Gemma 4 wrote a haiku about orbit this week on a Google tensor processing unit, or TPU, flying aboard a satellite, and the poem reached Earth as the spacecraft passed over Svalbard, Norway. Google calls the result the first AI response from a data-center-class TPU in orbit and the "Hello World" milestone for Project Suncatcher, its research program to test whether AI compute can run on solar-powered satellites. The satellite, built with Planet, launched October 1 on a SpaceX Falcon 9, and Google announced the response on October 30.

I cover the entire AI stack because a contract signed at one layer creates dependencies three layers down. Suncatcher sits at the bottom of that stack, in compute and power, and it tests a new answer to the question every hyperscaler faces on the ground: where the power for the next wave of AI capacity will come from.

Google Proved the Full Loop From Chip to Ground Station

Google published the Suncatcher research plan in November 2025, with a goal of flying two prototype satellites with Planet by early 2027. The team flew a single-satellite test first, in October 2026, which gives Google a year of orbital data before the paired mission.

Pre-flight testing gave the team reason to try. In a proton beam, Google's Trillium TPUs showed no hard failures at the highest dose tested, and the high-bandwidth memory, the most sensitive part, showed its first irregularities at close to three times the radiation a shielded chip would absorb over five years (Google Research, 2025).

In orbit, the team spent about three weeks commissioning the satellite, powering up the host computer and the TPU trays and running early machine learning tests before it sent the first prompt. Travis Beals, the Google senior director who leads the project, wrote that the team followed the haiku with limericks and with prompts asking the model to report the payload's technical status. A model that summarizes its own spacecraft's health points to an early use for onboard AI that does not wait on the larger scale questions.

Engineers will learn more from the commissioning logs than from the poem.

Google Can Turn Its Flight Data Into the Next Chip Design

Google has company in orbit. Starcloud, a startup backed by NVIDIA, launched a satellite carrying an H100 graphics processor in November 2025 and ran a Gemma model on it a month later, according to BigGo Finance. Two chip architectures now have flight data, which gives buyers and investors a comparison that did not exist a year ago.

Google holds an advantage in that comparison because it designs the TPU, trains the Gemma model and sells TPU capacity through Google Cloud. Thermal and radiation readings from this satellite go to the teams that design the next TPU and size the next model for a one-kilowatt power budget. A startup flying a supplier's chip depends on that supplier's roadmap for the same changes.

Google owns the chip and the model, so its flight data goes straight to the next TPU design.

Heat Sets How Much Work Each Satellite Can Return

The satellite carries four TPUs with about the computing power of one data center server, runs on about one kilowatt of solar power, and has a design life of about a year (Yahoo Finance, 2026). On the ground, air and chilled water carry heat away from chips. In a vacuum, heat leaves through radiator panels, so Google built a path of thermal interface material, aluminum and copper to carry it there.

With the current design, the chips run for about 15 minutes before they power down so the radiators can release the heat (NPR, 2026). Beals told NPR the radiators rank among the heaviest parts of the satellite and that his team is working on more efficient cooling. Radiator mass ties the thermal problem to the cost problem, because Google pays to launch every kilogram on the satellite.

Google says its next tests will measure TPU performance under thermal stress and radiation exposure, and that reliable workloads in orbit clear the way for twin interconnected satellites in 2027. That mission adds the network layer to the test, because a cluster in orbit needs satellites that exchange data with each other at speeds close to links inside a data center.

Google's Cost Case Depends on Launch Prices Below $200 a Kilogram

Google's argument for orbit starts with power. In the right orbit, a solar panel collects up to eight times more energy per year than the same panel on the ground, with near-continuous sunlight (Google Research, 2025). That figure measures energy collected, and the cost of an AI task in orbit also includes hardware, launch, cooling mass, communications and replacement.

Google's 2025 paper projected that launch prices could fall below $200 per kilogram by the mid-2030s. At that price, the paper estimated the yearly cost of power capacity in orbit at $810 to $7,500 per kilowatt, a range that overlaps the $570 to $3,000 per kilowatt that U.S. data centers spend on electricity each year (Google Research, 2025). Launch providers set the price side of that comparison, and this satellite produces Google's first orbital data on the thermal and chip side, which Google controls.

Your AI Capacity Plan Faces the Same Power Constraint Sooner

Google describes Suncatcher as a long-term moonshot, and your planning horizon is shorter. The useful signal for you is Google's reason for funding it: a hyperscaler researching power that sits outside the grid interconnection queue expects power on the ground to stay tight for the next decade.

If your AI roadmap assumes GPU or TPU capacity in a specific cloud region, the power available in that region affects when you get the capacity and what you pay for it. You have the most bargaining power to get those answers while you are still negotiating.

Two Results Will Show How Fast Suncatcher Can Scale

• Published thermal and radiation results from this year in orbit, including any gain in how long the chips can run between cooling pauses.
• The 2027 twin-satellite mission, which will show whether two satellites can exchange data fast enough to work as one cluster.