Project Suncatcher: AI in Space and the Thermodynamics of Cooling Orbital Data Centers

Direct answer: space is not a freezer

Vacuum contains no air to convect heat away from a satellite's exterior. A computing chip still dissipates electrical power as heat. That heat has to cross thermal interfaces and be conducted or transported to an emitting surface; the radiator then rejects it as infrared radiation.

The low temperature associated with deep space is not a cold breeze over the processor. Radiator temperature, area, surface properties, orientation and the Sun and Earth in its field of view determine the net heat flow. A chip can overheat in vacuum when the thermal path or emitting area is insufficient.

Project Suncatcher thus raises a measurable system question: can orbital solar energy support useful computation while its heat is rejected continuously and reliably? The answer has yet to be established at scale.

What Project Suncatcher is, and what it is not

Google introduced a long-term research moonshot for scalable space-based machine-learning infrastructure in November 2025. Its 24 September 2026 update described an initial orbital learning mission with Planet on Transporter-18 to study chips and thermal hardware. As of this article's 30 September check, the official update still described that flight as upcoming and published no mission outcome; schedules may change.

A learning prototype is not an operating commercial orbital data center. Google plans a separate two-satellite demonstration of high-bandwidth laser links in 2027. Ground tests, a first flight, a network demonstration and scaled infrastructure are different stages.

Why put AI in orbit?

A dawn-dusk sun-synchronous low Earth orbit can offer nearly continuous solar exposure. Researchers also examine computing near space-generated data, reducing direct reliance on a terrestrial grid connection, and potentially lowering local land and cooling-water demand. These are conditional advantages, not measured benefits of a commercial system.

Mass, launch loads, shielding, network access, storage and end-of-life disposal add constraints. A valid comparison must account for the entire system and its lifecycle.

Thermodynamics: Earth versus orbit

A terrestrial campus moves heat with air, water and refrigerants to chillers, cooling towers or dry coolers, ultimately rejecting it to the surroundings. An orbital craft can still conduct heat or circulate fluid internally, but its external rejection is primarily radiative. Vacuum removes external convection, not heat generation.

From TPU to deep space: the heat path

A conceptual chain runs from TPU to a thermal interface, a structural or cold-plate transfer stage, a heat pipe or fluid loop, a radiator and infrared photons. This describes functions, not an undisclosed Google flight drawing.

Heat transport and heat rejection are distinct. A heat pipe moves energy to an emitting surface; the radiator emits it. An interface bottleneck, solar loading or Earth infrared exposure can compromise the chain even if another component is enlarged.

A spacecraft energy balance

Conceptually, absorbed sunlight + Earth-reflected sunlight (albedo) + Earth infrared + electrical dissipation = stored thermal energy + emitted thermal radiation. Storage changes temperature during transients; at a quasi-steady state the inputs must balance the rejected output.

Solar absorptivity and infrared emissivity are different surface properties. Attitude, shadows and view factors determine what the radiator sees. A radiator facing Earth or obscured by another craft is not equivalent to an unobstructed view of deep space.

Stefan–Boltzmann and radiator area

For an idealized illustration, P ≈ εσA(T_rad⁴ − T_sink⁴), where P is thermal power, ε emissivity, σ the Stefan–Boltzmann constant, A effective emitting area and T absolute temperature. With an ideal deep-space sink near 0 K and no solar or planetary load, P/A ≈ εσT_rad⁴.

The fourth-power dependence is powerful, but chip limits, working fluids and interfaces constrain allowable temperatures. This stripped-down equation does not design radiator mass, deployment, orientation or margins; a real orbit needs a detailed thermal model.

Illustrative example: rejecting 1 MW

With ε = 0.9 and σ ≈ 5.67 × 10⁻⁸ W·m⁻²·K⁻⁴, ignoring external loads, a 300 K radiator emits about 413 W/m² and needs about 2,419 m² per megawatt. At 350 K the figures are 766 W/m² and 1,306 m²; at 400 K, 1,306 W/m² and 765 m². These rounded values use effective emitting area and do not represent Google's hardware or its power.

