4 Minutes
Imagine training an enormous artificial intelligence not in a chilled hangar on Earth but under constant sunlight, floating hundreds of kilometres above the atmosphere. That is the image Elon Musk is selling as SpaceX and xAI merge in a move he says will create the world’s most valuable private company—about $1.25 trillion by some tallies—and shift the centre of high-performance computing off-planet.
The pitch: unlimited solar power and vast real estate
Musk has argued that terrestrial data centres are hitting a hard limit. They guzzle electricity and need complex cooling systems; they strain local grids and can worsen environmental impacts if scaled aggressively. His proposal is stark: the only practical long-term path to run exponentially larger AI models is to take the compute into orbit, where solar energy is abundant and thermal management can be engineered differently.
The plan leans on SpaceX’s successes—rockets that can be reused, the Starlink constellation’s communications backbone, and launch economies that keep improving. In Musk’s timeline, within two to three years orbital computing could be cost-competitive for AI workloads. The architecture he envisions is effectively a ring or constellation of specialised compute satellites: orbital datacentres that tap near-constant sunlight, beam data via laser or radio links, and form a global, low-latency fabric for training and inference.
How it would work, in practical terms
Solar irradiance in space is roughly 30–40% stronger than at Earth's surface, and there’s no night-side atmospheric absorption. That energy advantage lets arrays generate more power per square metre. But power is only part of the puzzle. Heat rejection—getting rid of waste heat from racks of processors—would be handled by radiators that radiate thermal energy into cold space, avoiding the need for massive chillers and water loops common on the ground.
SpaceX has reportedly asked regulators about launching very large numbers of satellites. A previous filing hinted at ambitions on the order of hundreds of thousands to millions of small satellites. If that is scaled to datacentre-class platforms, companies could train models at unusual speed and scale because the constraint becomes launch cadence and in-orbit maintenance rather than the terrestrial grid.
Economic logic and institutional friction
SpaceX is profitable. xAI, and the social platform X associated with its founder, have far higher cash burn and face regulatory scrutiny, especially in Europe. Folding xAI and SpaceX closer together would stabilise financing and concentrate R&D: rockets, global connectivity, mobile-to-orbit links and AI models under one roof. The revenues from orbital datacentres, Musk suggests, would also help bankroll grander goals—sustained lunar bases and a self-sufficient presence on Mars.
Still, this is not just engineering optimism. Launch costs must fall further for orbital datacentres to make sense economically. Orbital servicing, debris mitigation, cybersecurity of space-based compute, and international regulatory frameworks all present serious non-technical hurdles. Latency-sensitive applications might still prefer terrestrial nodes. But for GPU-heavy model training, where throughput trumps a few extra milliseconds, orbit becomes compelling.
This is a bet on radically different infrastructure for AI: power from the Sun, cooling to deep space, and compute freed from earthly limits.
Whether regulators, investors and engineers align behind that bet will determine if we see the first wave of orbital datacentres in the coming decade—or if the idea remains an audacious footnote in the story of AI.
Leave a Comment