Pings in space: Google puts first AI data centre into orbit

Google has confirmed the viability of running large-scale AI data centres in space, with its Project Suncatcher prototype sending an AI-generated haiku back to Earth several weeks after it was launched into orbit.

That launch, which took place on 1 October, saw Google join forces with space technology firm Planet to put the prototype Suncatcher satellite into orbit around Earth’s poles using the Transporter-18 rideshare mission with SpaceX.

Powered by tapping the large amounts of solar energy available in orbit 24/7, Suncatcher is a precursor to large-scale data centres that will be built by stacking Google’s modular, purpose-built Trillium Tensor Processing Units (TPUs).

Each satellite has dozens of TPUs, and Google envisions linking large numbers of satellites into AI compute constellations that will eventually form large-scale AI data centres – a configuration Google will test next year after launching a second satellite.

The short poem – which read ‘Blue marble below / Data streams in silent grace / Starlight feeds my core’ – was created after researchers asked the TPUs’ onboard Gemma 4 generative AI (genAI) model to write a haiku about being in space.

Google is calling it Suncatcher’s ‘Hello World’ milestone, with Travis Beals, senior director for paradigms of intelligence announcing that “our chips have officially earned their space wings.”

A joint Google-Planet project team celebrates the receipt of the most complex and expensive haiku ever written. Source: Google

“Everyone’s reaction when they hear this idea is ‘that’s crazy’, and there must be a reason it won’t work,” said Google senior director and Project Suncatcher lead Dr Travis Beals.

“The first thing we tried to do is to find that reason – and it turns out that it is possible.”

Engineering AI for the space environment

Proving TPUs’ resilience in space was a significant engineering effort, with the Suncatcher team working closely with counterparts from Planet, an established orbital imaging company that operates over 200 satellites, to stress-test the AI compute modules before launch.

Engineers closely monitored how well the equipment handled the launch – which subjects the TPUs to up to 100 times the force of gravity – and are now tracking its ability to survive an orbital environment saturated with space junk, solar flares and other radiation.

That radiation can cause ‘bit flips’ that impede calculations, while ionising radiation can cause ‘silent data corruption’ events (SDCs) that affect the operation of AI workloads – a phenomenon that Google and Planet engineers were able to replicate during lab testing.

Orbiting at an altitude of 800km along a pole-to-pole route that maximises sun exposure, Suncatcher is now operating without a net as Google engineers monitor its ability to handle both radiation and thermal stress.

Thermal stress is a particular challenge: despite space’s extreme low temperature, satellites can’t use it to cool their components since the lack of air means heat doesn’t radiate away from data centre components, as it does on Earth.

Instead, Suncatcher uses a network of heat pipes, closed pumped fluid loops and flat panel radiators that draw heat away from the TPUs and emit it into space as infrared radiation.

The haiku sent from space. Source: Google

TPUs can generate over 100,000 watts of heat per square metre while radiators can only dissipate around 300 W/m2 – which means Suncatcher has to pause every 15 minutes to let the radiators siphon off the excess heat generated by the high-powered servers.

An early salvo in the new space race

Private firm Starcloud was the first company to put AI into space, last year launching a conventional NVIDIA H100 ‘AI factory’ GPU that was able to run Gemma, but Google believes the specialised design of its TPUs will scale more easily and efficiently.

With satellites now cheaper and faster to launch than ever, new technologies promise to accelerate deployment of orbital data centres, whose solar panels can source five to eight times as much power as ground-based panels due to the lack of atmospheric attenuation.

“We can leverage the advances in the commercial industry,” Planet director of systems engineering Eric Stevens explained, “and immediately send those to space instead of waiting for long design cycles to make it ‘space rated’.”

And while Google has proved the concept works, it’s far from the only company with its eyes on the skies: SpaceX CEO Elon Musk, for one, has previously rhapsodised about the promise of orbiting data centres.

Musk, who poured hundreds of millions of dollars into unsuccessful Republican midterm election campaigns, posted that “data centres in space are looking pretty appealing right now” – a nod to fears Democrats will saddle tech firms with restrictive planning and energy obligations.

With its infinite energy supply and lower cost of deployment, space could prove to be a more viable option to host high-capacity AI computing than many once thought, although the idea has also raised environmental concerns for some.

Coming years will see a flurry of activity rapidly progress projects including Blue Origin’s Project Sunrise, SpaceX’s Starmind, and startups like Orbital, Lonestar, Axiom Space, Cowboy Space Corporation, Satlyt and the European Space Agency’s ASCEND.

“Everything seems too hard until someone tries it,” said Stevens, “and my first reaction was ‘wow, that seems big even for us’ – but it’s an experimental moonshot, and an important first step towards moving data centres off planet and into Earth’s orbit.”