Google's First Orbital TPU Is in Space, but Its Data Center Plan Needs 1,800 Starship Flights

Google put one of its Tensor Processing Units into orbit for the first time today, sending a prototype compute satellite up on a SpaceX rocket launched from California.

The satellite, built by Planet Labs, will test whether Google's TPU — its answer to Nvidia's GPUs — can function in the vacuum of space. The trial requires supplying a kilowatt of continuous power, cooling the chip, and running a series of models through their paces to see if anything breaks.

"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 managing Project Suncatcher, the company's plan to develop large-scale compute clusters in orbit around the Earth.

Once commissioned, the satellite will fire up its TPU in 15-minute bursts to avoid straining its power and thermal management systems. The spacecraft uses a standard Planet Labs platform. The two companies are already working on a more advanced demo, expected to fly next year, that will put two purpose-built compute satellites into orbit connected by a laser communications link.

A long-term bet, not a product

The Google ride-share carries more than one hundred payloads, including missions from Satlyt and Cowboy Space Company. What separates Google from those startups — and from SpaceX itself — is the timescale.

Beals calls Suncatcher a "long-term moonshot" aimed at space infrastructure and AI workloads that do not exist yet. Google envisions a network of 81 satellites flying in close formation and processing in parallel.

"The bandwidth and the latency between TPUs really, really matters when you're trying to run a multi-rack workload…we're trying to look ahead to not just what workloads exist today, but where they will be in five years," Beals said.

The reason for the long horizon is simple: the rockets required to scale orbital data centers cost-effectively do not exist yet.

The math behind the moonshot

On Thursday, Google released a peer-reviewed version of its white paper on orbital data centers, one of the most rigorous public analyses of how compute gets to orbit. The paper will be published in Joule.

The authors stress the analysis is not an economic feasibility study, but it lays out how Google expects launch prices to fall. They argue SpaceX has achieved a price-reducing "learning curve" of roughly 20% a year since the Falcon 1, and they consider it reasonable to expect launch prices close to $200 per kilogram by 2035.

Getting there demands a lot of hardware. Based on Falcon 9 payload volumes, the researchers calculate that sustaining a similar cost-reduction trajectory would require Starship to fly 370,000 tons of payload into orbit. That works out to roughly 1,800 launches over the next ten years — 180 per year — assuming each mission carries 200 metric tons.

Starship has never flown more than five times in a year. Elon Musk has suggested the rocket could reach an hourly flight rate by 2029. Musk, as the source notes, says a lot of things.

Google is also a major investor in SpaceX, so its orbital data center ambitions and its launch provider's flight cadence are tightly linked.

The chips should survive

Google's updated research carries better news on radiation. The company had to redo its particle-accelerator tests after realizing the chips' original test configuration provided more shielding than they would actually experience in orbit. The revised tests produced slightly more errors in the chips' logic circuitry, but Google remains confident the hardware can handle large inference workloads for the full five-year lifespan of a satellite.

"The error rate is very low if you're thinking about typical inference operations, right? Like one in a million," Beals said. "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."

That distinction frames the near-term opportunity: orbital inference looks feasible, while orbital training at scale remains out of reach until both the chips and the launch economics change. The next real signal arrives next year, when the two-satellite laser-linked demo takes flight.