Space Computing: The Physics and Challenges of Orbital Data Centers

Orbital data centers face a “physics tax” that makes them at least 10 times more expensive to operate than terrestrial facilities, according to an analysis by ABI Research. While companies including Google, SpaceX, and Starcloud are planning satellite constellations for AI compute, fundamental challenges in radiative cooling and ionizing radiation currently limit their economic viability to niche applications like real-time collision avoidance and Earth-observation preprocessing.

Nvidia CEO Jensen Huang announced in March at the Nvidia GTC conference that “space computing, the final frontier, has arrived.” This shift is reflected in current industry plans: Google announced Project Suncatcher, a partnership with Planet to launch two satellites equipped with Google Tensor Processing Unit (TPU) AI chips by early 2027. Simultaneously, Elon Musk’s SpaceX has acquired xAI to plan a constellation of space-based data centers, and the startup Starcloud has filed a proposal with the Federal Communications Commission for a constellation of 88,000 satellites.

These proposed fleets would house racks of AI-grade GPUs interconnected via free-space optical links, communicating with Earth through microwave links. However, Chris Philpot, an aerospace analyst at ABI Research, reports that the cost of delivery and space-hardening makes general-purpose orbital data centers difficult to justify economically today.

Why is cooling a primary barrier for space-based AI?

Cooling in the vacuum of space relies entirely on radiation because conduction and convection are impossible without an atmosphere. According to Philpot, this is governed by the Stefan-Boltzmann Law, which dictates that the power a system can radiate is proportional to the radiator’s area and its temperature to the fourth power.

This creates a massive geometric penalty for high-power hardware. A single Nvidia H100 GPU drawing 700 watts requires 1.4 square meters of radiator surface to maintain a stable operating temperature of 60 °C. When scaled to a common AI rack containing 32 GPUs and drawing 40 kilowatts, the required radiator surface expands to 80 square meters, which Philpot compares to the size of a pickleball court.

Environmental degradation further increases these requirements. Ultraviolet light and atomic oxygen in low Earth orbit (LEO) degrade thermal coatings over time. Philpot’s model indicates that after five years in space, the required radiator area for a single chip must increase by 40 percent, from 1.4 to nearly 2.0 square meters, to maintain the same temperature.

How does radiation affect orbital silicon?

Standard terrestrial chips like the H100 or Google TPU are “soft” targets for ionizing radiation, which can cause bit-flips in memory or circuit-frying “latch-ups.” While radiation-hardened processors exist, Philpot notes they lack the processing power required to run modern large language models (LLMs) and are significantly more expensive.

To counter this, architects are using software-defined resilience. Instead of a single hardened computer, they deploy clusters of commercial chips and run the same calculation on multiple nodes to detect corruption. This redundancy approach is already utilized in SpaceX flight computers, Hewlett Packard Enterprise edge servers for the International Space Station, and the Artemis II moon mission.

What is the total cost of ownership for space GPUs?

An ABI Research total-cost-of-ownership comparison shows that launching and running a GPU in space for one year costs at least an order of magnitude more than a terrestrial equivalent. The model assumed a highly optimistic launch cost of $44 per kilogram using SpaceX’s Starship and a terrestrial energy cost of $0.20 per kilowatt hour.

Energy collection also adds complexity. While solar energy is abundant at 1,361 watts per square meter, panels degrade by 1 to 3 percent annually. Furthermore, maintaining the necessary three-way alignment—panels facing the sun, radiators facing the void, and antennas facing Earth—requires complex, high-torque attitude control systems that add mass and potential failure points to the spacecraft.

Which applications justify the cost of orbital compute?

Despite the costs, Philpot identifies two “killer apps” where the ability to generate insight in orbit outweighs the expense:

• Downlink Bottleneck Reduction: Earth-observation satellites generate hundreds of terabytes of raw data daily. Onboard processing allows satellites to downlink only relevant data, bypassing congested radio-frequency pipes.
• Real-time Collision Avoidance: SpaceX’s annual report indicates Starlink executes a collision avoidance maneuver every two minutes on average. Moving the “OODA” (observe, orient, decide, act) loop onboard can reduce analysis turnaround from minutes to milliseconds.

How will future orbital data centers evolve?

Industry engineers are exploring two primary thermal solutions to mitigate the “physics tax.” One involves origami-inspired radiators, similar to those on the James Webb telescope, which fold for launch and unfurl into lightweight thermal wings. Another is the liquid-droplet radiator, which sprays coolant oil directly into the vacuum of space to maximize surface area before collecting the fluid.

The transition to radiation-tolerant perovskite solar panels could further alter economics, though Philpot suggests this technology is at least five years away. Ultimately, the move toward “autonomous logistics”—using servicing vehicles to swap degraded panels and fried servers—will be necessary to prevent multimillion-dollar data centers from becoming space junk.