Google's own engineers have concluded that putting AI data centers into orbit will only make economic sense once SpaceX's Starship rocket has flown roughly 1,800 times over the next decade, according to a Google analysis tied to the company's Project Suncatcher effort. That works out to roughly 180 Starship launches a year, each carrying about 200 metric tons, for a combined 370,000 tons delivered to orbit by 2035.
Google's estimate extrapolates from SpaceX's historical launch-cost "learning curve" — a roughly 20% annual price decline dating back to the original Falcon 1 — to project that getting a kilogram into orbit could fall to around $200 by 2035. That is the rough threshold, Google's researchers argue, at which the lifetime cost of running a satellite-based data center becomes comparable to the energy and water costs of running the equivalent hardware on the ground.
The analysis landed the same day Google and the satellite company Planet launched a prototype into low Earth orbit carrying four of Google's custom Tensor Processing Units, flying as a rideshare payload aboard a SpaceX Falcon 9 on the Transporter-18 mission. It is the first real-world test of hardware for Project Suncatcher, Google's research effort to determine whether solar-powered orbital data centers could eventually supplement — or outcompete — terrestrial ones.
Promising, but years away
"We've done testing on the ground, but there's no test that's completely as good as the real thing," said Travis Beals, the Google executive leading the project. Early radiation data, he said, has been encouraging for simple inference workloads: "The error rate is very low if you're thinking about typical inference operations, right? Like one in a million." Training large models in orbit remains a harder problem, since the heavier computation involved is far more sensitive to radiation-induced errors.
Google's roadmap calls for two more satellites, equipped with laser links to communicate with each other, to launch next year, building toward a longer-term vision of a cluster of 81 satellites flying in tight formation and sharing compute. The company has previously said it hopes to have a TPU "somewhere in space" generating useful results by 2027, though the current cooling system limits each chip to roughly 15 minutes of continuous operation before it must shut down.
Google is reportedly in discussions with SpaceX about using Starship specifically for future Suncatcher launches, while SpaceX has separately filed paperwork with federal regulators for a constellation of up to a million orbital data-center satellites of its own. Not everyone in the industry is convinced the economics will arrive on schedule: Amazon Web Services' chief executive said this year that a vision of a million space-based data centers is still "just not economical," underscoring that Google's own 1,800-launch threshold depends on Starship achieving a launch cadence and reusability it has not yet demonstrated.
