The Data Center in Orbit Has a Heat Problem

Google's first orbital test of AI hardware is scheduled for the coming week. It is a small experiment with an unusually large implied question: could a data center someday be assembled from satellites that draw sunlight directly in orbit? Google's September 24 update says a prototype built with Planet is due aboard SpaceX's Transporter-18 rideshare. The mission will expose Trillium tensor processing units to launch vibration, radiation and thermal conditions. It is a hardware-survival test, not the opening of a commercial orbital compute service. Google plans to test a pair of laser-linked satellites in 2027; the first flight cannot establish that a cluster will communicate or run a large model economically.

The idea is compelling because power is a central constraint on terrestrial AI expansion. A spacecraft in a suitable low-Earth orbit can see near-constant sun and, Google estimates, collect up to eight times as much solar energy as an equivalent Earth installation over comparable periods. But a source of electricity does not remove the need to reject the heat generated by using it. Earthly data centers move heat into air or water. An orbital computer must ultimately radiate it into space, while carrying the mass and exposed surface area needed for that job. The experiment is best understood as an attempt to determine whether the underlying physical ledger can ever balance.

First, survive the trip

Launch imposes an intense mechanical test before the chip sees a single inference request. Google reports up to roughly 10 g sustained acceleration during the ascent and local loads on components reaching 50 to 100 g. It shook the satellite along three axes in ground testing. A successful bench test says the assembled hardware survived a specified simulation; actual launch will tell engineers whether mounting, connections and packaging withstand the ride together. A fault in an ordinary data center can prompt a technician visit. A fault in orbit changes the economics of redundancy and replacement.

Radiation is a second distinction. Charged particles can flip bits, accumulate damage in electronics, or interrupt a system long enough to spoil a workload. Google says it ran Trillium TPUs under a proton beam at the University of California, Davis, while monitoring computational errors. It reports that the tested chips tolerated a total ionizing dose greater than its estimate for a five-year mission. That is the company's result under its specified test conditions, not independent proof of a five-year reliable data-center service. Dose tolerance is only one question; single-event upsets, error detection, memory protection, solar weather and the shielding of an assembled satellite all shape actual performance.

The September update is relatively candid about why it wants flight data. A thermal-vacuum chamber simulates some conditions, but the combination of orbital cycles, sunlight angles, radiation and real workloads cannot be reproduced in full on the ground. This is why the first mission matters even if it never runs an economically useful model. It can replace some assumptions with measured failure rates and thermal behavior. It will leave many scaling questions open.

Sunlight in does not mean heat disappears

Solar panels produce electrical power from part of the sunlight they absorb. Chips then turn almost all their consumed electrical energy into heat. In a terrestrial facility, fans and liquid loops take that heat to the outside environment. Space is a vacuum: there is no surrounding air to carry it away by convection. Heat pipes can spread energy from a hot chip to a radiator, but a radiator must present enough surface and reach a suitable temperature to emit the heat as infrared radiation. The thermal architecture becomes a large, vulnerable part of the computing machine.

Google says it has tested heat pipes and radiator approaches in a thermal-vacuum chamber, and the flight will check how the new cooling system behaves in orbit. The power density of an accelerator makes this difficult: a small region can get hot faster than heat can be transported across the structure. An engineering design may throttle the processor, add larger radiators, raise allowable temperatures, or limit how many chips operate simultaneously. Each option changes output per kilogram and therefore the economics of launch.

There is a geometry problem too. A solar panel wants to face the sun; a radiator generally wants a cold view of space and must avoid absorbing too much solar energy itself. A cluster must maintain that orientation while its orbit changes relative to Earth and sun. Earthshine and eclipses complicate the thermal cycle. More radiator area can mean more launch mass, larger structures and a harder spacecraft attitude-control task. An eightfold solar-energy comparison does not translate into eight times the useful computation, because the conversion, storage, distribution and waste-heat constraints each take a share.

One can put the energy into a simple thought experiment without pretending to know Google's design. If a satellite draws one kilowatt continuously to run chips and supporting electronics, approximately one kilowatt of heat has to leave the system on average, apart from small energy sent away as transmitted signals. A burst workload can raise the local thermal load faster than the average radiator can respond. For a constellation with dozens of accelerators per satellite, the thermal bill scales with every added chip. The question is not whether space is cold; it is how fast a physical surface can radiate the energy away.

