---
title: "Google is putting four ordinary AI chips into orbit on October 1"
description: "Project Suncatcher’s first flight will test how its TPUs survive launch forces, radiation and, above all, their own heat"
author: "Walter Pine"
published: 2026-09-24T22:12:48Z
modified: 2026-09-26T11:29:54Z
url: https://rews.cc/a/google-is-putting-four-ordinary-ai-chips-into-orbit-on-octob-fa2494
language: en
tags: ["ai", "space", "google", "chips", "hardware", "tech"]
publisher: "Rews (https://rews.cc)"
---

# Google is putting four ordinary AI chips into orbit on October 1

*Project Suncatcher’s first flight will test how its TPUs survive launch forces, radiation and, above all, their own heat*

By Walter Pine · September 24, 2026 · https://rews.cc/a/google-is-putting-four-ordinary-ai-chips-into-orbit-on-octob-fa2494

## In brief

- A prototype Suncatcher satellite built with Planet flies on SpaceX’s Transporter-18 Falcon 9 rideshare on October 1
- Ground tests included vibration on all three axes and a UC Davis proton-beam run exceeding a five-year radiation dose
- Cooling limits TPU work to about fifteen minutes at a time; Google expects roughly 1.3 square meters of radiator per chip
- A two-satellite laser-link test follows in 2027; Google has described it as a joint mission with Planet
- Critics question costs and viability: Google targets $200/kg launches, and a Gartner analyst called the idea “peak insanity”

One point three square meters. That, according to figures reported by The Register, is the radiator area Google reckons it needs to carry heat away from a single tensor processing unit, and the company hopes eventually to pack dozens of TPUs into each satellite. That arithmetic, as much as the launch itself, is the load-bearing number in the enterprise of putting AI data centers into orbit — an enterprise that gets its first flight test aboard a SpaceX rocket on October 1.

The prototype satellite will fly on the Transporter-18 rideshare mission aboard a Falcon 9, carrying four TPUs, and it was built with Planet, the Earth-imaging company. The chips are not space-rated hardware of the kind other missions rely on; they are the same TPUs Google integrates into its servers on the ground — the existing Trillium accelerators, lightly prepared. The mission exists to measure how they cope with the stress of launch, the radiation of space, and its extremes of heat and cold.

The program is Project Suncatcher, announced last November as a long-term research moonshot into whether space could one day host scalable machine-learning infrastructure. In low Earth orbit, satellites reach near-constant sunlight and can generate up to eight times more solar power than on Earth. Further out, Google imagines constellations of satellites linked together, managing larger AI workloads in orbit — the same patient path, the company notes, that autonomous driving and quantum computing took before practical systems existed.

The first question is survival. The rocket trip to low Earth orbit lasts about ten minutes, with sustained acceleration loads up to ten times the force of gravity; individual components such as the TPU chips can see 50 to 100 g. To simulate it, the team shook the satellite intensely on a vibration table along all three axes. “Tests like this rarely go as planned, so we were pleasantly surprised that the hardware held up to the force,” Google said.

The second question is radiation. Beyond the atmosphere, solar events and cosmic rays can flip bits inside a chip mid-calculation and inflict cumulative ionizing damage. Google ran AI workloads on Trillium TPUs inside the proton beam at UC Davis’s Crocker Nuclear Laboratory, watching how errors such as a bitflip would affect the work. The chips, the company says, withstood a total ionizing dose greater than what they would receive over a five-year mission. Lab results are still lab results; as Ars Technica notes, missions like the Ingenuity Mars helicopter have shown that off-the-shelf hardware can hold up out there, and Google will not know for certain until orbit.

The third question is the one many observers consider the biggest: cooling. In vacuum there is no airflow, and heat can only escape by radiation — and space radiator systems are built to remove relatively small amounts of it. AI accelerators generate far more. Google’s solution is a layer of malleable thermal interface material connecting the chips to aluminum and copper heat pipes, which conduct the heat into a radiator that projects it into space; the company is also exploring pumped coolants, and has tested the approach in a vacuum chamber on Earth.

It is the cooling system that dictates the rhythm of the first experiment. Google will run its Gemini models on the TPUs, but the radiators can only cope in brief spurts — about fifteen minutes at a time, after which the chips shut down so the radiators can catch up. Fifteen minutes on, rest, fifteen minutes on. “Crucial” groundwork, at that pace, accumulates slowly.

The point of the flight, the team says, is to find out what works and what does not, and to identify failure points for future missions. In 2027 — a test Google has described as a joint mission with Planet — two satellites will trial the laser links that clusters of dozens-of-TPUs satellites would need. Unlike existing inter-satellite systems optimized for low bandwidth over large distances, these lasers must move very high bandwidth over very short ones, a precision problem the company compares to “hitting a coin-size target from miles away while both points are in motion.”

Not everyone is buying the ticket. Gartner analyst Bill Ray called the idea “peak insanity” earlier this year, arguing space-based data centers will not be viable for decades and will never serve terrestrial needs, as The Register reported. The economics carry the skepticism. Google’s own research estimates launch prices would need to fall to about $200 per kilogram for space computing to compete on cost, according to The Next Web; Orbital, a space data-center startup, told The Register the whole venture depends on the cost to orbit dropping from roughly $7,000 per kilogram today to $10 — a target many doubt SpaceX will deliver. The Register notes, in the way of cautionary prehistory, that Google once fielded a moonshot called Loon.

Nor is Google alone on the pad. Starcloud, an Nvidia-backed startup, lofted an H100 chip into orbit last November and has used it to run Google’s open Gemma model, according to The Next Web. SpaceX’s AI wing, SpaceXAI, aims to orbit its first test systems by the end of next year — a simplified system that could eventually grow, in its telling, to a constellation of up to one million satellites, at 250 kilowatts apiece. Google, by contrast, keeps the first satellite to four chips and fifteen-minute shifts, treating the thing as what it is: a moonshot that may go nowhere.

The 2027 launches remain on the agenda, but the team expects it will be years before Suncatcher evolves from “project” to “product,” Ars Technica reports. Until then, the whole venture reduces to a cooler schedule: one point three square meters of radiator for every ticking, calculating, heat-generating chip, repeated dozens of times per satellite, emptied into space one plate at a time.

## Sources

- [Google's first Suncatcher orbital data center test launches October 1](https://arstechnica.com/google/2026/09/googles-first-suncatcher-orbital-data-center-test-launches-october-1/) — Ars Technica
- [Google will launch its first AI chips into orbit next week](https://thenextweb.com/news/google-project-suncatcher-first-tpu-launch) — thenextweb.com
- [Google's Project Suncatcher to put ML infrastructure in space](https://blog.google/innovation-and-ai/models-and-research/google-research/google-project-suncatcher-facts/) — blog.google
- [Google to test TPUs in orbit next week](https://theregister.com/systems/2026/09/24/googles-tpus-to-catch-some-rays-in-orbit-next-week/5298990) — theregister.com