One hundred times more powerful than any processor ever flown in space. Two thousand teraflops of AI computing power. Energy costs: one-tenth that of terrestrial data centers. Water consumption for cooling: zero. These are the numbers. Starcloud-1, the satellite that Nvidia and the startup Starcloud will launch in November 2025 with an H100 chip on board. The goal is to test whether the orbital servers They can truly replace terrestrial data centers, exploiting continuous solar energy and radiative heat dissipation in a vacuum.
The startup Crusoe will launch the first public space cloud in 2027. Axiom Space is building a network of orbital nodes. And China has already launched twelve satellites from the constellation. Xingsuan, set to reach 2.800 units by 2035. Our data is leaving the planet before we do.
A refrigerator full of silicon goes into orbit
The satellite weighs 60 kilograms, about the same as a refrigerator. Inside, shielded by a silver casing, is the chip: a two thousand teraflop Nvidia H100, the most powerful AI processor ever sent beyond the atmosphere. It will depart in November on a ride courtesy of SpaceX (not true, you have to pay for it), bound for low Earth orbit.
This isn't a niche experiment: It's the first industrial attempt to move data centers off Earth, where solar power is free, cooling costs zero, and regulatory constraints, well… simply don't exist.
Philip JohnstonStarcloud CEO, believes that in ten years, most new data centers will be built in space. And he's not alone: Jeff Bezos shares the vision. Also Crusoe, a startup with a $10 billion valuation, has already signed on to launch orbital servers in 2026. The November test will tell whether they're right or if it's just an expensive idea with too much ambition.
Orbital servers, cosmic cooling, and four-kilometer panels
Land-based data centers have a structural problem. They consume water, a lot of it. A single 40-megawatt plant can use 1,7 million tons of water over ten years for cooling alone. In space, this limit disappears. The cosmic vacuum acts as an infinite heat sink. Heat is expelled via infrared radiation into deep space, without the need for evaporative cooling towers, water circuits, or fans.
According to a study published in Advanced materialsRadiative dissipation in space can handle high heat loads without active systems, provided radiators are designed with high emissivity. Starcloud-1 uses panels with a coefficient greater than 0,9, optimized for far-infrared radiation. The system is designed to operate even under maximum load, when the H100 chip consumes 700 watts.
The long-term plan is even more ambitious. Starcloud wants to build a five-gigawatt data center with solar panels and radiators four kilometers on each side.Sixteen square kilometers of surface area to capture energy and disperse heat. A similar architecture has also been proposed by China, which aims for a network of 2.800 satellites with a total capacity of a thousand petaoperations per second by 2035. My head is starting to spin.
Radiation, debris and failures at 7 kilometers per second
Ok, I told you my ambitions, but are they legitimate? I have doubts, I don't know about you.Satellites operate in a hostile environment: cosmic radiation, extreme temperature fluctuations (from minus 120 to plus 120 degrees Celsius), micrometeorites, space debris. A study published in Light: Science & Applications demonstrated that electronic components in space undergo accelerated degradation due to exposure to cosmic rays and atomic oxygen, which is particularly aggressive in low orbit.
Nvidia and Starcloud claim the H100 chip is sufficiently insulated to withstand the effects. It's unclear whether it's been hardened against radiation or whether the passive shielding is sufficient. The November test will provide concrete answers: operational stability during radiation events, thermal stability under maximum load, and high-speed optical networking to terrestrial and satellite constellations.
Orbital servers, really 10 years to move everything off the planet?
Johnston, as I wrote above, declared that "in ten years, almost all new data centers will be built in space." It's a bold prediction. On Earth, building a 100.000-GPU data center like xAI's in Tennessee It requires months of permits, dedicated electrical infrastructure, and complex cooling systems. In space, regulatory constraints are minimal. No planning permission is required, there's no opposition from local communities, and there's no limit to physical growth.
But there are other obstacles. Launch costs, even with SpaceX Starship promising $50-100 per kilogram, remain significant. Maintenance is complex: no technician can climb a ladder to replace a faulty component. And then there's Kessler syndrome, the risk that a collision will cause a debris flow which renders entire orbits unusable.
Eric Schmidt, the former CEO of Google who acquired space launch startup Relativity Space, told Congress that the growth of AI requires “industrial infrastructure on an unprecedented scale.”
Space, according to Schmidt, can provide those resources without the constraints of Earth. But advanced robotic technology is needed to assemble and maintain large-scale orbital structures, and that technology doesn't yet exist in a mature form.
Starcloud-1 is a test. If the H100 chip holds up, processes data reliably, and demonstrates that radiative cooling works, then the orbital servers They'll go from an experiment to a real infrastructure. If it fails, it will remain an interesting attempt in a sector that has always loved impossible ideas.
We'll know which way we're going in November.
