Google's Project Suncatcher tests AI chips in orbit

Google spent years worrying about radiation - and for good reason, since the TPUs flying on this mission are standard commercial Trillium (TPU v6e) chips, not radiation-hardened silicon. But once the prototype reached orbit, the harder problem turned out to be heat. There is no air or water in space to carry waste heat away from a chip, so the satellite's only cooling option is thermal radiation into the vacuum. That physical limit forces the TPUs to run AI inference on Gemini and Gemma models in bursts of roughly 15 minutes before shutting down so the satellite's radiators can shed the heat that built up ^[1]. It is a workable workaround for a four-chip prototype, but the same 15-minute duty cycle becomes the defining engineering question the moment Google scales toward the dozens-of-chips-per-satellite designs it envisions for the future ^[2].

Google is not the only one betting on this. Nvidia-backed startup Starcloud already flew an Nvidia H100 GPU to orbit on its roughly 60 kg Starcloud-1 satellite in November 2025 and used it to run Google's own Gemini model from space ^[3][4]. Nvidia itself has since announced a Space-1 Vera Rubin module that it claims delivers up to 25 times the H100's inference performance in orbit, positioning itself as both chip supplier and rival to Google's TPU approach ^[3]. The competition extends well beyond hardware vendors: Elon Musk has said SpaceX expects to start deploying its own orbital AI compute satellites as early as 2028, Jeff Bezos has floated gigawatt-scale data centers in space within a decade or more, and Eric Schmidt has reportedly acquired rocket company Relativity Space specifically to pursue putting data centers in orbit ^[5]. Several of tech's most powerful figures are independently chasing the same bet, and Google's TPU prototype adds its own flight data to a field where Starcloud already has an operational head start.

Google's own ground testing, firing a proton beam at the chips inside UC Davis's Crocker Nuclear Laboratory, found the Trillium TPUs could absorb a total ionizing radiation dose greater than what they would receive during a five-year space mission ^[2]. The in-orbit results so far back that up and then some: the chips survived a 15 kilorad dose, far beyond the roughly 750 rad(Si) a shielded five-year mission in the target orbit would actually accumulate ^[1]. The headline number is reassuring. The detail underneath it is not: most radiation-induced bit flips were recoverable simply by restarting the chip, but testing also detected one silent data corruption event ^[1]. A crash is an annoyance; silent corruption is the failure mode that is hardest to catch and most dangerous to trust, because the chip keeps running and reports nothing wrong while quietly returning bad results. For a four-chip, 15-minute-burst prototype that is a footnote. For a future cluster of dozens of TPU chips running unsupervised machine learning workloads, it is the kind of reliability question that will need answers before anyone trusts the output.

Not everyone is convinced this moonshot clears the bar. Astrophysicist Neil deGrasse Tyson has dismissed orbital data centers outright as a failed business model, arguing that ground-based renewable energy will stay cheaper than launch economics can ever overcome ^[6]. Scientific American reports that experts broadly caution there are numerous engineering and cost efficiency challenges still unsolved, including the prospect that maintaining hardware in orbit could cost 10 to 100 times more than repairing a terrestrial data center ^[7]. Astronomers quoted in the same reporting warn that large orbital data-center deployments could worsen light pollution and make the night sky harder to observe ^[7]. Google's own illustrated endgame for the program is an 81-satellite cluster roughly a kilometer across ^[9]. Separately, University of Texas aerospace engineering professor and orbital-debris expert Moriba Jah has raised concern that large-scale orbital compute deployments generally could add meaningfully to collision risk in an already congested orbital shell, and prefers thinking in terms of orbital carrying capacity rather than a strict Kessler-syndrome framing ^[8]. That split mirrors the reaction online, where enthusiast communities framed the launch as a first step toward a solar-powered Dyson swarm while more skeptical audiences questioned whether it is anything more than a flashy way to justify AI capital spending.