Google just put a TPU in orbit — and says Starship must fly 1,800 more times to make space data centers real

A Google Tensor Processing Unit is in space for the first time. The prototype orbital compute satellite launched today aboard a SpaceX rocket from California, marking the debut flight of the silicon at the heart of Google's plan to build large-scale data centers in orbit.

The satellite, built by Planet Labs on a standard platform, carries a Google TPU — the chip Google positions against Nvidia's GPUs. Its mission is straightforward but hard: prove that the accelerator can actually work in space. That means sustaining a kilowatt of continuous power, keeping the chip cool, and running a series of models through their paces to see what 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 program to develop orbital compute clusters.

Once commissioned, the satellite will fire up its TPU in 15-minute bursts. The short duty cycle is deliberate. Longer runs would strain the satellite's power and thermal management systems. This first spacecraft is a modified off-the-shelf design; the real test of Google's architecture comes next year, when the company and Planet Labs plan to fly a demo with two satellites purpose-built for advanced compute. Those two spacecraft will attempt to collaborate via a laser communications link.

One launch, more than a hundred payloads

Google's experiment is not riding alone. The SpaceX rocket is carrying more than one hundred different payloads, including missions from Satlyt and Cowboy Space Company. Several of those startups are chasing the same idea: putting AI compute in orbit, where solar power is abundant and cooling comes cheap.

What separates Google from the startups — and from SpaceX itself — is timescale. Beals calls the effort a "long-term moonshot," aimed at space infrastructure and AI workloads that do not exist yet. Google's end-state design is 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 stakes are considerable. Data center operators across the industry face tightening power and land constraints on the ground, and orbital compute is one of the more radical proposals for escaping them. If the economics ever close, the market for launch services, spacecraft manufacturing, and AI infrastructure would all shift. Google is also a major investor in SpaceX, giving it a financial as well as technical interest in cheap launch.

The peer-reviewed math: $200 per kilogram by 2035

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

The paper's most notable section concerns access to space. The researchers stress the analysis is not an economic feasibility study, but it sketches how Google expects rocket prices to fall. The authors argue SpaceX has achieved a price-reducing "learning curve" of roughly 20% a year since it launched the Falcon 1. On that basis, they consider it reasonable to expect SpaceX to deliver launch prices close to $200 per kilogram by 2035.

Getting there is the problem. Based on the payload mass launched by Falcon 9, the authors estimate that sustaining a similar cost-reduction trajectory would require Starship to fly 370,000 tons of payload into orbit. At 200 metric tons per mission, that works out to roughly 1,800 launches over the next ten years — about 180 flights a year, sustained, for a vehicle that has never flown more than five times in a single year.

SpaceX itself predicts a far steeper ramp. Elon Musk has suggested Starship could reach an hourly flight rate by 2029. That claim has no precedent in the history of launch, and the gap between 180 flights a year and the handful Starship currently manages defines the schedule risk for the entire orbital data center concept.

Radiation: good enough for inference, not for training

The paper also carries updated radiation test results, and for Google they are mostly reassuring. The company originally blasted its chips with particles in an accelerator, then realized the test configuration gave the chips more shielding than they would actually experience in orbit. Google redid the tests under more realistic conditions.

The results showed slightly more errors in the chips' logic circuitry. Google remains confident the TPUs can handle large inference workloads in orbit across 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 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 matters for how the architecture gets used. Inference — serving trained models to users — tolerates rare errors and can run close to wherever demand exists. Training frontier models cannot tolerate them at scale, which suggests orbital clusters will serve AI workloads before they create them.

What happens next

The immediate milestones are concrete. First, the satellite in orbit today must commission and run its TPU without failure. Second, the two-satellite laser-linked demo scheduled for next year must show that distributed compute in formation flight is more than a white-paper exercise. Third, Starship's flight rate must climb by orders of magnitude if Google's $200-per-kilogram scenario is to hold by 2035.

Each milestone gates the next. Today's launch answers the narrowest question — can one TPU survive and compute in space. The 81-satellite constellation Google envisions depends on all the rest.