- Joined
- Jan 20, 2026
- Messages
- 345
- Reaction score
- 2,491
There is a problem that cannot be solved with money.
Companies that need all the new AI capacity have started to look at the orbit. At first glance, the idea looks logical: there is a lot of solar energy in space, you do not need to buy land for a data center, there is no water consumption for cooling and you can process data next to satellites. But the main question does not rest on chips and not in missiles. Servers release heat, and in a vacuum it is difficult to get rid of it.
At the Nvidia GTC conference, CEO Jensen Huang called computing in space a new direction for AI infrastructure. Google is already developing Project Suncatcher: the company, together with Planet, plans to launch two test satellites with TPU tent processors by the beginning of 2027. Starcloud’s startup has submitted to the U.S. Federal Communications Commission for a constellation of up to 88 000 satellites for orbital data centers. This is not yet a ready-made industry, but a set of early projects, tests and applications, but the topic has clearly gone beyond the science fiction.
In the basic version, the orbital data center should look like a network of satellites with computing modules on board. The devices exchange data between them on laser communication lines, and the Earth is communicated directly or through other satellites. Inside may be GPU, TPU or other accelerators similar to those used in terrestrial AI clusters.
Proponents of such projects usually talk about three advantages: solar energy, no terrestrial limitations and proximity to space sensors. But solar panels, servers, batteries, radiators, antennas and orientation systems need to be withdrawn by a rocket. Each extra part increases the mass, the cost of the launch and the risk of failure.
The weakest point is cooling. The space is cold but empty. On Earth, heat from the servers can be carried by air, water or other liquid. In a vacuum, the air does not move, so the usual ventilation does not work. There is a radiator: a large surface that emits heat into space.
The more powerful the chip, the larger the radiator area. The IEEE Spectrum analysis provides a calculation for the Nvidia H100, one of the popular AI accelerators. One such GPU consumes about 700 watts. To keep it at about 60 °C, under ideal conditions, a radiator of about 1.4 square meters is needed. For one server rack with 32 such GPUs, processors, memory and network equipment, the cooling area grows to about 80 square meters.
This is not a stock of a huge data center, but a calculation for one rack. If you try to assemble an orbital data center with a capacity of tens or hundreds of megawatts, the radiators turn into one of the main parts of the structure. They need to be packed in a rocket, deployed in orbit, correctly directed and protected from gradual aging.
The low-Earth orbit is not like a clean laboratory. Ultraviolet, atomic oxygen and charged particles spoil the coatings of radiators. Over time, the surface radiates heat worse, so the area has to be laid with a margin. For one H100, the required area by the end of the five-year service life grows from about 1.4 to 2 square meters. In other words, part of the mass of the satellite should be spent simply so that the equipment does not overheat in a few years.
Solar energy also does not solve the problem on its own. In orbit, the solar radiation flow is really high - about 1361 W per square meter. But the panels need to be delivered into space, deployed, kept at the right angle to the Sun and take into account the degradation. Space solar panels are gradually losing efficiency due to radiation, usually about 1-3% per year.
There is another nuance: almost all the energy that the server spent on calculation turns into heat. If the satellite cluster receives 1 MW of electricity, about 1 MW of heat then need to be discharged through radiators. Therefore, next to the solar panels inevitably appear comparable in importance cooling systems.
Managing such a design is difficult. The panels should look at the Sun, the radiators - in the cold dark side, the antennas - to Earth or neighboring satellites. This requires powerful orientation systems. They take a place, consume energy, add moving nodes and create new points of failure. On Earth, a broken fan or pump can be replaced. In orbit, even a simple breakdown quickly becomes an expensive problem.
A separate difficulty is related to the chips themselves. Conventional satellites often use radiation-resistant processors. They are reliable, but strongly inferior to modern GPU and TPU in performance. For large AI models, such processors are too weak. Therefore, orbital data centers will have to take commercial accelerators, which were originally designed for terrestrial server.
