Google’s First Space Data Center Satellite Runs Gemma 4

Introduction

Google’s vision of space-based data centers has moved beyond a concept presentation. The first Suncatcher prototype has completed its initial commissioning, established communication with the ground, and begun running Google DeepMind’s Gemma 4 model in orbit. For the first prompt, engineers asked the satellite to write a haiku about space. It produced a conventional verse about Earth, data streams, and starlight—but the poem is less important than the hardware test behind it.

The spacecraft carries four Google Tensor Processing Units, or TPUs. It can generate up to 1 kilowatt of electricity, yet the chips consume enough power that only one TPU can be switched on at a time, for around 20 minutes. That constraint affects both the amount of work the satellite can perform and the scheduling of its space-to-ground communications.

Key points

• The poetry is a simple proof of operation. Google also asked the spacecraft to report orbital conditions, and the company says the satellite is operating nominally.
• Radiation is the central unknown. High-energy particles can cause memory “bit flips,” corrupting stored values without necessarily triggering ordinary monitoring systems.
• Inference may be more practical than training. Ground testing suggested radiation would have a smaller effect on inference workloads than on model training, but the orbiting experiment will provide a more realistic assessment.
• The next phase is cooperative computing. If the current hardware performs well, Google plans to launch two more advanced-chip satellites in 2027 and demonstrate optical links between spacecraft.

Why it matters

Suncatcher is designed around a network of closely flying satellites that could share large inference workloads. This distinguishes Google’s approach from concepts focused mainly on running inference independently on a single spacecraft. High-speed optical links could allow the satellites to behave more like a distributed system, but they also introduce difficult requirements for alignment, communications, power management, cooling, and fault tolerance.

The economics are just as challenging. Space hardware must operate with limited energy and face thermal stress and radiation, while launch costs remain a major barrier. A peer-reviewed analysis cited in the report estimated that SpaceX would need to fly Starship 1,800 times to achieve learning-driven cost reductions comparable to those seen with earlier launch vehicles. Both Google’s and SpaceX’s space-data-center ambitions therefore depend partly on heavy-lift rockets becoming rapidly and repeatedly reusable.

The Suncatcher mission should not be read as a commercial data-center launch. It is an infrastructure experiment aimed at the 2030s. Its immediate question is whether a TPU can deliver reliable inference in orbit despite intermittent power and radiation. The larger question is whether a cluster of satellites can provide enough computational value to justify the complexity of putting AI infrastructure above Earth.

Source: TechCrunch AI