Google's first data center satellite is operational in orbit
Google's prototype space-based data center, under Project Suncatcher, is now online and in contact with the ground after launching on a SpaceX Falcon 9 earlier in October. After roughly three weeks of commissioning, engineers this week sent the first prompt to the pre-installed Gemma 4 model on board — the answer was a haiku about being in space. But the power budget currently limits the chip to around twenty minutes of operation at a time, and an ongoing test campaign against radiation and thermal stress will determine whether two follow-on satellites can launch in 2027.
What happened
According to TechCrunch (Tim Fernholz), which reported the milestone on October 30, 2026, the satellite is now communicating with ground stations after launching earlier this month alongside more than a hundred other spacecraft on a Falcon 9 (TechCrunch, syndicated via MSN).
Commissioning took around three weeks and involved testing the satellite's subsystems. Engineers on the ground then sent the first prompt to the model loaded on board: Google DeepMind's Gemma 4. They asked the model to write a haiku about being in space — an inference performed on TPU hardware in orbit as part of a space-based computing program.
A Google spokesperson told TechCrunch that the satellite is "operating nominally." It is worth spelling out what that means: the model's actual response, and the satellite's health status, are information supplied by Google through TechCrunch. There is no independent verification of the inference result in the available source material, and no primary documentation from Google itself, such as a blog post or press release.
What the satellite actually is
The prototype carries a single Google TPU (Tensor Processing Unit) and generates up to 1 kW of electrical power. That may sound like a lot for one satellite, but it is modest measured against what the program is, according to TechCrunch, meant to pave the way for: networks of closely flying satellites performing large inference workloads in orbit.
TechCrunch describes the craft as Google's first data center satellite. That is what makes the milestone concrete: not just that Google has sent a chip into space, but that the chip is now powered up, communicating, and has performed an actual model inference on board.
The hardest constraint: power
The most revealing number in the story is twenty minutes. The TPU's power demands mean it can only be switched on for around twenty minutes at a time on the available power budget, TechCrunch writes.
That complicates, according to TechCrunch, an already difficult task: communication between space and the ground. In practice, a computing job has to both fit within the power budget and be coordinated with when the satellite can actually downlink data. For a haiku, that is manageable. For "large inference workloads" — what Suncatcher is, according to the reporting, meant to demonstrate the path toward — the combination of short operating windows and limited ground contact is one of the main challenges the program must solve.
What is being tested now: radiation and heat
The ongoing test campaign is examining how the chip behaves under thermal stress and in the radiation environment of space. Google has, according to TechCrunch, tested its TPUs extensively in particle accelerators on Earth, but must now see how radiation's real-world effects impact the processor in actual orbit.
One specific concern is so-called "bit flips" — individual bits flipped by radiation that can corrupt data without triggering the usual monitoring tools. For a system meant to run inference whose answers must be trustworthy, silent data corruption is a serious problem: you get wrong results without knowing it.
Google has, according to the report, assumed — based on ground testing — that radiation effects will be minimal for inference, while training will be a bigger problem. But this is an expectation from ground tests, not a finding: the whole point of the ongoing campaign is to check whether the assumption holds in real space conditions. For now it remains an open question, not an answered one.
The road ahead: two satellites in 2027
If the on-board TPU proves capable, the next step in Google's plan is two new satellites with more advanced chips in 2027. These are intended to support higher power generation and demonstrate optical links between satellites — a key technology for building networks of satellites that can work together in orbit, which is the precondition for Suncatcher to handle larger workloads than a single chip alone.
The timeline and the condition are worth noting: the follow-on launch is explicitly conditional on today's prototype demonstrating that the chip can withstand the conditions. Nothing in the sources suggests this has already been decided.
The economic assumptions behind it
The Suncatcher researchers themselves have calculated what the program presupposes. In a peer-reviewed analysis published earlier in October, they estimate that SpaceX would have to fly Starship 1,800 times to achieve the same learning-driven cost reductions the company saw with its earlier rockets. The estimate is the researchers' own cost modeling, not an achieved target — it describes how large a volume increase they assume is necessary for space-based computing to become cost-competitive.
At the same time, the number puts the milestone in perspective: one satellite writing a haiku on twenty minutes of power is an early technology demonstration step in a program whose economic foundation still rests on assumptions about future, extremely cheap and extremely frequent launches.
What we don't know
The source base for this story is narrow. TechCrunch's reporting is the only substantive basis; a version syndicated via MSN is the exact same story, not independent confirmation. That means every detail — the satellite's condition, the haiku, the 2027 plans — originates from Google, relayed through one journalist. Several details, such as the exact launch date, are not given in the sources ("earlier this month" is all that can be said).
The most important open questions follow from the test program itself: How well does the TPU withstand real radiation over time, and will radiation's effect on inference remain as small as the ground tests suggested? And can the tension between a 1 kW power budget and the requirement for large inference workloads be resolved? The answers will determine not only whether two satellites launch in 2027, but whether the entire Suncatcher concept has a path from demonstration to operation.