Orbital Data Centers Take Center Stage

SpaceX wants to park up to one million data centers in orbit. The audacious plan is pitched as AI’s answer to Earth’s climate and energy limits—move the compute, and the grid stress, up into space.

The MIT Technology Review explainer outlines a bold constellation of ambitions: SpaceX has filed an FCC application to launch a vast fleet of orbital data hubs, a quantum leap beyond today’s Earthbound centers. The broader ecosystem isn’t waiting: Amazon’s founder has floated a future where large-scale computing goes orbital, Google is eyeing a test constellation of 80 satellites as soon as next year, and a Washington-based startup, Starcloud, already tested an orbital Nvidia H100 GPU payload. The goal, by some forecasts, is for orbit-based data centers to grow to Earth-sized scale by 2030. Proponents argue the move would unlock AI without loading Earth’s energy grids or water supplies with cooling demands.

The core appeal is clear: space-based compute could separate AI’s power draw from terrestrial resource constraints. On Earth, cooling large data centers consumes substantial water and electricity—pressing concerns for communities near mega-facilities and for utilities stretched by the AI boom. In space, supporters say, cooling can be decoupled from rivers and lakes, and sunlight could power orbiting hardware rather than burning through ground-based grids. The piece points to a potential win-win: faster AI, lower local environmental impact, and a more elastic global compute fabric.

But the road from proposal to production is steep and studded with questions. Feasibility depends on hardware resilience in radiation-rich, vacuum environments, lifetime of components in orbit, and the ability to service or upgrade fleets of satellites at scale. Launch costs, maintenance cadence, and the economics of replacing or upgrading thousands—or millions—of micro-datacenters would dominate the business case. Then there’s the thorny issue of orbital debris and collision risk, regulatory complexity, and the need for reliable, low-latency links to ground networks. The promise is dramatic; the hurdles are equally stark.

For practitioners watching ship-and-build cycles, a few concrete takeaways stand out. First, cost structure will be decisive. Even with cheaper launchers and modular hardware, the economics of operating a million micro-datacenters in orbit hinge on repeatable, low-cost replacement and upgrade cycles, power budgeting, and thermal management in a vacuum. Second, reliability and maintenance will constrain feasibility. In-orbit hardware must tolerate radiation, micro-meteoroids, and long service intervals—making modularity and fault-tolerance non-negotiable design choices. Third, the governance of orbital assets will shape timelines. Debris mitigation, spectrum access, and debris-removal pathways will define how quickly fleets can scale and how aggressively operators can push upgrades. Fourth, practical performance will depend on the ground network’s ability to handle latency, synchronization, and data ingress/egress from space-hubs—a nontrivial systems engineering challenge.

Analysts liken the concept to moving the factory floor from a climate-controlled warehouse to a moving, sunlit orbiter—an analogy that clicks once you realize the entire premise rests on reimagining power, cooling, and maintenance as orbital problems, not terrestrial ones. If a fraction of the vision comes true, we could see a dramatic shift in how and where compute is provisioned, with downstream effects on datacenter supply chains, energy markets, and AI deployment patterns in the coming decade.

What to watch next: the outcomes of Starcloud’s tests with high-performance GPUs in orbit, the pace of Google’s constellation experiments, and any concrete economic models SpaceX or peers publish about scaling to “Earth-sized” orbital capacity. Until demonstrations materialize, this remains a provocative blueprint rather than a proven blueprint for the AI era.

Sources

Four things we’d need to put data centers in space