Google’s AI chip reached orbit, but space data centers still hinge on a Starship miracle
Google’s AI chip reached orbit, but space data centers still hinge on a Starship miracle
Alphabet first framed Project Suncatcher as an experiment in whether space could eventually host scalable machine-learning infrastructure. Its pitch is straightforward: low-Earth-orbit satellites can draw near-constant sunlight, potentially generating far more solar power than terrestrial installations. But the company also acknowledged that it had not yet seen how its TPUs would behave in orbit.
On Transporter-18, SpaceX launched a Planet Labs-built, solar-powered satellite carrying one of Google’s Tensor Processing Units—the first in-space trial of the chip. Elon Musk amplified SpaceX’s announcement that the Falcon 9 mission would carry 130 payloads from California, underscoring the rideshare scale that made the test possible.
The immediate mission is deliberately modest. Google will run the TPU in 15-minute bursts, testing whether it can sustain power, shed heat and cope with radiation. Travis Beals, who leads Suncatcher, called the flight the necessary reality check: “there’s no test that’s completely as good as the real thing.” Google says its radiation tests point to a low error rate for ordinary inference work, but not for the months-long, thousands-of-chips training runs that define frontier AI.
Next year, Google and Planet Labs hope to launch two more purpose-built spacecraft that can exchange data by laser. The eventual vision is an 81-satellite cluster working in parallel, where bandwidth and latency between TPUs become as important as raw processing power.
That is where the optimism collides with launch economics. Google’s updated analysis suggests SpaceX would need to put roughly 370,000 tonnes into orbit—about 1,800 Starship launches over a decade—to drive costs toward $200 per kilogram by 2035. Musk promotes Starship as a tool for putting much larger hardware into space, but Google’s own math makes clear that orbital data centers will not arrive on a single successful payload. They require a rocket revolution, repeated hundreds of times a year.
Write a comment