TL;DR · 30-second read
The Short Version
Space is freezing, yet keeping a computer cool there is surprisingly hard. With no air or water to carry heat away, the heat has to radiate off huge panels.
One analyst calculates that a single cabinet of artificial intelligence chips needs a panel about the size of a pickleball court, and every panel has to ride up on a rocket.
SpaceX and Google are both pursuing computers in orbit. The analyst’s rough estimate: running one of these chips in space for a year costs at least ten times what it costs on Earth.
IEEE Spectrum reported on June 11, 2026, in a feature by ABI Research aerospace analyst Andrew Cavalier, that a 40-kilowatt AI server rack in orbit would need an 80-square-meter radiator to stay at 60 °C, and that a 100-megawatt orbital data center would need at least 2,500 such radiators. Cavalier also describes an ABI Research estimate that launching and running a GPU, the specialized chip behind most AI computing, in space for a year costs at least an order of magnitude more than running one on Earth.
The analysis arrives as SpaceX, Google and the startup Starcloud advance orbital computing plans ranging from a two-satellite Google demonstration planned by early 2027 to a SpaceX proposal for up to 1 million satellites.
Executive Summary
Orbital data centers are often pitched on three advantages: abundant solar power, free cooling, and freedom from problems on the ground such as earthquakes, floods and protesters. Cavalier’s calculations challenge the cooling claim most directly. Space is cold, but it has no air or water to carry heat away, so a satellite can shed heat only by radiating it from a surface. That surface must be built, launched and kept working for the life of the mission.
This matters because the pitch has moved beyond concept. Nvidia CEO Jensen Huang declared at the company’s GTC conference in March that ‘space computing, the final frontier, has arrived.’ SpaceX has acquired xAI and is planning a constellation, Google has announced Project Suncatcher with Planet, and Starcloud has filed with the Federal Communications Commission for an 88,000-satellite constellation.
ABI’s conclusion is narrower than a blanket dismissal. It finds general-purpose orbital data centers difficult to justify economically today, while identifying niches where the premium could pay off, such as preprocessing Earth-observation data, tracking hypersonic missiles and avoiding collisions in low Earth orbit.
Free Cooling Becomes Cargo
The governing physics is the Stefan-Boltzmann law: the power a surface can radiate depends on its area and on its temperature raised to the fourth power. Cavalier’s model holds a 700-watt Nvidia H100 chip at 60 °C, described as the sweet spot for GPU longevity and stability, with the radiator facing deep space. The result is 1.4 square meters of radiator per chip. A typical 32-GPU rack, drawing about 40 kilowatts once CPUs, memory and networking are included, needs 80 square meters, roughly a pickleball court.
Scaling is where cooling becomes a logistics problem. At least 2,500 of those radiators are needed for 100 megawatts, which multiplies out to about 200,000 square meters of radiating surface. Ultraviolet light and atomic oxygen, a corrosive form of oxygen found in low Earth orbit, then degrade the surface over a typical five-year satellite life. Holding the same temperature at end of life takes nearly 2.0 square meters per chip, a 40 percent increase. Because maintenance is difficult in orbit, that margin has to be launched on day one, which on these figures implies about 280,000 square meters for the same 100 megawatts.
That is the mechanism behind the headline. On Earth, cooling is mostly an equipment and operating expense. In orbit, it also becomes launch mass, launch volume and additional atmospheric drag, costs that are incurred before a satellite has processed a single AI query. Those affected first are the spacecraft architects and financiers who must pay to fly the extra hardware.
Power and Heat Claim Nearly Equal Real Estate
Solar power does not avoid the problem, because the electricity a satellite collects eventually leaves as waste heat. Cavalier puts usable generation at about 400 watts per square meter and heat rejection at about 450 watts per square meter, so each square meter of solar array demands approximately another square meter of radiator. At 400 watts per square meter, 100 megawatts implies about 250,000 square meters of solar panels alongside the radiators. Those panels typically lose 1 to 3 percent of their efficiency each year.
