SpaceX wants to put AI data centers in orbit with STARMIND

Terrestrial data centers are running into hard limits. Between grid bottlenecks, water consumption for cooling, and local community pushback, scaling AI compute on Earth is becoming an administrative nightmare.

SpaceX wants to bypass all of that by moving compute clusters into orbit.

The company has officially introduced STARMIND, a space-based AI satellite architecture designed to tap continuous solar energy and off-planet cooling.

A 75-meter data center in the sky

The first iteration, dubbed AI1, is massive. Once deployed, each satellite spans 75 meters in width and reaches 30 meters in height.

It operates as a self-contained orbital compute module capable of delivering up to 250 kW peak compute power, with an average draw of 175 kW.

Placed in a sun-synchronous orbit, the satellites soak up uninterrupted sunlight 24/7 without weather degradation or atmospheric filtering.

To lift these mega-structures into orbit, SpaceX relies directly on the heavy payload capacity and full reusability of Starship.

Radiating heat into vacuum

Cooling is one of ground data centers' biggest operating costs. In orbit, heat radiates naturally into the deep vacuum of space.

SpaceX says this eliminates energy-hungry chillers, cooling towers, and fans, cutting cooling energy overhead by an order of magnitude.

On the hardware front, SpaceX describes the architecture as chip-vendor agnostic. Reports note integrations with Nvidia's Rubin GPUs and Vera CPUs, while SpaceX also plans a joint fab facility with Tesla to build proprietary chips at scale.

The satellites interconnect via high-speed laser links, and inference results beam straight back to Earth over the existing Starlink constellation.

To build them at scale, SpaceX is already constructing the Gigasat Factory in Bastrop, Texas, aiming to manufacture and deploy thousands of units starting as soon as late 2027.

Why it matters

If this works, it rewrites the economics of high-density AI infrastructure.

Deploying data centers in space completely circumvents local zoning approvals, terrestrial power grid queues, water rights disputes, and community protests. It turns a political and bureaucratic problem into a purely capital and engineering equation.

It also closes the loop on Elon Musk's industrial ecosystem: Starship supplies mass to orbit, Starlink provides the low-latency communications backbone, and Tesla collaborates on compute and manufacturing.

Still, we think orbital compute faces brutal physics and operational realities.

Cosmic radiation causes single-event upsets and rapid hardware degradation. Unlike a terrestrial server rack, an orbital node cannot be serviced when a chip fails; it has to be written off or deorbited entirely.

Furthermore, even with advanced optical cross-links, inter-satellite latency and bandwidth cannot easily match the tightly coupled InfiniBand fabrics of terrestrial clusters.

A late 2027 rollout sounds characteristically optimistic, but STARMIND makes one thing clear: frontier AI is officially looking beyond the limits of planet Earth.