Space is cold. So if we put an AI data center in orbit, cooling should be easy, right?
There is no hot summer afternoon. No crowded server room. No need to pull cool outside air through giant cooling equipment.
At first, the idea sounds almost obvious: move the computers into cold space and let nature do the cooling.
But there is a catch.
Space is cold, but a vacuum is not a giant air conditioner.
The problem is not whether space is cold. The real problem is much simpler:
How do you actually move the heat away from the computer?
Once that question is clear, one of the hardest engineering problems behind space data centers becomes much easier to understand.
Start with a Laptop on Your Desk
Think about your laptop.
When the processor works hard, it gets hot. A heat sink pulls heat away from the chip. A fan moves air across the hot parts. That warmer air then carries the heat into the room.
A large data center does the same basic job on a much bigger scale.
The details can become complicated, but the heat still follows a path:
Chip → cooling system → outside environment
Modern AI servers often use liquid cooling because high-density GPUs can produce too much heat for ordinary air cooling alone. The liquid picks up heat near the chips and carries it somewhere else.
On Earth, that heat can eventually be transferred to air, cooling water, a cooling tower, or another part of the facility.
Now remove the atmosphere.
That is what changes in space.
Why Vacuum Changes the Cooling Problem
Heat can move in three familiar ways:
- Conduction: heat moves through a solid material.
- Convection: heat is carried by a moving fluid such as air or water.
- Radiation: heat leaves as electromagnetic radiation.
Inside a spacecraft, conduction still works. A pumped liquid loop can also carry heat from one component to another.
But outside the spacecraft there is almost no air.
NASA's 2026 SmallSat thermal-control guide explains the key point directly: in vacuum there is no convection to the outside environment, so external heat exchange depends largely on thermal radiation.
That gives an orbital computer a very different heat-removal chain:
GPU → cold plate → coolant → heat exchanger → radiator → infrared radiation → space
The last step is the difficult one.
You cannot simply blow the heat away with a bigger fan.
A Space Radiator Is Not Like a Car Radiator
The word radiator can be misleading.
A car radiator is cooled mainly because air flows through it. A spacecraft radiator works differently. Its surface emits thermal energy, mostly as infrared radiation.
The basic relationship is described by the Stefan-Boltzmann law:
P = εσAT4
You do not need to memorize the equation.
For this story, remember only three things:
- A larger radiator can reject more heat.
- A hotter radiator can reject much more heat.
- More computing power usually means more waste heat that must be rejected.
The temperature term is raised to the fourth power, so radiator temperature matters a lot.
That sounds helpful. Why not just run the radiator very hot?
Because the computer, coolant, electronics, seals, pumps, and structural materials all have temperature limits. Keeping the radiator hotter can make heat rejection easier, but keeping the chips cool enough becomes harder.
That is the trade-off.
What Does 1 MW of Waste Heat Look Like?
Let us make the scale visible.
An ideal black surface at about 300 K, roughly 27°C, radiates about 459 watts per square meter. A real radiator with an emissivity around 0.85 to 0.90 would be closer to roughly 390 to 410 watts per square meter before we account for real orbital conditions.
If we divide 1 megawatt by that heat flux, we get a simple first estimate:
1 MW of heat → roughly 2,400 to 2,600 m² of radiator area
This is only a back-of-the-envelope calculation. Real spacecraft must also deal with sunlight, radiation from Earth, geometry, temperatures, plumbing, view factors, and safety margins.
But interestingly, a 2026 orbital-data-center study reached almost the same scale. In one representative 1 MW design case, it calculated about 2,500 m² of radiator area.
That is the important point.
A one-megawatt computer does not simply need a one-megawatt power source.
It also needs a way to get rid of almost one megawatt of heat.
Now Compare That with a 100 MW AI Data Center
This is where the numbers become uncomfortable.
In our earlier Contexta articles, we used 100 MW to build intuition about modern AI data-center scale. A terrestrial facility of that size is already a major industrial load.
How Much Power Is 100 MW? An AI Data Center Compared with Entire Cities showed how a single large facility can use electricity on the scale of a city.
We should not simply multiply the 1 MW radiator estimate by 100 and call that a real spacecraft design. Engineers can change operating temperature, architecture, chip efficiency, radiator geometry, and workload.
But the scaling problem does not disappear.
As computing grows, heat rejection becomes a major part of the spacecraft.
And that leads to a new infrastructure chain:
Compute → Heat → Radiator Area → Mass → Launch Cost
This may be one of the most useful ways to think about orbital computing economics.
The Radiator Has to Ride the Rocket Too
A radiator is not a weightless sheet floating beside the computer.
It needs structure. It may need pumps, coolant lines, valves, heat exchangers, sensors, deployment mechanisms, and backup paths.
Large surfaces must survive launch loads and then deploy reliably in orbit. They must also keep working through repeated hot-cold cycles.
More radiator area therefore usually means more spacecraft hardware.
More hardware means more mass.
And every kilogram must somehow reach orbit.
This is why cooling becomes an economic problem, not just a thermal-engineering problem.
NASA's own spacecraft work shows how important radiator mass can become. A NASA study of a multi-megawatt nuclear-electric-propulsion concept described about 2,500 m² of total radiator area for roughly 4 MW of rejected thermal power, with a heat-rejection-system mass measured in tonnes.
Different mission, different temperatures, different technology—but the lesson is useful:
Large heat loads create large physical systems.
We Already Have a Real Example: The ISS
The International Space Station has been solving this problem for years.
Inside the station, water loops collect heat from equipment and the cabin. Heat exchangers transfer that energy to external loops that use liquid ammonia. Pumps move the ammonia through large radiators, where the heat is finally rejected into space.
