Google Tests Space-Based AI Chips with SpaceX Launch, Eyes 1,800 Starship Flights for Orbital Data Centers

First orbital compute satellite demonstrates Google TPU in space

Google launched a prototype of its orbital compute satellite today aboard a SpaceX rocket departing from California, marking the tech giant’s first mission to send an advanced chip into orbit. The spacecraft, built by Planet Labs, will validate whether a Google Tensor Processing Unit (TPU)—the company’s rival to Nvidia GPUs—can operate in the harsh environment of space. That involves delivering a kilowatt of continuous power, managing thermal cooling, and running a suite of models to detect any anomalies.

“We’ve done testing on the ground, but you know, there’s no test that’s completely as good as the real thing,” said Travis Beals, the Google executive overseeing Project Suncatcher, the company’s initiative to develop large‑scale compute clusters orbiting Earth.

Once activated, the satellite will power its TPU in 15‑minute bursts to prevent overloading the power and thermal management systems. While this satellite rides on a standard Planet Labs platform, Google and Planet Labs are already planning a follow‑on demo slated for next year. That mission will feature two purpose‑built satellites capable of heavier workloads, linked by a laser communications system to enable inter‑satellite collaboration.

Project Suncatcher aims for a network of 81 satellites

Suncatcher is not the sole AI payload on this launch; the rocket carries more than 100 payloads, including ventures from Satlyt and Cowboy Space Company. What distinguishes Google’s effort is its long‑term vision. Beals describes the program as a “long‑term moonshot” focused on building the infrastructure and AI capabilities that will be needed years from now.

The ultimate goal is an orbital data center consisting of 81 satellites flying in tight formation, processing tasks in parallel. Bandwidth and latency between TPUs are critical for multi‑rack workloads, and Google is planning for both current and future demands, anticipating that workloads will evolve significantly over the next five years. The company acknowledges that the rockets required to scale orbital data centers cost‑effectively do not yet exist.

Launch cadence and cost assumptions behind the 1,800 figure

Google also released a peer‑reviewed white paper on orbital data centers, set to appear in Joule. The paper, while not an economic feasibility study, outlines how the company expects launch costs to decline. Google relies on SpaceX—both as a launch provider and a major investor—to make its spacecraft viable.

Researchers argue that SpaceX has achieved a price‑reducing “learning curve” of roughly 20% per year since the Falcon 1 era. Extrapolating this trend, they anticipate launch prices approaching $200 per kilogram by 2035. To reach that price point, the analysis suggests Starship must deliver about 370,000 tons of payload into orbit over the next decade. Assuming each mission can carry 200 metric tons, that translates to roughly 1,800 launches—approximately 180 per year.

This demand is steep for a vehicle that has never flown more than five times annually. SpaceX projects a much higher cadence, with Elon Musk hinting at an hourly flight rate by 2029, though such statements remain speculative.

Radiation resilience and chip performance

A bright spot in Google’s updated research is the apparent ability of its chips to withstand space radiation. Initial tests in a particle accelerator had to be redone after engineers discovered the chip configuration provided more shielding than the actual space environment. This adjustment introduced slightly more errors in the logic circuitry, yet Google remains confident the chips can handle large inference workloads throughout a satellite’s five‑year lifespan.

“The error rate is very low if you’re thinking about typical inference operations, right? Like one in a million,” Beals explained. “On the other hand, it was already problematic for doing, say, some mega‑scale training run where you’re going to have many thousands of chips running for months.”

Correction: The headline for this story originally misstated the learning curve estimate for Starship launches as 1,600; it is 1,800.