# Google ran Gemma 4 on a TPU in orbit. Here's what Project Suncatcher actually proved

> Google's Suncatcher prototype satellite returned its first AI response from a data-center-class TPU in orbit. It's a real hardware milestone, but it says very little yet about whether space data centers can compete on cost.

Published Nov 2, 2026 by TechMeasured Staff · Science · Canonical: https://techmeasured.com/science/google-suncatcher-gemma-4-tpu-in-orbit-what-was-proven

Google says the prototype satellite for **Project Suncatcher**, its research effort to run AI compute on solar power in space, has produced its first AI response from a **data-center-class TPU in orbit**. In an October 30 post, Travis Beals, a senior director at Google, wrote that the team uplinked a prompt asking the onboard **Gemma 4** open model to write a haiku about being in space. The answer came back as the satellite passed over Svalbard, Norway. It opens with "Blue marble below."

This matters because it's the first time Google has shown the chips it uses in its data centers working in orbit, not just in a lab. But a haiku is a smoke test. The questions that decide whether orbital data centers make sense (radiation, heat, links between satellites and launch costs) are still open, and Google says so.

## What was actually demonstrated

| Item | What's been reported |
|---|---|
| Launch | October 1, 2026 (Google), on a SpaceX Falcon 9 alongside more than 100 other spacecraft (TechCrunch) |
| Builder | Planet, the Earth-imaging company (TechCrunch, Hardware Busters) |
| Compute | Four Google TPUs (TechCrunch); Hardware Busters identifies them as Trillium-generation |
| Power | Up to about 1 kW (TechCrunch) |
| Operating limits | Only one TPU runs at a time, for roughly 20 minutes each (TechCrunch) |
| Commissioning | About three weeks of subsystem testing before operations (TechCrunch) |
| Workload | Gemma 4 inference: a haiku, space-themed limericks, and prompts that report payload status (Google) |

Google hasn't said which Gemma 4 size runs on the satellite, and it hasn't published performance, throughput or power-efficiency figures from orbit. A Google spokesperson told TechCrunch the satellite is operating nominally.

So the proven part is narrow but real. A commercial-style satellite powered up data-center TPUs, ran an open model, and returned useful text through a normal ground-station pass. The limits matter too: one chip at a time, for about 20 minutes, on about a kilowatt. A terrestrial AI data center draws many megawatts.

## What hasn't been tested yet

**Radiation.** TechCrunch reports that Google tested the TPUs in particle accelerators on Earth. Those ground tests suggested little impact on inference but more risk for training, because radiation can flip bits and quietly corrupt data. According to Hardware Busters, Google's 2025 research found that Trillium TPUs coped better than expected in a 67 MeV proton beam. InfoQ, summarising the same work, reported that the chips should withstand the dose of a five-year low-Earth-orbit mission. Google's post says the next phase will test how the TPUs behave under radiation and thermal stress in actual orbit.

**Networking.** A data center is many chips talking to each other quickly. This satellite is a single node. Google says its next milestone is **twin interconnected satellites in 2027**. TechCrunch adds that those satellites will carry more advanced chips, generate more power and test optical links between spacecraft. In 2025, InfoQ reported that Google had reached up to **1.6 Tbps** with a single transceiver pair in early lab experiments.

**Economics.** This is the biggest open question, and nothing on this flight addresses it. When Google announced Suncatcher in 2025, InfoQ reported that the company's analysis depended on launch costs falling **below $200 per kilogram by the mid-2030s**. At that price, Google said, orbital compute could roughly match the energy cost of data centers on the ground. The same coverage cited sun-synchronous orbits collecting solar power up to **eight times** more effectively than ground arrays, and a design of up to **81 satellites** at about 650 km altitude. Hardware Busters notes that whether launch prices will ever fall that far is still unresolved.

## Plan versus reality

In 2025, InfoQ reported Google's target as two prototype satellites by early 2027. What has flown so far is a single satellite, with the interconnected pair scheduled for 2027. That timeline is roughly on track, but the step that tests the "data center" part of the idea is still ahead.

## What to watch

- **Radiation results from orbit.** Watch especially for whether Google says anything about training, not just inference.
- **The 2027 twin-satellite mission** and any optical-link bandwidth measured in space.
- **Real numbers** on throughput per watt and on uptime per orbit. Without those, cost comparisons with ground data centers are guesswork.

For now, Suncatcher shows that the hardware can survive the trip and switch on. It doesn't yet show that space is a cheaper or better place to run AI.


## Sources

1. [Google — Project Suncatcher: first AI response from an orbital TPU](https://blog.google/innovation-and-ai/models-and-research/google-research/project-suncatcher-satellite-response/)
2. [TechCrunch — Report on Google's Project Suncatcher prototype running Gemma 4 in orbit](https://techcrunch.com/?p=3183584)
3. [Hardware Busters — Project Suncatcher's orbiting TPU sends back its first AI answer, and it's a haiku](https://hwbusters.com/news/project-suncatchers-orbiting-tpu-sends-back-its-first-ai-answer-and-its-a-haiku/)
4. [InfoQ — Google Project Suncatcher and space-based AI compute (Nov 2025)](https://www.infoq.com/news/2025/11/google-suncatcher-space)