In space, such chips work in a more dangerous environment. High-energy particles can change bits in memory, cause log faults, or damage the chains. Thick radiation protection reduces the risk, but dramatically increases the mass. Another option is to run the same task on multiple nodes and compare the answers. If one node is wrong, the system excludes it and restarts. Reliability is growing, but part of the power goes to the reserve and checking.
The economy is not yet formed for the usual AI-loads. ABI Research compared the year of the GPU on Earth and in orbit. The model used the Nvidia H100 rack, the necessary panels and radiators, the very optimistic Starship launch price of $ 44 per kilogram and the terrestrial electricity price of $0.20 per kilowatt-hour. Even under such conditions, GPUs in space costs at least an order of magnitude more expensive than in a conventional data center.
Other calculations inevitably lead researchers to a similar conclusion. For the orbital cluster with a capacity of 1 MW, only solar panels, energy storage and radiators give tens of kilograms of mass per kilowatt of useful computational load. To this mass you need to add a body, communication, control, protection, start, maintenance and limited service life. Therefore, the transfer of an ordinary cloud for AI to orbit still looks too expensive.
But computing in space may still be needed. Just the first useful tasks will not look like an earthly data center with chatbots. More sense is the processing of data where they appear - next to observation satellites, radars, optical sensors and navigation systems.
Modern Earth observation satellites collect huge amounts of data. Hyperspectral chambers and radiators with synthesized aperture can give hundreds of terabytes of raw information per day. It is difficult to transfer everything to Earth: radio channels are limited, ground stations are occupied, and some of the data is needed quickly. If the satellite itself selects important areas, finds changes and sends a ready-made result, the communication channel will be used much more efficiently.
This scenario is suitable for monitoring earthquakes, infrastructure, ships, fires, military facilities and rapid changes on the surface. Here the value is not the fact of computation in space, but the reduction of delay. The satellite saw the event, processed the data and transmitted the bottom not all the flow, but a specific warning or the desired fragment.
Another practical task is to avoid clashes. The low-Earth orbit is quickly filled with satellites. If the two devices collide, the wreckage can damage other satellites and trigger a chain of accidents. This risk is known as Kessler syndrome.
SpaceX is already regularly performing evasion maneuvers for Starlink. According to the company, the entire group on average performs such a maneuver about once every two minutes. Now a significant part of the calculations is still associated with ground systems. But with the growth in the number of satellites, it will be increasingly difficult to wait for a solution from the Earth.
Larger groups need a shorter cycle: to notice the risk, calculate the trajectories, choose a maneuver and perform it. For a dense orbital network, the account can go not for minutes, but for seconds and milliseconds. Therefore, more powerful on-board computers or computing nodes near the group look more practical than trying to launch conventional AI services in orbit.
Engineers are already looking for ways to cope with heat. One option is foldable radiators. They can be compactly laid at launch, and in orbit to wide light panels. This idea is close to other deployable space structures, including elements of the James Webb telescope.
Another option is drip radiators. The system sprays the coolant in vacuum with small drops, each drop emits heat with the entire surface, then the flow is collected back and pumped again through the contour. The idea is complex, but with megawatt heat loads, hard panels can be too heavy.
There are few new radiators. If orbital data centers become an expensive infrastructure, they can not be simply started and written off after degradation. You will need service devices that can change the radiators, update server modules and extend the life of satellites. Without maintenance, a large orbiting computing network will quickly turn into an expensive set of disposable devices.
Therefore, the near future of cosmic computing is likely to be narrow and applied. The orbit will be carried out those calculations that can not be conveniently performed on Earth due to the delay, congested communication channel or the need for an autonomous solution. Training large models and mass-producing answers to users remain tasks for earth data centers.
The main limit is set by heat. Chips can be made faster, launches - cheaper, solar panels - more efficient. But every watt spent by the server needs to be diverted. On Earth, air, water and serviced infrastructure help with this. In space, you have to run large radiators, monitor their aging and pay for every kilogram of the structure.
