"Starlight Fills My Core"... A Poem Written by AI in Space [Weekend Money]

Google's "space data center" prototype satellite activated for the first time
The goal is power... Solar panels can be up to eight times more productive than on Earth
SpaceX and others join the race... Heat dissipation, radiation and other challenges remain

"A blue marble below

Data streams in silent grace

Starlight fills my core

(Blue marble below

Data streams in silent grace

Starlight feeds my core)"

The "Earthrise" photo taken by Apollo 8 in 1968. NASA

This haiku, a traditional Japanese short poem, was written by artificial intelligence (AI) aboard a Google satellite orbiting Earth and sent down to the ground. When Google researchers asked the open AI model Gemma 4 aboard the satellite to "write a haiku about what it feels like to be in space," the satellite sent back this response as it passed over a ground station in Svalbard, Norway.

Google announced on October 30 that it had received its first AI response from a prototype satellite for Project Suncatcher, its space data center concept. This was the first time a data-center-class TPU, Google's custom-designed AI chip, performed inference in orbit and sent back a response. Google named the test "Hello World," a reference to the first step in programming.

Jeong Hoon Jeong, a researcher at Eugene Investment & Securities, said, "The significance of this in-orbit demonstration is not a performance record, but a validation of the premise that data-center-class chips can actually operate in the space environment."

Solar power can be up to eight times more efficient than on Earth... But for now, enough to run one TPU for 15-20 minutes

The moment AI (Gemma) finished responding from space after being asked on the ground to "write a poem." Screenshot from a video released by Google.

The satellite used in the test was built by Planet Labs. It launched aboard a SpaceX Falcon 9 rocket on October 1. After reaching orbit, it underwent three weeks of commissioning and subsystem checks before beginning full-scale AI computing tests.

The satellite carried four TPUs. However, the maximum power it can generate, including through solar power, is just 1 kW (kilowatt) - about as much as it takes to run a hair dryer. Because the chips consume a lot of electricity, the available power is enough to run one TPU at a time for about 15-20 minutes.

Preparations were made for the space environment in advance. Space radiation can cause semiconductors to malfunction, but the chips reportedly operated normally during ground tests before launch, even after being exposed to protons equivalent to five years' worth of space radiation.

Power is the reason Google wants to build data centers in space. AI data centers consume enormous amounts of electricity, and Google has emphasized that space could offer a solution to this power challenge.

In dawn-dusk orbits, which keep the satellites facing the sun as they orbit, solar panels receive continuous sunlight and can be up to eight times more productive than on the ground.

The plan will proceed in stages. Google will first continue testing whether TPUs can compute reliably without data errors amid extreme temperature fluctuations and space radiation. Once these tests are complete, the company plans to launch a pair of satellites next year, in 2027, equipped with high-performance TPUs and optical laser communications. The technology will allow the satellites to exchange data with each other by laser.

The long-term goal is a cluster, or basic unit, of 81 satellites flying in formation across an area about 1 km wide. The idea is to link multiple satellites and use them as a single data center.

Starcloud and SpaceX join the race... Heat dissipation, radiation and maintenance remain challenges

The race to build data centers in orbit is not limited to Google. Starcloud, a startup backed by Nvidia, ran Gemma aboard its "Starcloud-1" satellite, which carried Nvidia's H100 GPU AI chip, and received a response beginning, "Hello, Earthlings," in December 2025. It is preparing an 8 kW-class second satellite for early 2027.

SpaceX is preparing to deploy Starmind, an orbital data center co-designed with Nvidia, in 2027. SpaceX CEO Elon Musk said the company would launch AI satellites powered by Nvidia chips in the fourth quarter of 2027 and expected the network to reach substantial scale in 2028.

Many challenges remain. Experts point to space radiation, the difficulty of maintenance, data center cooling and space debris, and expect commercialization to take more time.

Among these challenges, cooling is a structural problem in space. Data centers on Earth use air conditioning and coolant to remove heat, but in space, where there is no air, a separate heat-dissipation design is needed to release the heat generated by the chips. Heat pipes, which transfer heat, and radiators, which release it into space, are key components.

Jeong Hoon Jeong, a researcher at Eugene Investment & Securities, said, "The companies standing out in the global competition for orbital data centers are SpaceX, Google and Starcloud, among others." He added, "Google is also expected to expand its orbital data center business next year."

This content was produced with the assistance of AI translation services.