Tech • AI • Robotics • Game
Google Just Sent Gemma 4 Into Space and It Answered Back
9.2/10TL;DR
Google has received the first complete AI-generated response from a TPU operating in orbit, a milestone in Project Suncatcher that tests whether future AI data centers could work in space.
KEY POINTS
First AI reply from orbit
Engineers sent a prompt to a satellite carrying a Google Tensor Processing Unit and received a complete answer back on Earth. The model, Gemma 4 from Google DeepMind, wrote a short poem about space and returned it when the satellite passed near Svalbard in the Arctic. Sundar Pichai described the event as the “hello world” moment for Project Suncatcher.
What is flying in space
The satellite was built with Planet and launched on October 1 aboard SpaceX’s Transporter-18 rideshare mission. After roughly three weeks of system checks, the team brought up the main computer and TPUs, ran early machine-learning tests, and then attempted a generative AI task. The spacecraft carries four TPUs, but only one can run at a time.
A very limited prototype
The current system produces only about 1 kilowatt of power, roughly the draw of a microwave oven. Because the chips are power-hungry, each can operate for only about 20 minutes at a stretch. That makes continuous computing impossible for now and complicates contact with ground stations, since the hardware is available only in short windows.
Why Google wants orbital computing
The core attraction is energy. Google research says solar panels in the right orbit can generate up to eight times as much power as they would on the ground. In a dawn-dusk sun-synchronous orbit, a spacecraft can receive near-continuous sunlight, reducing the need for large batteries and making space-based computing theoretically attractive.
The long-term design
The concept is not a single giant satellite but a swarm of smaller ones flying in tight formation. A research paper in Joule outlines a cluster of 81 satellites at about 400 miles altitude, spread across a radius a little over half a mile. Some neighboring spacecraft would sit only 330 to 660 feet apart so they could exchange data quickly.
The networking challenge
To mimic a terrestrial AI data center, satellites would need extremely fast interconnects because AI training depends on constant exchange of intermediate results between chips. Google estimates those links must reach about 10 terabits per second. Ground tests of one laser-link pair achieved 800 gigabits per second one way, or 1.6 terabits per second bidirectional, still well short of the target.
Rivals are already in orbit
Google is not first to place AI hardware in space. Startup StarCloud launched StarCloud-1 with an Nvidia H100 GPU and later ran language models and trained a small model in orbit. In August, SpaceX and Nvidia announced StarMind AI-1, using Rubin GPU and Vera CPU hardware for a larger computing payload. Google’s main distinction is its push for laser links fast enough for multiple satellites to work on one shared job.
Heat is the hardest physics problem
In orbit, waste heat cannot be removed by air or conventional liquid cooling into the environment. It must be carried to radiators and emitted as thermal radiation. Google is using heat pipes and radiators tested in vacuum chambers, but scaling that system to dozens or hundreds of chips could add major mass, complexity and launch cost.
Radiation remains a major risk
High-energy particles from the Sun and cosmic rays can flip bits in memory or silently corrupt calculations. Google tested its sixth-generation Trillium TPU with a 67 million electron-volt proton beam. The most sensitive area was high-speed memory, but no permanent dose-related failures were identified up to 15 kilorads, and issues did not appear until about 2 kilorads, near three times the expected five-year mission exposure with shielding.
Economics may decide everything
The project depends on launch costs dropping sharply. The paper assumes about $90 per pound, implying roughly $810 per kilowatt per year, compared with estimates of $570 to $3,000 per kilowatt per year for U.S. data-center electricity. But those economics rely heavily on rapid progress in reusable rockets, especially Starship, which remains under development.
CONCLUSION
The first AI response from a Google chip in orbit is a real technical milestone, but it is still far from proving that space can host practical large-scale AI computing. Power, heat, radiation, networking and launch costs remain the barriers that will determine whether orbital data centers become infrastructure or remain an expensive experiment.
Ask a question