A cluster needs a network, not just a launch slot

A modern AI workload spreads computation across chips. Training in particular involves heavy exchanges of intermediate values and model state; some inference services also demand fast communication among accelerators. A group of satellites could in principle act as one compute fabric, but only if its links offer enough bandwidth, low enough delay and reliable enough synchronization for the chosen workload. A collection of isolated processors in orbit is a different product from an orbital data center.

Google proposes optical links between nearby spacecraft. Laser communication has been demonstrated in space, but Google says its intended links must combine unusually high bandwidth with short distances and precise tracking between moving satellites. Its September update describes a 2027 mission with two satellites to test that step. A demonstration of a link will itself be narrower than a proof of a many-node training cluster: scaling introduces routing, contention, loss recovery, time synchronization and reconfiguration when a satellite fails or drifts.

The orbital geometry also affects ground service. A low-Earth-orbit satellite passes above a ground terminal for a limited portion of each orbit. Serving terrestrial customers continuously would require relay links, a constellation, multiple ground stations or a workload that can tolerate delays. Data ingress and egress may make an orbital facility particularly ill-suited to jobs that depend on rapid movement of large Earth-based datasets. A more plausible early application might involve computation on data already generated in space, or experiments where the value of available solar power outweighs communication cost. These are our inferences, not declared Google products.

Orbit is not free real estate. Each launch adds spacecraft that need collision avoidance and eventual disposal. A large formation increases tracking, debris and regulatory demands. Radiation tolerance and redundant systems increase mass; a small satellite still needs power management, thermal control, communication and propulsion or other collision-avoidance provisions. Those supporting systems compete with compute chips for every kilogram launched. Terrestrial data centers face land and grid limits, but can be repaired, upgraded and connected by fiber much more easily.

The calculation is relative, not absolute

Google's earlier Project Suncatcher system-design discussion explored whether declining launch prices and strong orbital solar production might someday change that balance. It described a research pathway, with assumptions about how hardware can be packaged and interconnected. The 2026 prototype is a way to test some of those assumptions. Neither the paper nor the upcoming launch establishes a production cost per useful GPU or TPU hour that beats a terrestrial alternative.

Any economic comparison needs matching boundaries. Count the satellite bus, radiators, solar panels, batteries or storage, launch, ground network, replacements, loss rate and deorbit costs against the useful compute delivered to a customer. Compare that with a terrestrial facility including grid connection, energy, cooling, maintenance and chip replacement. A fair comparison also asks what job each system performs. A satellite with abundant sunlight that spends most of its time waiting for data or throttling from heat can be more expensive per useful computation than a terrestrial machine with intermittent renewable power.

The environmental claim has a similar boundary problem. Moving computation away from a local water system could reduce that particular burden. It does not erase the emissions and material footprint of manufacturing and launching a constellation, or the effects of adding objects to crowded orbits. Conversely, terrestrial data centers are not all equally water-intensive or powered by the same grid mix. A credible sustainability assessment would publish a life-cycle inventory and compare it with facilities serving the same workloads. At present, Google's public update is a technology experiment, not such an assessment.

What a useful result would look like

The near-term scorecard is specific: does the chip run correctly after launch; how often do radiation-induced errors appear under real workloads; how well does heat move to the radiators across orbital cycles; and how much power remains available after the satellite's basic operations? The 2027 scorecard adds optical throughput, pointing stability and the ability to sustain a distributed workload between two nodes. A later, larger demonstration would need to report useful compute per unit mass, time and money.

One satellite cannot answer the final question. It can identify a physical showstopper or give engineers reason to proceed. Google itself says the initial flight is about finding points of failure, while the two-satellite laser test belongs to the next stage. As of September 25, the orbital test is planned, not completed. Launch timing can move, and the measured results are still ahead. A headline declaring that Google has put a data center in space would be premature.

The real promise of Project Suncatcher is a testable idea: if orbital sunlight, reliable chips, heat rejection and fast inter-satellite links work together at a sensible cost, part of computing infrastructure might move above the grid. The real discipline is keeping all four conditions in the same sentence. Solar power alone cannot make the data center orbit.