Space Data Centers: AI Compute Goes Orbital

As Elon Musk addressed the World Economic Forum in Davos this January, he laid out a bold timeline for artificial intelligence infrastructure. The lowest-cost place to run AI compute would soon shift from Earth to orbit. Space data centers, he argued, could harness constant solar power and the vacuum of space for cooling.

Here is the kicker: SpaceX has already filed FCC plans for a million-satellite network to support this vision, with xAI integration in the works. Yet AWS CEO Matt Garman told industry leaders in early February that orbital data centers remain "pretty far" from economic reality.

The Orbital Vision

Proponents envision compact satellite constellations in sun-synchronous low-Earth orbits. These clusters would draw near-continuous sunlight with minimal batteries. Early tests include Starcloud's 2025 GPU launch and Google's planned 2027 prototypes under Project Suncatcher.

Google's research outlines modular designs using free-space optical links for tens of terabits per second between satellites. Radiation-hardened TPUs and precise orbital formations aim to match terrestrial performance for distributed machine learning.

Technical Realities

Challenges abound in the vacuum. Heat dissipation requires massive radiators since traditional air or liquid cooling fails. Radiation demands shielding or hardened chips, while inter-satellite coordination must overcome orbital dynamics and potential bit-flip errors.

Launch economics dominate the math. Current costs hover near $3,600 per kilogram. Viability likely requires drops to around $200 per kilogram, expected with next-generation vehicles in the 2030s. A 1 GW orbital facility could cost nearly three times its ground equivalent today.

Impact on Legacy Providers

Why this matters for the finance market: AWS, Microsoft, and Google continue committing hundreds of billions to terrestrial expansions through 2030. Garman highlighted insufficient rocket capacity even for Musk-scale ambitions.

Legacy providers view space as a distant complement at best, not a near-term replacement. Short-term, ground-based hyperscale growth drives cloud revenues while space remains niche for latency-tolerant inference workloads.

The Road Ahead

What changed next could hinge on Starship success and prototype results from 2027 onward. Small-scale demonstrations may arrive by decade's end, but full hyperscale orbital systems likely wait until the 2030s or beyond.

Investors should track launch economics and regulatory filings closely. For now, the cloud market stays firmly Earth-bound, yet the orbital race signals long-term pressure on traditional infrastructure models.

Key Comparisons: Space vs Terrestrial Data Centers

• Power Cost per kW/Year: Orbital ~$14,700 (current) vs Ground $570–$3,000
• Energy Source: Constant solar in optimal orbits vs Grid-dependent with variability
• Cooling: Radiative vacuum vs Water/air systems
• Timeline to Scale: Prototypes 2027, potential viability 2030s vs Immediate hyperscale builds