A million “orbital data center” satellites is not simply an AI scaling plan: this is an orbital design problem, too. SpaceX has petitioned the Federal Communications Commission to license a constellation of an up to million solar-powered satellites which it termed as artificial intelligence space-based data centers. In the filing, the company places the concept as compute infrastructure that can ride near-continuous sunlight and avoid terrestrial constraints as “a first step towards becoming a Kardashev II-level civilization” as well as carrying AI workloads on its behalf on “billions of people.”
It is a geometry, rather than rhetoric, that makes the greatest impression. SpaceX defines the narrow shells at a depth of up to 50 km, distributed on 500 km to 2000 km altitude and inclined at 30° to 30° sun-synchronous orbits. The same low-Earth-orbit bands are overlapped by that range, with which the communications systems already vie over spectrum, collision-avoidance margins, and predictable operation. The FCC has also demonstrated its readiness to approve only limited segments of large constellation plans, and the result of that is a regulatory fact in which “one million” can become an opening bid instead of a final condition.
The sky is already crowded even in the absence of the AI payload. According to the European Space Agency, there is an estimated tracking environment of approximately 40,000 objects of which there are approximately 11,000 active payloads in addition to a larger number of debris that are too small to track in full detail. At the altitude neighborhood of 550km, ESA modeling shows the threat density of debris is now comparable to active satellites- that is, routine operations are now not only including aviation of avoidance, coordination, and end-of-life disposal discipline.
It is not the only spacecraft safety problem because of congestion. Astronomy now has an operational limitation of optical and radio interference (particularly in wide-field surveys, which view at twilight and at the beginning and end of night). One series of simulations analyzing future low-Earth-orbit space telescopes discovered that when proposed constellations are completely deployed (approximately 560,000 satellites), the fraction of exposures containing at least one satellite trail increases to over 96 percent in a variety of proposed missions, and the fraction of images affected by Hubble increases to approximately 40 percent at a one-million-satellite population. These figures are important since AI satellites that represent “data center” would probably have large surfaces of power-generation, and the brightness of a satellite is highly dependent on orientation, area of the panel and the geometry between Sun, satellite and viewer.
This is more of an engineering irony. The orbits which render solar-powered computing appealing, such as sun-synchronous ones which maintain arrays visible, are also the orbits which concentrate when it is time to see the sun rise and fall, the very time that some of the scientific programs are incapable of just taking a break and peeking the observations. In the meantime, satellites can be pushed downwards to decrease the duration of sunlit visibility on the surface, but these decreases are at the cost of increasing drag and causing more frequent reentry events, which alters the sustainability cycle to a debate on collision risk to one of atmospheric deposition and reentry rate.
Space-based AI compute is not exclusive to a single company. Project Suncatcher of Google has outlined closely clustered structures with optical inter-satellite connections and initial test objectives and other ideas have propagated internationally. The filing by SpaceX is a wakeup call, though, that on a megaconstellation scale, “data center” is a claim to bundle: compute as well as power and thermal rejection and downlink capacity and maneuverability and credible end-of-life behavior must all fit within a set of shared orbital lanes that everyone must share. With filings piling up, the feasible question is no longer whether space can power AI, but whether space can remain usable when doing so.