Locating data centres in space is being sold as a solution to the rising electricity and water consumption by AI. SpaceX has requested regulatory officials to allow up to one million satellites in the field that are not necessarily intended to provide internet, but rather to compute, making low-Earth orbit effectively a huge, solar-powered machine room.
The proposal that was presented to the Federal Communications Commission describes “orbital data centres” as a method to provide computing power to “billions of users worldwide” and not be limited to land, grid and cooling of traditional server farms. The satellites would be in the low-Earth orbit just like in Starlink 5002,000km. In arg. X of the same, Elon Musk said that their distance would make them hard to discern of each other, and added: “Space is so vast as to be beyond comprehension.”
Hanging over the ambition is a terrestrial bottleneck which is now difficult to overlook. Gigaworthy data centres already place a significant load on local infrastructure, and the most rapidly advancing AI workloads only exacerbate the problem by putting more heat into each rack. Certain facilities may require up to 5 million gallons of water per day to cool them and the national numbers are in billions of gallons per year. One article mentioned 449 million gallons of water per day of U.S. centres (as of 2021), explaining the reason why water has become an aspect of the AI discussion, instead of a side note.
Space-based computing in principle avoids some of that. Photovoltaic arrays assembled above the weather are able to gather nearly uninterrupted sunshine in the right orbit and heat can be dissipated using radiators instead of evaporating water. The filing by SpaceX refers to spacecrafts in sun-synchronous orbits as well as lower inclinations, with the intention to align the always-on capacity to the highs and lows of demand. The company also outlines a network that is constructed based on inter-satellite optical connections, in which Starlink spacecrafts are used to relay traffic to the ground when necessary, and the Ka-band communications serve as the backup channel.
But orbit is substituting the usual with the unfamiliar engineering. Exposes of radiating forces compel designs to depend on shielding, strong components, or software-based error detection; heating large radiators needs to be lofted along with the processors they cool, since big radiators need to be lofted also. On systems where upgrades must be launched instead of a forklift swap, component life cycles analysis is in a different form: the experts mentioned in broader industry deliberation have explicitly indicated that sophisticated chips are typically substituted every five to six years, a pace that complicates the logistics of supporting large constellations.
The issue of whether or not greener in orbit will still be “greener” when the rockets are taken into account is also present. Studies highlighted in the orbital data centres debate have indicated emission of launches and hardware burning down in the atmosphere during re-entry, which also included the issue of atmospheric chemistry and the ozone effects. The allure is obvious – use the Sun itself, no longer construct power plants on Earth, but the entire footprint is determined by the frequency of changing payloads and the number of flights.
Then there is traffic. The altitudes put forward are not empty space, space is overrun by active spacecraft and debris. According to an estimate given by debris experts in 2025, the debris in orbit is estimated to be in the form of almost 130 million fragments, including whole rocket bodies and paint particles. As the number of objects increases so do the number of close approaches and so does the amount of operational load, which is the tracking, coordination and collision avoidance, between and among competing fleets.
Research in the industry on the dynamics of megaconstellation has characterized the frequency with which satellites fly within a kilometre of each other, and how solar storms can enhance the situation by bloating the upper atmosphere, reducing drag and affecting navigation. Resiliency is not an idealistic concept in that context: it is an engineering imperative of any system that anticipates accommodating valuable compute hardware at scale.
Despite SpaceX requesting permission to launch an unprecedented constellation, smaller and more practical experiments are already underway to see what “computing in space” can be. The International Space Station has a prototype orbital data centre supported by the ISS National Laboratory, which is concentrated on near-real time processing when the downlink bandwidth is constrained. Its reasoning is less broad than that of SpaceX process information near the source of the work, less reliance on Earth but it highlights one major idea: the jump to orbit is a systems engineering task and operations issue as much as it is a concern of sheer computing power.
