Google just proved you can run inference in orbit — now the question is whether you should.

The Summary

The Signal

Google launched a satellite equipped with Tensor Processing Units and successfully ran its Gemma 4 language model from orbit. The first prompt: write a haiku about being in space. The model delivered. "Blue marble below / Data streams in silent grace / Starlight feeds my core." Google also tested space-themed limericks and telemetry prompts, confirming the hardware works under orbital conditions.

The satellite is part of Project Suncatcher, Google's experimental initiative exploring whether AI data centers belong in space. TechCrunch confirms this is the first TPU operating in orbit, running a Google DeepMind model in an environment where cooling, power, and latency constraints look nothing like a terrestrial rack.

"The first TPU in space marks a shift from whether compute can survive orbit to whether it should live there."

Why put inference in space? Three reasons:

  • Power: Solar arrays work 24/7 above the atmosphere, no weather, no nighttime gaps
  • Cooling: Radiative cooling in the vacuum of space eliminates the need for water or air-based systems
  • Geography: Satellites bypass fiber routes, regulatory borders, and real estate costs

The haiku is cute. The real test was whether the TPUs could boot, stay stable, and handle model weights in a radiation-heavy, zero-gravity environment. They did. That means Google now has hard data on power draw, thermal performance, and error rates for inference hardware operating outside Earth's protective magnetosphere. Radiation-hardened chips are standard for satellites, but running a modern AI accelerator at scale in orbit is new territory.

Latency is the obvious problem. Geostationary orbit adds 500ms round-trip. Low Earth orbit cuts that to 20-40ms, but you need constellation coverage and handoffs between satellites. For real-time applications like voice or interactive agents, that's a dealbreaker. But for batch inference, model training on vast datasets, or serving models to remote regions with no ground infrastructure, the latency trade-off starts to pencil.

The Implication

This is a capability demo, not a product launch. Google isn't moving GPT-scale training runs to orbit tomorrow. But if Project Suncatcher proves that space-based TPUs can run inference reliably and cheaply, it opens a new axis for hyperscalers: vertical expansion. Build up, not out.

Watch whether Google shares performance metrics, cost-per-inference comparisons, or power efficiency data in the next six months. If they go quiet, it's a science project. If they start talking about constellation deployment timelines, someone did the math and it worked.

Sources

The Verge AI | TechCrunch AI