AI Data Centers in Space: Google, SpaceX, and Starcloud Are Already Racing to Orbit
AI Data Centers in Space: Google, SpaceX, and Starcloud Are Already Racing to Orbit
On October 1, 2026, a SpaceX Falcon 9 lifted off from Vandenberg Air Force Base in California carrying something that had never flown before: Google's tensor processing units, its custom AI chips, riding aboard a Planet Labs satellite prototype as part of a program called Project Suncatcher. By 1 p.m. local time, all payloads were deployed. The chips were in orbit. The race to put AI computing in space had its most prominent entrant yet.
This is not a distant vision. It is a situation report. In less than 12 months, orbital AI computing went from a fringe concept to a field with real hardware in orbit, real capital deployed, and real timelines attached. The driving force is not just engineering ambition. It is the political reality documented in the previous article in this series: Earth-based communities, by margins of 69 to 71 percent, do not want AI data centers near them. Space removes that problem entirely. No ratepayers. No zoning fights. No water rights disputes. The engineering challenges that remain are solvable in ways the political ones are not.
The Players, as of October 2026
The field is small but moving fast. Here is where each major participant stands.
Starcloud is the furthest along and the best funded. In November 2025 the company launched Starcloud-1, carrying an Nvidia H100 GPU — the first of its kind in orbit. Since then the satellite has trained an AI model in space and run a version of Google's Gemini. The company has raised $450 million in total, carries a valuation of $2.3 billion, and counts Nvidia and Cisco among its investors. Nvidia is also collaborating with Starcloud on the Space-1 Vera Rubin Module, a chip architecture designed specifically for orbital conditions. Starcloud's next satellite will carry Nvidia Blackwell chips, an AWS server blade, and the largest deployable radiator ever flown on a private satellite. The company's long-term plan is a constellation of 88,000 satellites delivering 20 gigawatts of computing capacity. To put that in context, a single hyperscale data center on Earth draws roughly 100 to 200 megawatts. Starcloud's planned network would represent the equivalent of 100 to 200 of the largest such facilities ever built.
Axiom Space reached a different kind of milestone in January 2026. It launched two orbital data center nodes to low Earth orbit, operating through Kepler Communications' optical relay network at 2.5 gigabits per second. These are not demonstration units. They are the first operational foundation for space-based cloud computing, built to Space Development Agency interoperability standards. Their primary customers today are defense and intelligence users who need secure, real-time data processing without routing through ground stations.
Google's Project Suncatcher, announced in November 2025 and now flying as of October 1, 2026, is Alphabet's formal research initiative into space-based AI computing. Google describes the effort as a step toward learning whether space could one day host scalable machine learning infrastructure. Satellites in low Earth orbit receive near-constant sunlight, which Google says could generate up to eight times more solar power than equivalent panels on Earth. The October launch is a test. How the TPUs hold up under orbital radiation conditions is still being evaluated.
SpaceX is approaching the field from two directions. The company has asked the U.S. government for permission to build and operate up to one million satellites for distributed computing. COO Gwynne Shotwell has stated that SpaceX plans to deploy supercomputers in space in 2027. Starcloud CEO Philip Johnston expects SpaceX will primarily serve its own internal workloads rather than selling capacity to others — which, if accurate, leaves the commercial orbital computing market open.
Orbital, backed by a16z Speedrun, is pre-launch and targeting April 2027 for its first test mission on a Falcon 9. The company is opening a research and development facility in Los Angeles and is currently filing with the FCC for a satellite constellation. Its focus is AI inference, not training — inference workloads can be distributed across many independent satellite nodes without requiring tight synchronization between them.
In less than 12 months, orbital AI computing went from a fringe concept to a field with real hardware in orbit, real capital, and real timelines. The political pressure on Earth is part of what is driving the clock.
Why Space Actually Works for This
The case for orbital data centers is not just about avoiding community opposition, though that is a real factor. The physics of space create genuine infrastructure advantages that do not exist on the ground.
Power is the first one. Data centers on Earth consume electricity from the grid and generate enormous heat that requires water-intensive cooling systems. In orbit, solar power is available around the clock with no atmospheric interference. Heat, the chronic enemy of dense computing, radiates directly into the vacuum of space. No cooling towers. No water consumption. No utility hookups. The two most politically toxic byproducts of a terrestrial data center simply do not apply.
The second advantage is latency for specific use cases. Satellites in low Earth orbit are closer to other satellites than any ground station is. Processing data where it is generated — aboard imaging satellites, weather platforms, or defense systems — eliminates the round-trip to Earth entirely. Axiom Space built its first nodes specifically for this use case.
The Honest Constraints
The engineering challenges are real and should not be minimized.
Launch cost is the central constraint. Starcloud's CEO has said plainly that orbital data centers will not be competitive on energy costs until SpaceX's Starship rocket is flying frequently at commercial rates. His target is roughly $500 per kilogram to orbit. Starship has not yet reached that price point at scale. Until it does, orbital computing is expensive relative to terrestrial alternatives.
Radiation is the second challenge. The space environment degrades electronics in ways that ground-based hardware never encounters. Radiation hardening adds cost and complexity. Starcloud-1 lost an Nvidia A6000 GPU during its launch. Learning how long orbital hardware actually survives under real conditions is part of what these early missions are designed to test.
Synchronization is the third challenge, and the least solved. Training large AI models requires many GPUs working in tight coordination, exchanging enormous volumes of data in near real time. Doing that across satellites separated by hundreds of kilometers requires optical inter-satellite links — OISLs — of a kind that are only now being tested at scale. Axiom Space's nodes use Kepler's optical relay network at 2.5 gigabits per second, which is an early proof point, but the bandwidth requirements for synchronized AI training dwarf what any current OISL can deliver. Orbital inference — running a trained model — is far more tractable than orbital training, which is why most players are starting there.
The Timeline That Matters
The unlock that makes orbital data centers commercially viable at scale is Starship. SpaceX has targeted 2027 for deploying its own space-based supercomputers. Orbital is targeting its first test mission in April 2027. If Starship reaches frequent commercial flight in the 2028-to-2029 window — which Starcloud's CEO expects — launch costs drop, payload capacity explodes, and the economics of orbital computing change fundamentally.
The projected market reflects that trajectory. Orbital computing is estimated at $1.77 billion by 2029 and $39 billion by 2035. Those numbers assume the Starship cost curve materializes. They also assume that the political pressure on terrestrial data centers continues to intensify — which, given the polling data and the legislative activity documented in this series, appears likely.
The race is early. The hardware flying today is proof-of-concept. But proof-of-concept was where cloud computing was in 2006, and the companies that dismissed it then spent the following decade catching up to the ones that did not.
Aaron Rose is the founder and publisher of Tech Reader Magazine, covering system architecture, cloud platforms, and AI policy.