The artificial intelligence industry is beginning to come up against the limitations of on-premises infrastructure. Global energy consumption by data centres is set to rise from around 415 TWh in 2024 to 945 TWh in 2030. AI-optimised servers account for almost half of this increase. The problem is no longer solely the availability of processors. Energy, network connections, cooling, land and the pace of permit approvals are becoming equally important.
In this context, space-based data centres are no longer merely a technological curiosity. In February 2026, SpaceX acquired xAI, bringing together rockets, a satellite network and one of the world’s largest computing clusters under a single organisation. Google is developing Project Suncatcher: a constellation of satellites equipped with Tensor Processing Units and connected via optical links. Two prototypes are set to be launched into orbit in early 2027.
The scale of the applications submitted shows that companies are not merely competing for experimental missions. Starcloud has applied to the US Federal Communications Commission for approval of a system comprising up to 88,000 satellites positioned at altitudes of between 600 and 850 kilometres. The constellation would support AI models and cloud services. For the time being, this is a regulatory proposal, not an approved investment project.
Nvidia has also identified orbital computing as a market worth entering. In March, it unveiled the Space-1 Vera Rubin module, designed for data processing, AI inference and the autonomous operation of spacecraft. The company claims that the new chip delivers up to 25 times more computational power for inference than the H100. However, the product does not solve the most challenging part of the problem: heat dissipation.
A vacuum does not cool servers. The absence of an atmosphere precludes convection, which plays a fundamental role in terrestrial data centres. Thermal energy can be dissipated mainly through radiation, and its efficiency depends on temperature and the surface area of the heat sink. According to ABI Research, a single Nvidia H100 processor would require approximately 1.1 square metres of cooling surface area. A DGX H100 system would need around 16 square metres of heat sinks and 33 square metres of solar panels.
As computing power increases, so too do the size of the installation, its mass and the cost of launching it. A megawatt-scale cluster may require several thousand square metres of photovoltaic panels and heat sinks. Every kilogram of shielding, structural components, control systems and power supply reduces the amount of computing equipment that can be placed in orbit in a single launch.
Solar energy is available almost continuously only in carefully selected orbits. The panels must remain oriented towards the Sun, the heat sinks must not be heated by them, and the antennas must maintain communication with Earth or other satellites. Meeting these requirements calls for sophisticated attitude control systems. Furthermore, the panels and coatings degrade under the influence of radiation, ultraviolet radiation and atomic oxygen.
The same environment affects processors. High-energy particles can alter bit values, disrupt calculations and permanently damage circuits. Radiation-hardened electronics are expensive and are usually significantly slower than commercial AI accelerators. Alternatives include shielding, error correction and redundancy – that is, performing the same operations in parallel across several devices. The system’s resilience increases at the expense of mass and usable performance.
ABI Research estimates that the total cost of an orbital data centre may currently be more than 78 times higher than that of a comparable ground-based facility. An additional challenge is the short technology cycle. AI accelerators become obsolete within a few years, whilst replacing a server in orbit requires a separate space mission.
The economics may be more favourable where data is already being generated in space. Local processing of hyperspectral imagery, radar data and meteorological observations reduces the amount of information transmitted to Earth. It also enables faster detection of fires, damage to infrastructure, ship movements and the risk of collisions. In such applications, it is not only the cost per computing unit that matters, but also bandwidth savings and response times.
However, large-scale constellations come at a systemic cost. In June 2026, there were approximately 16,100 satellites in orbit, whilst observation networks were tracking over 46,000 objects. Adding tens of thousands of large computing platforms would increase the risk of collisions, the amount of space debris, astronomical interference and pressure on the most useful orbits.
Orbital data centres are not currently in competition with the AWS, Google or Microsoft clouds. They represent an attempt to build a new layer of infrastructure: more expensive, more difficult to maintain, but capable of analysing data directly at the point of origin. The most immediate market does not lie in training yet another large language model from above the Earth. It lies in satellite intelligence, autonomous spacecraft and the processing of data that would be too slow or too costly to bring back to the surface. Space does not remove the constraints of data centres. It simply replaces energy and land shortages with mass, radiation, thermodynamics and launch costs.
