The most revealing question about putting AI infrastructure in orbit is not whether it can be done. Engineers have already demonstrated that meaningful computing can operate in space, and companies are now designing systems intended to push much further. The more interesting question is why the industry would move the machine at all. That question matters because AI infrastructure increasingly looks less like software and more like an industrial system. It needs processors, electricity, cooling, networking, physical space and a supply chain capable of manufacturing, transporting and replacing components as computing systems evolve. Moving those requirements into orbit does not eliminate them. It changes their operating environment.

Google’s Project Suncatcher makes the proposition unusually concrete. The company is researching networks of solar-powered satellites equipped with its Tensor Processing Units and says it plans to launch two prototype satellites with Planet by early 2027 to test hardware in orbit. Starcloud, meanwhile, says its Starcloud-1 satellite launched in November 2025 carrying an Nvidia H100 GPU and subsequently ran AI workloads in space. The technological novelty is obvious. The industrial logic is less settled.

The orbital data center changes what “infrastructure” means

Terrestrial data centers hide an extraordinary amount of physical complexity behind a familiar building. Power arrives through an electrical grid, replacement hardware arrives by truck or aircraft, technicians can reach a failed server, and cooling systems can move heat into an environment that engineers can continuously manage. An orbital system gets none of those conveniences for free. Space offers an attractive energy source because sunlight remains available outside Earth’s atmosphere, and companies developing orbital computing architectures point to solar generation as a potential advantage. Starcloud, for example, argues that orbital facilities could combine solar power with radiative cooling while avoiding some terrestrial constraints.

But electricity is only one part of the computing equation. A processor does not simply consume energy. It converts much of that energy into heat, and the system must dispose of it. In a terrestrial facility, engineers can move heat through increasingly sophisticated cooling loops and ultimately reject it into the surrounding environment. In orbit, radiative cooling becomes central because there is no surrounding atmosphere through which conventional convection can carry heat away. That does not make cooling impossible. It makes thermal engineering a fundamental part of the computing architecture.

The same applies to radiation. NASA notes that ionizing radiation can produce single-event effects in spacecraft computers, ranging from data errors to system failures and permanent damage. NASA’s work on radiation-tolerant computing illustrates why conventional computing hardware cannot simply be assumed to behave in orbit as it does inside a terrestrial server rack. The irony is hard to miss: AI hardware development increasingly emphasizes performance and energy efficiency, while an orbital architecture introduces an environment where reliability, radiation protection, redundancy and fault recovery become important design variables.

The supply chain does not disappear when the server rack does

The strongest argument for orbital computing often starts with scarcity on Earth. Data centers require enormous quantities of electricity, land, cooling capacity and grid infrastructure. An orbital architecture appears to bypass several of those constraints. Yet it introduces another supply chain: launch. Every kilogram of computing hardware placed into orbit has to leave Earth first. That means processors, memory, storage, power systems, radiators, communications equipment and structural components become payloads in a transportation network whose economics depend on launch capacity, orbital insertion and mission reliability.

ESA has already treated space-based data centers as a serious feasibility question rather than pure science fiction. Its studies have examined architectures in which satellites process Earth-observation data in orbit, reducing the need to transmit large volumes of raw information back to Earth. ESA has also examined the infrastructure and system requirements that future space-based data centers could require. Processing satellite data closer to where it originates can have a straightforward systems rationale.

If an Earth-observation satellite generates large quantities of raw information, processing that information in orbit can reduce downlink requirements and return only useful results to the ground. ESA’s research specifically identifies limited data-downlink capacity as a driver for onboard and space-based processing. That use case is fundamentally different from putting general-purpose AI training infrastructure in orbit simply because terrestrial data centers have become difficult to expand. The first is a network optimization problem. The second is an infrastructure relocation strategy.

Upgrading an orbital GPU is a different economic problem

AI hardware also creates an uncomfortable question for orbital economics: What happens when the chip becomes obsolete? The terrestrial AI industry operates within a hardware market where processor architectures and semiconductor technologies continue to advance rapidly. A data center can decommission a server, install a replacement and redirect workloads with relatively little drama. An orbital cluster cannot assume the same upgrade cycle. Replacing a failed or outdated accelerator could require a launch, rendezvous, robotic servicing or an entirely new spacecraft. The economic calculation therefore shifts from the price of a chip to the lifecycle cost of the platform carrying it.

That could encourage a different philosophy of AI hardware. Orbital systems may favor modular designs, longer-lived processors, redundancy and remote reconfiguration to accommodate the greater difficulty of replacing hardware in space. NASA’s current high-performance spaceflight computing work points in a similar direction at the spacecraft level, emphasizing fault tolerance, system health monitoring and computing that can operate reliably in harsh environments. The result could be a new category of infrastructure where compute density is important, but maintainability becomes equally important.

Space may expose AI’s physical limits rather than remove them

This is where the orbital AI proposition becomes more interesting than the familiar promise of limitless solar power. Moving compute into space does not make computing immaterial. It makes the physical dependencies impossible to ignore. A terrestrial AI cluster can appear almost abstract to its users. They see an API, a model or a cloud service. Behind that interface sits a large industrial machine that consumes electricity, generates heat, requires networking and eventually needs replacement hardware. An orbital data center makes those dependencies visible. Its processors need protection from radiation. Its power system needs to survive the orbital environment. Its thermal architecture must reject heat through radiation. Its communications system must maintain reliable links. Its hardware must survive launch. Its computing capacity must justify the cost of getting every component into orbit.

There are already signs that the industry is testing that boundary rather than merely speculating about it. Google is pursuing prototype missions for Suncatcher, while Starcloud is developing a commercial satellite architecture that it says will provide GPU computing and storage in orbit. The eventual market may therefore look less like “data centers, but in space” and more like a specialized computing layer for workloads that benefit from being there. That is a considerably more compelling proposition. The real test for orbital AI will not be whether engineers can make a GPU work above Earth. They already can. The harder test will be whether the economics, reliability and upgrade cycles can support a computing business at meaningful scale. If that happens, space will not have solved AI infrastructure’s physical problem. It will have turned that problem into an even more visible one.