Playbooks · Coding tasks
Google put a TPU in orbit and got Gemma 4 to answer
Written November 4, 2026
One prompt went up, one haiku came back down. That is the whole milestone so far — a working loop between the ground and a data-center-class chip in orbit — and it is worth exactly that much, no more.
What changed
On 1 October 2026, Google launched a prototype satellite for Project Suncatcher, described in Google's announcement as a long-term research moonshot to explore powering AI compute in space. After several weeks of commissioning — powering up the host and TPU trays and running early machine learning tests — the team reached what Google calls its "Hello World" milestone: the first-ever AI response from a data center class TPU in orbit. During a ground-station pass, the team uplinked a prompt to the onboard Gemma 4 open model, asking it to generate a haiku about being in space. As the satellite passed by Svalbard, Norway, the downlinked poem confirmed success: "Blue marble below / Data streams in silent grace / Starlight feeds my core."
What the milestone shows
- The full loop works. A prompt was uplinked, a model ran onboard, and an answer was downlinked. That chain — ground station to satellite to ground station, with compute in between — is what the milestone proves.
- Follow-up prompts ran too. After the haiku, the team tested space-themed limericks and telemetry prompts reporting the technical status of the payload systems. The satellite did more than answer once.
- Commissioning came first. The response followed weeks of checklists and early tests, not a launch-day attempt. The order matters: the milestone rests on the satellite working normally before the model was ever asked anything.
What it does not show
- No performance figures. Google's post gives no chip count, no TPU generation name, no power figure, no latency number, and no throughput or benchmark score. None is stated here either.
- No durability result yet. The next step, in Google's own words, is a series of tests to evaluate how TPUs perform under thermal stress and radiation exposure in orbit. Those tests have not reported results. A chip that answers once has not been shown to run reliably over time in space.
- No economics. The post claims no cost, no price, and no commercial timeline. Nothing about orbital data centers being viable, cheap, or near follows from one answered prompt.
- No independent check. The account is Google's own. No outside party has reproduced the result, and the post publishes no telemetry or logs to check it against.
What to do with it
Read it as a feasibility marker, not a product signal. If you build on AI infrastructure, nothing about where to run workloads changes on the strength of this post. The two things worth watching are the ones Google names next: the thermal and radiation test results, and the twin interconnected satellites planned for 2027. Until those report, orbital compute remains a research question with one answered haiku attached to it. Headlines calling this the first AI data center in space go beyond what Google said; Google's claim is narrower, and the narrower claim is the one that holds.
Limits of this reading
- Everything above rests on a single Google blog post about a single prototype satellite. That is the whole public record for this milestone.
- The post describes a research moonshot. It is not a service announcement, and no part of it offers access to orbital compute.
- What the satellite does next — and whether the 2027 twin-satellite test happens on schedule — is not established by anything published so far.
Related playbooks
- Antigravity's remote agents run in Google's cloud
- Google ADK 2.14: A2A graduates, Teams joins in
- Google's AQuA watches for silent agent failures
- Google Jules ends Nov 11: export your work
Sources
Back to: Coding playbooks hub · Also: How MonoRace flies a drone on one camera and an IMU · The intelligence-explosion argument, and its evidence