Google’s first data center satellite is online and writing poetry

Google's Suncatcher prototype space data center is online and communicating with the ground, after launching earlier this month on a SpaceX Falcon 9 alongside more than 100 other spacecraft. The satellite carries four Google Tensor Processing Units (TPUs) and generates up to 1 kW of electricity. Its purpose is to lay groundwork for future networks of closely flying satellites that perform large inference workloads in orbit. A Google spokesperson told TechCrunch the satellite was operating nominally.
Commissioning took roughly three weeks of subsystem testing. This week, engineers on the ground sent the first prompt to the model pre-loaded on the spacecraft, Google DeepMind's Gemma 4: write a haiku about being in space. The result was a three-line poem about the Earth below, data streaming, and starlight feeding the model's core. Google said it also tested the model with space-themed limericks and, more substantively, asked the spacecraft to report on the orbital conditions around it.
The power budget is the defining limitation. Although four TPUs are aboard, the satellite's power supply means only one can be switched on at a time, and only for about 20 minutes per session. That complicates space-to-ground communications, which are already difficult because a satellite is visible to a ground station only in windows. In practice, this makes the setup a testbed rather than a service: a short burst of inference must be scheduled, run, and its results downlinked within a narrow slot. I'd expect early experiments to focus on batching small jobs and tolerating latency rather than serving anything interactive.
The ongoing campaign will examine how the chips behave under thermal stress and in the space radiation environment. Google has tested TPUs extensively in particle accelerators on Earth, but wants to observe real orbital conditions. Radiation can cause bit flips — a stored value in memory flipping state — which can corrupt data without tripping ordinary monitoring tools. Ground testing suggested radiation's impact would be minimal for inference workloads and more of a problem for training, and the team will now check that against reality rather than simulation.
If the onboard TPUs prove capable, the next step is launching two more satellites with the advanced chips in 2027. Those spacecraft are slated to support higher power generation and to demonstrate inter-satellite optical links.
That optical-link plan is what distinguishes Suncatcher from other orbital compute proposals. Google's project is described as largely alone in planning high-speed links between satellites to support coherent distributed workloads. SpaceX's data centers are also expected to use laser links but are projected to orbit too far apart to share processing effectively, while companies such as Starcloud and Cowboy Space expect to run inference on individual satellites. The distinction matters technically: splitting a workload across satellites requires low-latency, high-bandwidth connections and tight synchronization, so inter-satellite optics are the enabling piece, not a nice-to-have.
The two visions are closely linked. Alphabet, Google's parent, holds a significant investment stake in SpaceX, and both companies' orbital data center plans depend on Starship, the still-in-development heavy-lift rocket, reaching full and rapid reusability. In a peer-reviewed analysis published earlier this month, Suncatcher researchers estimated SpaceX would need 1,800 Starship flights to achieve the same learning-driven cost reductions seen with its earlier vehicles.
Suncatcher fits Google's pattern of long-horizon infrastructure bets: the design target is infrastructure that is economical in the 2030s. Near-term pressures are separate but relevant to how much patience the company has. Alphabet's stock fell this week after it reported that inserting AI into its search businesses is hurting ad revenue growth.
Why it matters: For developers, this is an early signal that inference hardware is moving off-planet, but the 20-minute single-TPU window and radiation unknowns mean orbital compute stays experimental for now. Teams building latency-sensitive or large-scale AI products should treat space-based inference as a 2030s option rather than a near-term deployment target; my inference is that the 2027 optical-link demonstration, not this haiku, is the milestone worth watching. Google's near-term AI economics, meanwhile, will likely shape how generously the moonshot is funded.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
The first TPU in space is running a Google DeepMind Gemma 4 model and writing cheesy poetry.