In simplified form:
Electronics → water loop → heat exchanger → ammonia loop → radiator → space
NASA describes the ISS External Active Thermal Control System as a pumped ammonia system that collects waste heat and rejects it through radiators.
The radiators even need to be pointed properly. NASA documentation describes rotating radiator assemblies to maintain useful heat rejection while controlling other thermal risks.
That is an important detail.
A radiator in orbit is not simply installed and forgotten.
Space Is Cold—and the Sun Is Hot
There is another reason the phrase “cold space” can mislead us.
A spacecraft in Earth orbit is exposed to strong sunlight. It can also receive reflected sunlight and infrared energy from Earth.
So one side of a structure can receive significant heat while another side has a much better view of cold space.
Thermal engineers therefore care about orientation, surface coatings, solar absorptivity, infrared emissivity, shadows, orbital geometry, and what the radiator can “see.”
NASA's thermal-control guidance notes that good radiator surfaces are designed to emit infrared energy efficiently while absorbing as little sunlight as practical.
So the real design question is not:
“Is space cold?”
It is:
“Can we create a reliable path from the hot chip to a radiator that can see a useful thermal sink?”
This Changes How We Value an Efficient AI Chip
There is a more interesting business insight here.
On Earth, reducing a server's power consumption saves electricity and usually lowers cooling demand.
In orbit, one watt saved can potentially help several systems at once.
Lower compute power can mean:
- less solar-array area,
- less stored energy for eclipse periods,
- less waste heat,
- less radiator area,
- less coolant and plumbing,
- less supporting structure,
- and eventually less launch mass.
This is why performance per watt may become even more valuable in space than it is on Earth.
Recent research is already exploring this direction. One 2026 paper on space-oriented compute-in-memory accelerators treats radiator capacity as a direct limit on usable AI performance. Another proposal integrates solar generation, computing, and radiator functions into the same large panel structure to reduce mass.
These are still research concepts, not proof that giant orbital AI campuses are ready.
But they show how the design problem is changing.
A successful space data center may not look like a normal server building placed on a satellite.
It may need to be designed as a spacecraft first and a data center second.
What About Google's Project Suncatcher?
The idea is no longer limited to science fiction.
Google announced Project Suncatcher in November 2025 as a research effort to explore networks of solar-powered satellites equipped with Tensor Processing Units, or TPUs.
Google said it plans a learning mission with Planet to launch two prototype satellites in early 2027 and test hardware in orbit.
This does not mean Google has solved orbital data centers.
It means the engineering questions—radiation, communications, formation flying, power, and thermal control—are moving from thought experiment toward hardware testing.
Cooling remains one of those questions.
The Number to Remember
If you remember only one chain from this article, make it this one:
Compute → Heat → Radiator → Mass → Cost
Every useful computation eventually becomes heat.
On Earth, we can use air, water, cooling towers, chillers, and large surrounding infrastructure to move that heat away.
In orbit, the final step is radiation.
Radiation works. The International Space Station proves that.
But it needs radiator area, thermal plumbing, structure, control, and mass.
So yes, space is cold.
That does not make cooling free.
A better way to think about it is:
Space gives a computer a huge place to radiate heat into, but almost nothing outside the spacecraft can carry that heat away.
Why This Matters
In Why Power, Not Chips, May Limit the AI Data Center Boom, we followed electricity from the grid to the GPU and then out again as heat.
Orbital computing removes part of the terrestrial grid problem, but it does not remove infrastructure.
It replaces one set of constraints with another:
grid connection and land on Earth → launch, power, thermal control, communications, and maintenance in space
That is the deeper lesson of the space-data-center debate.
Moving computing somewhere else does not make the physical world disappear.
What to Watch Next
There may be another way to reduce pressure on giant data centers.
Instead of sending more computers into orbit, what if we move more AI computing onto the devices already in our pockets and on our desks?
Phones and PCs are gaining increasingly capable AI accelerators. Smaller models can sometimes run locally without sending every request to a distant cloud server.
That leads to the next question in this series:
Will Edge AI Reduce the Need for Giant Data Centers?
Key Vocabulary
Vacuum
An environment with extremely little matter. In space, the lack of surrounding air removes the normal outside-air convection used by many cooling systems on Earth.
Thermal radiation
Energy emitted as electromagnetic radiation because an object has a temperature. Spacecraft radiators use this mechanism to reject heat.
Radiator
A spacecraft surface designed to release waste heat by thermal radiation.
Emissivity
A measure of how effectively a surface emits thermal radiation compared with an ideal blackbody.
Heat rejection
The process of moving unwanted heat out of a system and releasing it to the surrounding environment.
Related Articles
- Can We Put AI Data Centers in Space?
- Why Power, Not Chips, May Limit the AI Data Center Boom
- How Much Power Is 100 MW? An AI Data Center Compared with Entire Cities
- How to Understand a 100 MW Data Center Power Model
Sources
- NASA Small Spacecraft Systems Virtual Institute — Thermal Control
- NASA Technical Reports Server — International Space Station Active Thermal Control Sub-System
- NASA Technical Reports Server — Thermal Radiator Pointing for International Space Station
- Google Research — Project Suncatcher
- Turyshev, S. G. — Orbital Data Centers: Spacecraft Constraints and Economic Viability, 2026
- Mugdho, Hasan & Wang — Space-CIM: Enabling Compute-In-Memory Accelerators for Thermally-Constrained Space Platforms, 2026
- Gaalema, Indyk & Staley — Reduced-Mass Orbital AI Inference via Integrated Solar, Compute, and Radiator Panels, 2026
Published: August 2026 · Sources checked through: August 2026 · The radiator-area example is an order-of-magnitude illustration, not a design value for a specific orbital data center.