Google's first space data center satellite is live — and it writes haikus

Google's first data center satellite is online, communicating with the ground, and its first job was writing a haiku about being in space. The Suncatcher prototype, launched earlier this month on a SpaceX Falcon 9 rocket alongside more than 100 other spacecraft, carries a single Google Tensor Processing Unit (the company's custom AI accelerator chip) and now runs a pre-loaded language model in orbit.

That model is Google DeepMind's Gemma 4. When engineers sent their first prompt this week, they asked it to write a haiku about being in space. The satellite replied:

Blue marble below Data streams in silent grace Starlight feeds my core. ✨

A Google spokesperson told TechCrunch that the satellite was operating nominally. That one sentence carries a lot of weight here, because the ambitious parts of the mission are still ahead.

Why does one satellite matter?

Suncatcher is Google's bid to move large AI inference workloads — running trained models to generate answers — off Earth and into orbit. The prototype is intended to pave the way for future networks of close-flying satellites performing large inference workloads in space, where solar power is abundant and the constraints of terrestrial electricity grids do not apply.

The stakes are straightforward. AI data centers on the ground are consuming growing amounts of power and hitting grid limits; space-based alternatives promise effectively unlimited sunlight, if the engineering problems are solvable. The engineering problems, at this stage, look considerable.

The satellite generates up to 1 kW of electricity, but the TPU's power demands mean it can only turn on for about twenty minutes at a time. That constraint complicates the already tricky business of space-to-ground communications. A data center that runs in twenty-minute bursts is not yet a data center in any commercial sense.

In plain words: the hardware works, but only in short sprints, and nobody has yet proven the economics.

What is the satellite actually testing?

The spacecraft reached operational status after a commissioning process that tested its subsystems over about three weeks. The ongoing test campaign will examine how the chips perform under thermal stress and in the space radiation environment.

Radiation is the central worry. It can cause "bit flips" — a stored value in computer memory flipping — which in turn can corrupt data without triggering typical monitoring tools. A model that quietly produces wrong answers, with no error flag raised, is a serious problem for any inference service.

Google has extensively tested its TPUs in particle accelerators on Earth, but those tests are simulations of a sort. The team now needs to see how radiation's real effects on their hardware in space affect the processor.

Ground testing suggested the impact of radiation would be minimal for inference workloads and more of a problem for training a model, where corrupted values propagate through a learning process. The Google team will now get to assess those dynamics in reality, not a simulation.

Google also asked the spacecraft to report on the orbital conditions around the satellite — environmental telemetry that matters as much as the poetry. The company said it tested the model with additional space-themed limericks as well.

The takeaway: this mission is less about the haiku and more about whether Google's chips survive orbit without silently corrupting results.

What happens if the test succeeds?

If the TPU onboard this spacecraft proves capable, the next step will be launching two more satellites equipped with the advanced chips in 2027. Those spacecraft will support higher power generation and demonstrate inter-satellite optical links — laser connections between satellites that would let a cluster share work as if it were one machine.

Those links are the differentiating bet. Google's Starcloud project is largely alone among visions for space-based data centers in planning high-speed links between satellites to support coherent distributed workloads. SpaceX's data centers are also expected to use laser links, but the company's satellites are expected to orbit too far apart to share processing effectively.

Companies like Starcloud and Cowboy Space expect to perform inference loads on individual satellites. Google's plan, by contrast, treats the constellation itself as the computer.

That distinction may decide whether orbital data centers ever match the flexibility of terrestrial ones.

How dependent is this on SpaceX?

Google and SpaceX's visions are closely linked. Alphabet, Google's parent company, holds a significant investment stake in SpaceX. And to launch their orbital data centers, both companies will require Starship — the still-in-development heavy-lift rocket at the center of Elon Musk's space vision — to be fully and rapidly reusable.

The cost math is unforgiving. In a peer-reviewed analysis published earlier this month, Suncatcher researchers estimated that SpaceX would need to fly Starship 1,800 times to achieve the same learning-driven cost reductions it saw with its earlier vehicles. That figure comes from Google's own researchers, which cuts both ways: it is the best available estimate, and it describes a hurdle the company cannot clear on its own.

In plain words: Google's space data center plans currently depend on another company's rocket flying hundreds of times more often than it has to date.

Is this a business or a moonshot?

Suncatcher is a classic Google moonshot — the company's focus is on designing infrastructure that will be economical in the 2030s, not next quarter. A single TPU running twenty minutes at a time in low Earth orbit is a physics experiment, not a revenue line.

The near-term question is whether Alphabet sustains the patience. The company's stock plunged this week after it reported that efforts to insert AI into its search businesses are hurting ad revenue growth. A parent company under investor pressure for AI's near-term costs may find a satellite program whose payoffs arrive in the 2030s easier to trim than to defend — or may double down on it precisely because terrestrial compute is getting expensive.

Google's own radiation data, collected over the coming months, will likely settle the technical question before the financial one. Whether the program survives to its 2027 two-satellite follow-up depends less on the quality of the haiku than on whether Alphabet's board still believes in the 2030s timeline when budgets are set closer to home.