A hotter radiator reduces mathematical area, but the entire thermal path must tolerate that temperature and maintain reliability. Mass and deployable geometry are part of the system design.

Heat pipes connect chips and radiators

A heat pipe often evaporates a working fluid at the hot end, transports vapor to a cooler section, condenses it and returns liquid through a wick or another driving mechanism. Its passive thermal transport has capacity and operating limits.

Google reports thermal-vacuum tests with heat pipes and radiators. That does not reveal the final type or geometry of a constellation. The pipe carries heat; the radiator rejects it to space.

A satellite cluster is also a thermal geometry

The revised preprint uses an illustrative 81-satellite cluster within roughly a 1 km radius in sun-synchronous low Earth orbit. This is a studied architecture, not the initial flight configuration or final product. Close formation helps investigate dense optical links but complicates attitude control and collision avoidance.

A neighboring panel can obstruct a radiator's view or return infrared energy. Pointing, shadowing, optical links and heat rejection must therefore be designed together.

Solar energy and its conditions

For certain orbital and terrestrial comparison assumptions, the study estimates up to eight times the annual solar energy per unit panel area relative to a mid-latitude terrestrial panel. That is annual exposure, not an eightfold improvement in solar-cell conversion efficiency. Eclipses, degradation, orientation, storage and conversion losses still matter.

Radiation: survival is not data integrity

Total ionizing dose (TID) builds up over time. Single-event effects (SEE) may flip bits or corrupt computations. Google's proton-beam tests found Trillium TPU TID tolerance above a modeled five-year requirement, while high-bandwidth memory was more sensitive; SEE and silent data corruption remain engineering problems.

Training, inference and persistent data have different error tolerances. Error correction, checks, redundancy and recovery need application-specific design. A ground test is not five years of demonstrated orbital reliability.

Launch mechanics, laser links and formation

Google describes launch loading of around 10 g at spacecraft level and potentially 50–100 g at components. Three-axis vibration testing, mounting and connector qualification are part of the hardware problem.

Distributed training requires very high throughput and low latency. The preprint examines links on the order of 10 Tbps and reports an 800 Gbps one-way or 1.6 Tbps bidirectional bench demonstration. These are research results, not achieved rates between the initial mission satellites; pointing and orbital availability remain to be proven.

Maintenance, mass and economics

On Earth, technicians replace boards, pumps and transceivers. In orbit, a failed node may require isolation and workload redistribution. Fault tolerance, modularity, replacement, debris control and end-of-life planning all affect lifecycle costs.

The preprint models a launch-price scenario at or below US$200/kg to low Earth orbit in the mid-2030s under learning-curve and reuse assumptions. It is neither a contracted price nor a forecast. Every radiator, shield and cable adds mass to each watt of useful computing.

Environmental benefits require evidence

Orbital computation might displace local cooling-water and land demand for a particular terrestrial workload, but manufacturing, launches, replenishment and deorbiting have their own impacts. The European ASCEND feasibility work conditions potential climate benefits on far less emitting launch systems and full lifecycle accounting. Reduced local water use is not zero environmental impact.

What must work before a large orbital data center

Orbital tests must establish thermal rejection under actual solar and Earth loads, radiation tolerance with data integrity, power availability, optical links and safe formation flight. Multi-year reliability, failures, maintenance or replacement, cost per computation, ground connectivity, disposal and lifecycle impact follow. A successful TPU flight test would answer only part of that system question.

• Radiator area, orientation, view and mass
• TID/SEE tolerance and error detection
• Real orbital power and storage
• Measured laser availability and formation safety
• Fault tolerance, deorbiting and lifecycle economics

What orbital compute teaches terrestrial engineering

On Earth the bottleneck may be delivering megawatts to a campus; in orbit it becomes delivering heat to a radiator and rejecting it. This complements the questions of grid capacity, cooling and physical reliability faced by terrestrial AI data centers.

The cloud may leave Earth; thermodynamics does not. Suncatcher's first value is to turn a long-term vision into measurable questions: how much computing, what emitting area, what reliability and at what cost?

AI data centers and the grid →

Engineering AI data centers on Earth →