Google's TPU runs its first inference in orbit: Gemma 4 sends back a haiku, downlinked as the satellite passed over Svalbard

Sources first. This article rests on four primary documents: Google's September 24, 2026 pre-launch post, the October 1 post titled Our Project Suncatcher prototype satellite is in orbit, and the October 30 post We've received our first response from the Project Suncatcher prototype satellite, all three by Travis Beals; plus the peer-reviewed Joule paper that the October 1 post links to. Hardware details come from TechCrunch's October 30 report by Tim Fernholz and Ars Technica's late-September pre-launch coverage. Anything from those two outlets is labelled as such and kept out of the official account.

The timeline runs like this. On October 1 the prototype satellite reached orbit aboard SpaceX's Transporter-18 rideshare mission; the spacecraft was built in partnership with Planet. Google confirmed contact the same day and said it was operating as expected. The following weeks were a commissioning checklist: powering up the host and TPU trays, then running early machine learning tests. On October 30 the blog announced the Hello World milestone: the first-ever AI response from a data center class TPU in orbit. Twenty-nine days from launch to first words.

The actual operation was small enough to be underrated. During a ground-station pass the team uplinked its first prompt to the onboard Gemma 4 open model, asking for a haiku about being in space. As the satellite passed over Svalbard, Norway, the downlink returned this: a blue marble below, data streaming in silent grace, starlight feeding its core. Three lines in the original: Blue marble below / Data streams in silent grace / Starlight feeds my core. Google's post keeps the sparkle emoji.

The part with actual signal comes after the poem. Google says it then tested space-themed limericks and telemetry prompts reporting the technical status of the payload systems. The second one deserves a second look: having a model compose natural-language descriptions of telemetry means its role on the link is not mascot but interface, translating structured state into something a human can read. That is my reading, though; the post only says those prompts were tested, not that the model is intended as a ground-operations front end.

Hardware specs belong in separate piles by source. TechCrunch reported that the satellite carries four Google TPUs and generates up to one kilowatt of electricity, and that the power demands of the TPU mean it can only turn one on for about twenty minutes at a time. That report carries a correction stating plainly that the prototype has four TPUs onboard but can only operate one at a time. Ars Technica's pre-launch piece adds other details: the satellite is called MVP, about the size of a refrigerator, with roughly one kilowatt of solar power (a figure it attributes to The New York Times), and the spacecraft comes from Planet. The original plan was two custom satellites in 2027; Google accelerated by integrating its chips into an existing satellite. All of this is media reporting, not Google's wording.

Why orbit at all: the October 1 post gives the official reason. This is the first step in a long-term research moonshot exploring whether space could one day host scalable machine learning infrastructure, and some things can only be tested in space. Over the coming weeks the team will gather in-orbit data on how the TPUs handle the physical stress of spaceflight plus the radiation and thermal extremes. Note the phrasing: exploring whether one day, not about to deploy. The peer-reviewed paper in Joule is the academic version of that argument.

Radiation is the hardest technical point here. TechCrunch explains that radiation causes bit flips, silently changing a stored value and corrupting data without tripping ordinary monitoring. Ground testing suggested the impact would be minimal for inference workloads and more of a problem for training. With a prototype in orbit, that conclusion can finally move from simulation to measurement. So the real question this mission answers is how long this silicon can compute reliably in orbit, not how well the model writes poetry.

The 2027 next step is stated plainly: once TPUs are shown to execute workloads reliably in space, the next milestone is twin interconnected satellites. TechCrunch fills in the engineering: those two will support higher power generation and demonstrate inter-satellite optical links. What makes Google relatively alone here is that it plans high-speed links between satellites to support coherent distributed workloads; SpaceX's data centers also use laser links but orbit too far apart to share processing effectively, while Starcloud and Cowboy Space expect to run inference on individual satellites.

Cost cannot be skipped. TechCrunch notes that a peer-reviewed analysis published earlier in October estimated SpaceX would need to fly Starship 1,800 times to achieve the same learning-driven cost reductions it saw with earlier vehicles, and that both Google and SpaceX need Starship, a heavy-lift rocket still in development and not yet fully reusable, to launch orbital data centers. There is a capital layer too: Alphabet holds a significant investment stake in SpaceX. So the question of who gets to space compute first is not a pure contest at the ownership level.

Boundaries, so this is not mistaken for something usable. First, the official post gives no engineering parameter at all: no TPU model, no orbital altitude, no power, no latency, no throughput, no cost per inference. Second, this is inference, not training; the wording is AI response, and the model is the pre-loaded Gemma 4 open model, not a frontier model like Gemini 4 Argon. Third, one haiku is not a benchmark, and Google says the thermal and radiation measurements are what comes next. Fourth, beyond twin interconnected satellites in 2027 there is no deployment timeline and no statement about when any of this serves real traffic. Fifth, the four TPUs, one kilowatt, twenty minutes and 1,800 launches are media reporting or paper estimates, not Google commitments.

For developers there is nothing to build on here in the short term, but two things are worth tracking. The first is duty cycle: if a TPU can only run about twenty minutes before it has to stop, the natural workload in orbit is batch and asynchronous, not interactive, which is the opposite of the latency-first instinct on the ground. The second is the link: what decides whether orbital compute can scale may not be the chips but whether inter-satellite optical links can stitch several satellites into one machine, which is exactly what the two 2027 spacecraft are meant to test. Both are my judgment, not Google's conclusion.

One line to close: what succeeded is that a data-center-class accelerator can finish an inference pass in orbit and send the result home, nothing more, and that is already enough to move space data centers off the slide deck and onto the to-be-verified list. The two data points to watch are the thermal and radiation measurements and the 2027 twin-satellite optical link demo. Until then, any claim that orbital compute is cheaper than terrestrial has no public data behind it, including Google's own model in the Joule paper, which is still a model.