Companies that need all the new AI capacity have started to look at the orbit. At first glance, the idea looks logical: there is a lot of solar energy in space, you do not need to buy land for a data center, there is no water consumption for cooling and you can process data next to satellites. But the main question does not rest on chips and not in missiles. Servers release heat, and in a vacuum it is difficult to get rid of it.
At the Nvidia GTC conference, CEO Jensen Huang called computing in space a new direction for AI infrastructure. Google is already developing Project Suncatcher: the company, together with Planet, plans to launch two test satellites with TPU tent processors by the beginning of 2027. Starcloud’s startup has submitted to the U.S. Federal Communications Commission for a constellation of up to 88 000 satellites for orbital data centers. This is not yet a ready-made industry, but a set of early projects, tests and applications, but the topic has clearly gone beyond the science fiction.
In the basic version, the orbital data center should look like a network of satellites with computing modules on board. The devices exchange data between them on laser communication lines, and the Earth is communicated directly or through other satellites. Inside may be GPU, TPU or other accelerators similar to those used in terrestrial AI clusters.
Proponents of such projects usually talk about three advantages: solar energy, no terrestrial limitations and proximity to space sensors. But solar panels, servers, batteries, radiators, antennas and orientation systems need to be withdrawn by a rocket. Each extra part increases the mass, the cost of the launch and the risk of failure.
The weakest point is cooling. The space is cold but empty. On Earth, heat from the servers can be carried by air, water or other liquid. In a vacuum, the air does not move, so the usual ventilation does not work. There is a radiator: a large surface that emits heat into space.
The more powerful the chip, the larger the radiator area. The IEEE Spectrum analysis provides a calculation for the Nvidia H100, one of the popular AI accelerators. One such GPU consumes about 700 watts. To keep it at about 60 °C, under ideal conditions, a radiator of about 1.4 square meters is needed. For one server rack with 32 such GPUs, processors, memory and network equipment, the cooling area grows to about 80 square meters.
This is not a stock of a huge data center, but a calculation for one rack. If you try to assemble an orbital data center with a capacity of tens or hundreds of megawatts, the radiators turn into one of the main parts of the structure. They need to be packed in a rocket, deployed in orbit, correctly directed and protected from gradual aging.
The low-Earth orbit is not like a clean laboratory. Ultraviolet, atomic oxygen and charged particles spoil the coatings of radiators. Over time, the surface radiates heat worse, so the area has to be laid with a margin. For one H100, the required area by the end of the five-year service life grows from about 1.4 to 2 square meters. In other words, part of the mass of the satellite should be spent simply so that the equipment does not overheat in a few years.
Solar energy also does not solve the problem on its own. In orbit, the solar radiation flow is really high - about 1361 W per square meter. But the panels need to be delivered into space, deployed, kept at the right angle to the Sun and take into account the degradation. Space solar panels are gradually losing efficiency due to radiation, usually about 1-3% per year.
There is another nuance: almost all the energy that the server spent on calculation turns into heat. If the satellite cluster receives 1 MW of electricity, about 1 MW of heat then need to be discharged through radiators. Therefore, next to the solar panels inevitably appear comparable in importance cooling systems.
Managing such a design is difficult. The panels should look at the Sun, the radiators - in the cold dark side, the antennas - to Earth or neighboring satellites. This requires powerful orientation systems. They take a place, consume energy, add moving nodes and create new points of failure. On Earth, a broken fan or pump can be replaced. In orbit, even a simple breakdown quickly becomes an expensive problem.
A separate difficulty is related to the chips themselves. Conventional satellites often use radiation-resistant processors. They are reliable, but strongly inferior to modern GPU and TPU in performance. For large AI models, such processors are too weak. Therefore, orbital data centers will have to take commercial accelerators, which were originally designed for terrestrial server.
