Google’s Project Suncatcher Just Ran AI in Orbit: A Haiku From Space and What Comes Next

On October 30, 2026, a satellite flying over the Norwegian Arctic did something no computer in orbit had done before. It ran a full AI prompt on a data center class chip, generated an answer, and sent it back to Earth. The answer was a haiku.

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

Google calls the moment the "Hello World" of Project Suncatcher, its long running research effort to find out whether AI computing could one day run in space, powered directly by the Sun. The poetry was simple. The engineering behind it was not.

What Actually Happened Above Svalbard

The satellite, a prototype built with Earth imaging company Planet and known as MVP, launched on October 1 aboard a SpaceX Falcon 9 rocket on the Transporter-18 rideshare mission, sharing the ride with more than 100 other spacecraft. It then spent about three weeks in commissioning while engineers checked its systems.

During a ground station pass in the last week of October, engineers uploaded a prompt to the onboard model, Google DeepMind's Gemma 4, asking it to write a haiku about being in space. The reply came down as the satellite crossed Svalbard, home to one of the world's busiest polar ground stations. Google announced the result on October 30.

The team did not stop at poetry. According to reports, engineers also tried space themed limericks and prompts asking the model to report telemetry about the satellite's own systems, a practical check that the chip and the spacecraft bus are talking to each other properly. Google chief executive Sundar Pichai described the transmission as a hello world milestone, comparing it to the first basic test programmers run on a new system.

Four TPUs and One Kilowatt of Sunlight

The prototype carries four of Google's Tensor Processing Units, or TPUs, the chips it uses to run AI models in its data centers. These are Trillium generation chips, the same class of hardware that powers large scale AI workloads on the ground.

The catch is power. The satellite's solar panels produce roughly one kilowatt of electricity, and that is only enough to run one TPU at a time for about 20 minutes, drawing on energy stored in its batteries. Omar Sanseviero of Google DeepMind noted publicly that the model ran on stored sunlight.

Why Put AI Chips in Space at All

Data centers on Earth are running into familiar limits. They need enormous amounts of electricity, they need cooling, and they need land near power grids. In orbit, sunlight is constant and abundant, and there is no shortage of room. Google's idea is to test whether some of that computing could eventually move off the planet.

The logic has an obvious appeal and an equally obvious set of problems. Space is a harsh place for electronics. Radiation can flip bits in memory. Vacuum makes cooling hard, because there is no air to carry heat away, so the satellite has to conduct heat from the processors and radiate it into space.

Google has been working through those problems in the lab before risking them in orbit. In the research paper behind Suncatcher, the team reported irradiating its Trillium TPUs with a proton beam and found they survived a total dose equivalent to a five year mission without permanent failures, while the researchers measured the rate of bit flips. That lab result is what gave the team confidence to fly the hardware.

What the Haiku Proves, and What It Does Not

It is worth being honest about what a single short answer actually proves. It confirms the complete inference path works: a ground station sent the prompt, the orbital system activated the accelerator, the model generated text, and the satellite returned that text to Earth. That is a real first, and it matters.

It proves almost nothing about sustained computing. A haiku gives no benchmark for speed under load, cost per token, error rates over months, or what happens when a particle strike flips a bit in memory at the wrong moment. The radiation test that matters most has not happened yet in any meaningful way. Google says thermal and radiation tests come next.

The project was announced in November 2025 together with a preprint paper, now peer reviewed and published in the journal Joule. From the start, Google has framed Suncatcher as research, not a product. There is no public API, no pricing, and no date when developers could run workloads on such a system.

What Comes Next

Google says the next milestone is a pair of interconnected satellites in 2027, which will generate more power and demonstrate optical laser links between spacecraft. The longer term vision is a constellation flying at roughly 650 kilometers in a dawn dusk sun synchronous orbit, where each spacecraft would carry dozens of Trillium v6e Cloud TPUs. Some reports put the long term goal at a cluster of up to 81 satellites.

That is still a long way from a working orbital data center. Two satellites in 2027 would be an experiment, not a service. The real questions, how the chips handle months of radiation, how they shed heat at scale, and whether the economics ever make sense, will take years to answer.

Still, the first step is now on the record. An AI model ran on a data center chip in orbit and answered in verse. For a project named after catching sunlight, that feels like the right kind of beginning.

Frequently Asked Questions

What is Google Project Suncatcher?

Project Suncatcher is Google's research effort to test whether AI computing can run in orbit, powered directly by solar energy. A prototype satellite called MVP, built with Planet, launched on October 1, 2026, and ran its first AI prompt on October 30, 2026.

What did the satellite actually do?

Engineers uploaded a prompt asking Google's Gemma 4 model to write a haiku about being in space. The model ran on a Trillium TPU aboard the satellite and sent back three lines: "Blue marble below / Data streams in silent grace / Starlight feeds my core."

When will orbital data centers become real?

Not soon. Google plans two more interconnected satellites in 2027 to test laser links and higher power. A full constellation would take years of further testing on radiation, cooling, and cost before it could serve real workloads.