Confirmed 30 Oct 2026 · Compute A one-minute read

Google ran a model on a data centre chip in orbit and got an answer.

Google says its prototype satellite ran the Gemma 4 model on an orbiting processor and sent back a haiku. Caveat: Google gives no power, radiation or throughput figures, and does not date the pass.

Until now

Compute in space is a slide in a research deck.

Now

One data centre chip has now answered a prompt from orbit.

Why it matters

Power and cooling are the binding limits on the ground, and permits take years. Orbit offers constant sunlight and no neighbours. A haiku proves almost nothing about capacity, but it moves the question from whether the hardware survives to what it costs. Google's next step is 2 satellites talking to each other.

Established

A data centre class TPU in orbit produced an AI response.
The satellite launched on October 1 and was commissioned within weeks.
Google names twin interconnected satellites as its 2027 milestone.

Not established

That orbital compute works at scale. One prompt is not capacity.
That it is cheaper. Google publishes no cost per operation.
That the chips endure. Thermal and radiation tests come later.

The five-minute readThe record, the timeline, the weight and 4 sources

The record

Google published a post titled "We've received our first response from the Project Suncatcher prototype satellite" on October 30, 2026, by Travis Beals, senior director, Paradigms of Intelligence, read by us at blog.google. It says Google launched the prototype satellite on October 1 and calls it "the first step in our long-term research moonshot to explore powering AI compute in space". Over the following weeks the team completed a commissioning checklist, including powering up the host and TPU trays and running early machine learning tests. The post calls the result a "Hello World" milestone and "the first-ever AI response from a data center class TPU in orbit". During a ground station pass the team uplinked a prompt to an onboard Gemma 4 open model asking for a haiku about being in space. The reply, downlinked as the satellite passed Svalbard, Norway, reads: "Blue marble below / Data streams in silent grace / Starlight feeds my core." Follow-up tests used space-themed limericks and telemetry prompts reporting the status of payload systems. Next come tests of how the TPUs perform under thermal stress and radiation exposure in orbit, and a named next milestone of "twin interconnected satellites in 2027". The post does not name the TPU model, the rocket, the satellite, any partner or the date of the pass, and gives no power, radiation, latency or link rate figures. Reported, not read by us at a primary source: the satellite flew on SpaceX's Transporter-18 rideshare on October 1 and was built with Planet; the payload is described as carrying 4 TPUs on a power budget of about 1 kilowatt, able to run one of them for roughly 20 minutes at a time. We searched EDGAR full text for Suncatcher across October and found no filing. Dating: this card is dated October 30, the date of the only public record of the result. Google does not say when the pass happened and we do not assume it.

How it got here

Oct 1
Google launches the Project Suncatcher prototype satellite, which it calls the first step in a research moonshot on powering AI compute in space.
Oct 7
Google pays for 890 megawatts of nuclear uprates with Constellation rather than buying existing output, the first compute buyer to fund new generation. Edition 044's power card.
Oct 28
American Electric Power discloses 84 gigawatts of contracted large load, most of it data centres. Edition 065's power card.
Oct 29
Alphabet's Form 10-Q shows $130.5 billion of future payments on leases not yet commenced, primarily data centres and energy infrastructure. Edition 066's capital card.
Oct 30
Google publishes the first response from the satellite: a haiku from a Gemma 4 model running on an orbiting data centre class TPU, downlinked over Svalbard.

Who has to react

Nvidia, whose accelerators are not the ones in orbit; the launch providers selling rideshare capacity; the orbital data centre startups that have raised money on this thesis; spectrum and orbital slot regulators; the utilities and grid operators holding the queue Google is trying to leave; and Alphabet shareholders already paying for ground capacity.

What to watch

Would strengthen itGoogle launches the twin interconnected satellites, or a second company runs a data centre class AI accelerator in orbit, by December 31, 2027.
Would weaken itGoogle publishes thermal or radiation results showing the TPU degraded, or the twin satellite milestone slips past 2027 with no new date, by June 30, 2028.
Would strengthen itGoogle or a partner files for spectrum or orbital slots naming AI compute, or publishes a cost per operation in orbit, by December 31, 2027.

Why it weighs this much

Signal strength

CostMedium

CommitmentMedium

ReversibilityHigh

PersistenceYears

Cost is medium. Building a payload with host and TPU trays, buying a launch and running a ground station campaign is real money, but it is one prototype on a rideshare, not a constellation. Commitment is medium: the hardware is in orbit and cannot be recalled, and Google has named a 2027 milestone in public, which is harder to abandon quietly than a paper. Reversibility is high, because research programmes are cancelled all the time and nothing here is sold to a customer. Persistence is years, since the stated path runs to twin satellites in 2027 and the physics questions the post names, thermal stress and radiation, resolve over that horizon rather than this one.

The words

TPU: Tensor processing unit, Google's own AI chip. A data centre class part is the same silicon it runs on the ground, not a cut-down version.
Gemma 4: Google's family of smaller open-weight models, published for anyone to download and run on hardware Google does not operate.
Downlink: The radio transmission from a satellite to a ground station. A satellite can only answer while a station is in view.
Commissioning: The checklist a newly launched satellite works through before doing real work, powering each system up in turn.

Sources

Your call

Which explanation do you believe?

A call is final. Other people's calls show after yours.