Elon Musk noted at Davos that the most efficient radiator is one that never sees the sun. That points to the right configuration, but it adds another engineering burden: panels must face the sun, radiators the void and antennas the Earth, all at once. Holding that alignment takes complex, high-torque attitude control systems, the hardware that sets a spacecraft’s orientation. Those systems add payload and failure modes in a setting where repairs are hard.
Temperature offers one lever. Cavalier’s figures show the required area falling from just under 3 square meters per chip at 20 °C to 1 square meter at 85 °C, so hotter chips need less radiator. The trade-off is that 60 °C is the point associated with longevity and stability, and a chip that fails early in orbit cannot easily be swapped out. Radiator mass and chip life therefore become one engineering decision.
Redundancy Multiplies the Heat
Radiation adds a second cost that compounds the first. Radiation-hardened processors are reliable but expensive, and they lack the power to run a modern large language model. Orbital operators are therefore turning to the same commercial H100s and Google TPUs used on the ground, which are vulnerable to flipped memory bits and to ‘latch-ups,’ faults that can destroy a circuit. Shielding adds weight, so the leading alternative is redundancy: run three or more copies of a calculation and compare the answers, as Artemis II and SpaceX’s flight computers do.
Clusters of commercial chips could cost as little as one-tenth to one-hundredth as much as the radiation-hardened approach, a real saving. But every redundant chip draws power and generates heat like any other. If a workload runs in triplicate, the operator is powering and cooling three chips to get one chip’s useful answer. For a data center priced on useful computing rather than installed hardware, that increases the radiator and solar burden per result.
The Cost Models Point in Opposite Directions
ABI’s model is explicitly a back-of-the-envelope exercise. It assumes an H100 rack on a spacecraft resembling Starcloud’s pilot, SpaceX’s Starship at a highly optimistic launch price of $44 per kilogram, and terrestrial electricity at $0.20 per kilowatt-hour. Even with that launch price, a GPU-year in orbit comes out at least an order of magnitude more expensive. Google’s own paper, by contrast, places cost parity in the mid-2030s, when launch falls below $200 per kilogram. Starcloud models about $8 million per 40-megawatt cluster over 10 years, against $167 million on the ground. Musk expects parity in two to three years.
These estimates differ in scope, so they cannot be reconciled from the headline figures, and each deserves testing. The ABI model is simple and built around one current chip, while SpaceX proposes a custom D3 processor. A flat electricity price also cannot fully capture the value of power to builders in regions where they are scrambling to secure it. Still, the spread among the projections suggests that the assumptions matter more than any single result, and radiator mass and lifespan are among the first inputs to examine.
The near-term case is strongest where data already sits in orbit. Earth-observation satellites using hyperspectral and synthetic aperture radar sensors can generate hundreds of terabytes of raw data per day, and the radio links that send it down are congested. Processing it in orbit addresses a constraint that a cheaper data center on the ground cannot, which is a different proposition from moving mainstream AI workloads off Earth.
Background
A data center’s AI racks convert almost all the electricity they consume into heat, so power and cooling tend to limit how much computing a facility can deliver. On Earth, data center builders in many regions are scrambling for electric power, and that pressure has helped turn orbital computing from science fiction into a serious spending category. The proposals generally involve fleets of satellites that each carry one or more racks of AI chips, linked to each other by lasers known as free-space optical links and to the ground by microwave links.
The three leading plans differ widely. SpaceX has made a Federal Communications Commission filing for up to 1 million satellites and published an AI1 design. Starcloud launched its first H100-equipped satellite in late 2025, with a second due in October 2026. Google’s Project Suncatcher remains in a research phase ahead of a two-satellite demonstration with Planet. Most plans favor low Earth orbit, and Starcloud and Google specify dawn-dusk sun-synchronous paths; Starcloud says its path keeps satellites in sunlight more than 99 percent of the time. Source: The Real Cost Of Cooling GPUs In Space Might Shock You, an IEEE Spectrum feature by Andrew Cavalier of ABI Research on the physics and economics of orbital data centers.Sources