Google tests space-based AI chips as SpaceX chases orbital data center dreams

Elon Musk and SpaceX are betting big that space will eventually host nearly all computing workloads. Google, by contrast, is taking a step back to answer a more fundamental question: can AI chips even work reliably in orbit?

The two approaches reveal starkly different philosophies about the future of distributed computing. SpaceX and Nvidia are charging ahead with plans to place a fully equipped Vera Rubin AI system in orbit by late 2027, aiming to build out what they envision as orbital data centers. Google's Project Suncatcher, launching this October, is considerably more modest: a refrigerator-sized test satellite carrying four Tensor Processing Units (TPUs) to run experiments and gather real-world performance data.

The case for space computing#

The appeal is straightforward. Satellites enjoy abundant solar power throughout most of their orbits, theoretically making them ideal hosts for power-hungry AI infrastructure. A network of interconnected orbital processors, linked by high-speed laser communications, could in principle eliminate the transmission delays and land-based infrastructure costs that plague modern data centers.

Musk has already declared this future inevitable, stating that "the amount of compute in space will obviously round up to 100% of all compute." Google's framing is considerably more cautious, describing Suncatcher as a "long-term, research moonshot exploring whether space could one day host scalable machine learning infrastructure."

The harsh realities of orbital hardware#

Before Google can dream of a constellation of AI satellites, it first needs to understand whether current chips can survive the space environment at all.

Launch itself is punishing. During ascent, the spacecraft will endure extreme vibration and acceleration reaching 10 times Earth's gravity, with individual components potentially experiencing 50 to 100 g of force—far more violent than any truck delivery of earthbound servers. Once in orbit, the hardware faces continuous bombardment from high-energy radiation absent on Earth. As Carnegie Mellon professor Brandon Lucia explained to The New York Times, "Sometimes, you get a random high-energy particle strike that is like someone throwing a dart at the insides of your computer chip."

Thermal management adds another layer of complexity. Lucia noted that the "cooling problem is actually quite difficult. These computers will be basking in the sun all day." Google's first-generation TPUs can only operate in 15-minute bursts before they must shut down to cool—hardly a foundation for reliable, continuous workloads.