Google's first satellite data center, Project Suncatcher, has produced a work of poetry from orbit, according to reports on Friday. The model behind the feat is DeepMind's Gemma 4, confirming that artificial intelligence can run in space — even as questions remain over the economic viability of orbital computing. The milestone marks an early step in the effort to move parts of the AI infrastructure buildout off Earth.

One TPU, 20 minutes, and three lines of verse

Google launched the prototype on October 1 aboard a SpaceX rideshare, with the work executed in partnership with Planet. Roughly three weeks later, after a thorough checkout, engineers asked Gemma 4 to describe space. It responded:

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

The satellite carries four TPUs, but there is a catch: it can run only one at a time, and for about 20 minutes. It generates up to 1 kilowatt of electricity, TechCrunch reported. Engineers also asked the model to write limericks and report on conditions in orbit. It was a promising first test, but running a commercial data center in space involves far more than writing poetry.

There is logic to Google's turn toward the sky. According to a study published in Nature, electricity consumption by AI-dedicated data centers reached 118 terawatt hours in 2024 and could climb to between 239 and 295 terawatt hours by 2030. Goldman Sachs estimates that global data-center electricity demand will grow 170% between 2025 and 2030, while JLL notes that connecting to power grids can take anywhere from 2 to 10 years. Orbit offers solar power without those waiting times; the main challenge is getting the equipment there and keeping it running.

Which workloads fit, and who gains

Space is not ideally suited to every form of computing. JLL identifies the strongest candidates as AI training, batch processing, and the processing of data generated in space. Real-time transactions, by contrast, require proximity to users, which narrows the addressable market while costs remain high. BCG estimates the orbital cost premium today at roughly 2.5 to 3 times terrestrial levels, though it projects that space could eventually claim 10 to 15 percent of the AI data-center market by 2040 — a forecast, not a certainty.

To capture that projected share, operators must clear three major hurdles: high costs, the effects of heat and radiation on chips, and the difficulty of moving data between satellites. Google's haiku proves an AI model can run in orbit; turning that early success into a commercial business is another matter.

Heat, radiation, and physics that won't cooperate

Cooling is among the most serious technical obstacles. With no air to carry heat away, servers require large radiators, and Brookings cites an estimate of 2.15 million square feet of radiator area for an orbital data center. Cooling is not the only concern: ultraviolet exposure and space debris threaten equipment, and repairs in orbit are difficult. Radiation can also cause memory “bit flips” that corrupt results, TechCrunch noted, underscoring the need for radiation-hardened chips, as discussed by Semiconductor Engineering. Even the environmental case requires scrutiny — an ACM study examines the carbon trade-offs of computing in orbit.

SpaceX, Starcloud, and the 2027 twin-satellite test

Despite those challenges, companies are pressing ahead. Cryptopolitan reported that SpaceX is seeking about $40 billion for Nvidia chips to power AI data centers on Earth and potentially in orbit, though that financing remains at the proposal stage. Starcloud has flown an H100 into orbit, and Cowboy Space is pursuing single-satellite inference, TechCrunch reported.

Google's plan is different: close-flying satellites sharing work through laser links. If current tests succeed, the company intends to launch two higher-power spacecraft in 2027 to test those connections. The next breakthrough will not be another poem. It will be proof that the satellites can work together.