operation. Faults, bit flips, component degradation and maintenance limits all matter when the server rack is hundreds of miles above the ground.
SpaceX brings scale and a story
SpaceX has its own reason to make orbital compute sound credible. The company filed with the Federal Communications Commission in January 2026 for authority to launch and operate up to 1 million satellites as part of what it called the SpaceX Orbital Data Center system. The FCC accepted the application for filing on February 4, opening the proposal to review and comment.
That number should be treated carefully. Companies often ask regulators for more capacity than they ultimately use, and 1 million satellites would be far beyond anything in orbit today. Still, the filing shows where SpaceX wants investors and customers to look. Starlink proved that a private company could build a global satellite network at industrial scale. Orbital data centers would try to turn that launch machine into AI infrastructure.
Google is not just another prospective customer in this story. The company holds a 6.1% stake in SpaceX, according to the Journal, giving the talks an extra layer of strategic logic. Google needs more ways to scale AI compute. SpaceX needs large commercial use cases that justify its rocket capacity and support its broader growth narrative. Each side has something the other can use.
The risks are not small. Latency could limit which workloads make sense in orbit, especially where data has to move between Earth and satellites repeatedly. Space debris is already a major concern before anyone adds enormous compute constellations to crowded orbital shells. Regulators will also have to decide whether the environmental savings claimed on Earth are worth new congestion above it.
The practical takeaway is that AI infrastructure is becoming a full-stack industrial race. Chips still matter, but so do power contracts, cooling systems, launch capacity, optical links, orbital safety and regulatory approvals. If orbital data centers become real, they will not arrive as a single breakthrough. They will arrive through years of small tests that show which workloads can tolerate the cost, latency and risk.
For now, Google and SpaceX are selling possibility more than capacity. But possibility has value in this market. The companies that solve AI's power problem, whether on Earth or above it, will shape where the next generation of computing gets built.
Also read: Anthropic pushes Claude deeper into legal work. • Wispr Is Chasing A $2 Billion Valuation As Voice AI Moves Past Dictation • Isomorphic Labs raises $2.1 billion to test AI drug discovery