In space, such chips work in a more dangerous environment. High-energy particles can change bits in memory, cause log faults, or damage the chains. Thick radiation protection reduces the risk, but dramatically increases the mass. Another option is to run the same task on multiple nodes and compare the answers. If one node is wrong, the system excludes it and restarts. Reliability is growing, but part of the power goes to the reserve and checking.
The economy is not yet formed for the usual AI-loads. ABI Research compared the year of the GPU on Earth and in orbit. The model used the Nvidia H100 rack, the necessary panels and radiators, the very optimistic Starship launch price of $ 44 per kilogram and the terrestrial electricity price of $0.20 per kilowatt-hour. Even under such conditions, GPUs in space costs at least an order of magnitude more expensive than in a conventional data center.
Other calculations inevitably lead researchers to a similar conclusion. For the orbital cluster with a capacity of 1 MW, only solar panels, energy storage and radiators give tens of kilograms of mass per kilowatt of useful computational load. To this mass you need to add a body, communication, control, protection, start, maintenance and limited service life. Therefore, the transfer of an ordinary cloud for AI to orbit still looks too expensive.
But computing in space may still be needed. Just the first useful tasks will not look like an earthly data center with chatbots. More sense is the processing of data where they appear - next to observation satellites, radars, optical sensors and navigation systems.
Modern Earth observation satellites collect huge amounts of data. Hyperspectral chambers and radiators with synthesized aperture can give hundreds of terabytes of raw information per day. It is difficult to transfer everything to Earth: radio channels are limited, ground stations are occupied, and some of the data is needed quickly. If the satellite itself selects important areas, finds changes and sends a ready-made result, the communication channel will be used much more efficiently.
This scenario is suitable for monitoring earthquakes, infrastructure, ships, fires, military facilities and rapid changes on the surface. Here the value is not the fact of computation in space, but the reduction of delay. The satellite saw the event, processed the data and transmitted the bottom not all the flow, but a specific warning or the desired fragment.
Another practical task is to avoid clashes. The low-Earth orbit is quickly filled with satellites. If the two devices collide, the wreckage can damage other satellites and trigger a chain of accidents. This risk is known as Kessler syndrome.
SpaceX is already regularly performing evasion maneuvers for Starlink. According to the company, the entire group on average performs such a maneuver about once every two minutes. Now a significant part of the calculations is still associated with ground systems. But with the growth in the number of satellites, it will be increasingly difficult to wait for a solution from the Earth.
Larger groups need a shorter cycle: to notice the risk, calculate the trajectories, choose a maneuver and perform it. For a dense orbital network, the account can go not for minutes, but for seconds and milliseconds. Therefore, more powerful on-board computers or computing nodes near the group look more practical than trying to launch conventional AI services in orbit.
Engineers are already looking for ways to cope with heat. One option is foldable radiators. They can be compactly laid at launch, and in orbit to wide light panels. This idea is close to other deployable space structures, including elements of the James Webb telescope.
Another option is drip radiators. The system sprays the coolant in vacuum with small drops, each drop emits heat with the entire surface, then the flow is collected back and pumped again through the contour. The idea is complex, but with megawatt heat loads, hard panels can be too heavy.
There are few new radiators. If orbital data centers become an expensive infrastructure, they can not be simply started and written off after degradation. You will need service devices that can change the radiators, update server modules and extend the life of satellites. Without maintenance, a large orbiting computing network will quickly turn into an expensive set of disposable devices.
Therefore, the near future of cosmic computing is likely to be narrow and applied. The orbit will be carried out those calculations that can not be conveniently performed on Earth due to the delay, congested communication channel or the need for an autonomous solution. Training large models and mass-producing answers to users remain tasks for earth data centers.
The main limit is set by heat. Chips can be made faster, launches - cheaper, solar panels - more efficient. But every watt spent by the server needs to be diverted. On Earth, air, water and serviced infrastructure help with this. In space, you have to run large radiators, monitor their aging and pay for every kilogram of the structure.
