> ## Content Index
> Fetch the complete content index at: https://venture-block.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# Google's Suncatcher satellite runs its first AI prompt on a TPU in orbit
- URL: https://venture-block.com/google-suncatcher-satellite-first-ai-response-tpu-orbit/
- Published: 2026-10-31T14:08:17.000Z
- Updated: 2026-10-31T14:08:17.000Z
- Description: Google's Project Suncatcher prototype ran a Gemma 4 prompt on a TPU in orbit and downlinked the answer. Radiation and thermal tests come next, ahead of two linked satellites in 2027.
- Author: Venture Block
- Tags: Science, AI

Google says the prototype satellite for Project Suncatcher, its research effort to run AI computing in orbit, has [returned its first AI response from space](https://blog.google/innovation-and-ai/models-and-research/google-research/project-suncatcher-satellite-response/?ref=venture-block.com), generated on one of the company's own Tensor Processing Units (TPUs). Engineers uplinked a prompt to a copy of Google DeepMind's open Gemma 4 model stored on board, and the answer came back during a ground-station pass over Svalbard, Norway.

The prompt asked for a haiku about being in space. Google published the result: "Blue marble below / Data streams in silent grace / Starlight feeds my core." The company calls it the first AI response from a data-center-class TPU in orbit, and says it followed the haiku with space-themed limericks and telemetry prompts that reported on the payload's own technical status.

The poem is lighthearted, but the engineering milestone behind it is concrete. It shows that a commercial AI accelerator can be powered up, cooled and fed a workload on a small satellite, with the output making it back to the ground through the short, intermittent communication windows that low-Earth-orbit spacecraft get.

## What is on the satellite

The spacecraft was built with Planet and [launched on October 1](https://blog.google/innovation-and-ai/models-and-research/google-research/project-suncatcher-prototype/?ref=venture-block.com) on SpaceX's Transporter-18 rideshare, a Falcon 9 mission that carried more than 100 payloads. It is based on one of Planet's standard satellite platforms rather than a purpose-built compute design, [TechCrunch reported at launch](https://techcrunch.com/2026/10/01/google-thinks-spacexs-starship-has-to-launch-1600-times-before-space-data-centers-get-off-the-ground/?ref=venture-block.com).

It carries four TPUs and generates up to 1 kW of electricity. That is not enough to run them together: the satellite can power only one TPU at a time, for roughly 20 minutes per session, [according to TechCrunch](https://techcrunch.com/2026/10/30/googles-first-data-center-satellite-is-online-and-writing-poetry/?ref=venture-block.com), so power and heat budgets shape every test as much as the chips do. Google says the past few weeks went to a commissioning checklist that included powering up the host computer and TPU trays and running early machine learning tests.

Gemma 4 is a sensible first payload for that setup. As an open model that can be stored and run locally, it lets Google test inference entirely on board, without relying on a ground link for anything except the prompt and the answer.

## Radiation is the real test

The next phase, Google says, is a series of tests on how the TPUs hold up under thermal stress and radiation exposure. The concern is bit flips, where a charged particle changes a stored value in memory and can corrupt data without tripping normal error monitoring.

Google has already tested the chips in particle accelerators. At launch, TechCrunch reported that the team had to redo those beam tests after realizing the original setup shielded the chips more than space would, and that the repeat produced slightly more logic errors. Travis Beals, the Google executive running the project, told TechCrunch the error rate is "very low" for typical inference, around one in a million, but already a problem for a mega-scale training run across thousands of chips for months. The company still expects its chips to handle large inference workloads over a satellite's five-year life. Orbit is where that ground data gets checked against reality.

## From one satellite to a cluster

If the prototype holds up, Google plans to launch [two interconnected satellites in 2027](https://www.theverge.com/tech/1023528/google-project-suncatcher-ai-satellite-transmission?ref=venture-block.com). TechCrunch reports they will generate more power and demonstrate optical links between spacecraft. The long-term design described at launch is a cluster of 81 satellites flying in close formation and working on the same job in parallel, which is why Google cares so much about bandwidth and latency between chips.

That focus on tightly linked, distributed workloads sets Suncatcher apart from most other orbital computing efforts. TechCrunch notes that startups such as Starcloud and Cowboy Space plan to run inference on individual satellites, and that SpaceX's planned orbital data centers are expected to use laser links but fly too far apart to share processing effectively. Satlyt, which also flew on the October 1 rideshare, reported this week that it had [tested task-planning software in orbit](https://techcrunch.com/2026/10/30/satlyt-takes-another-step-toward-turning-satellites-into-space-data-centers/?ref=venture-block.com), using one physical computer on a spacecraft to plan work for a simulated four-node network.

## The launch math

Google is open about the fact that the economics do not work yet. Its peer-reviewed paper in the journal Joule, released alongside the launch, argues that launch prices near $200 per kilogram could be reachable around 2035 if SpaceX keeps a cost-reduction learning curve of about 20 percent. TechCrunch's reading of the paper is that getting there would mean Starship carrying about 370,000 tons to orbit, roughly 1,800 launches over 10 years at 200 metric tons each. The researchers stress that the paper is not an economic feasibility study, and Starship has never flown more than five times in a single year.

That keeps Suncatcher in the moonshot category, aimed at infrastructure for the 2030s while chipmakers on the ground, like [FuriosaAI with its inference chips](https://venture-block.com/furiosaai-620m-series-d-rngd-inference-chips/), race to meet AI demand today. It is also the second Google research bet this week to show up on our pages, after DeepMind's [Genie 4 world model](https://venture-block.com/google-deepmind-genie-4-world-model/). The next data point from orbit will be the radiation and thermal results, ahead of the twin-satellite demonstration planned for 2027.