Google sends first TPU compute satellite into orbit under Project Suncatcher

Google's first compute satellite is in orbit

Google has flown a Tensor Processing Unit to space for the first time. According to TechCrunch, the company's prototype compute satellite lifted off aboard a SpaceX rocket launched from California — the first time Google has sent one of its advanced chips into orbit, and an early test of whether the TPU, Google's competitor to Nvidia's GPUs, can function off-planet.

The satellite was built by Planet Labs on a standard platform, and its mandate is narrow but foundational: supply roughly a kilowatt of continuous power to the chip, keep it cool, and exercise a series of AI models to see whether anything goes wrong. "We've done testing on the ground, but you know, there's no test that's completely as good as the real thing," Travis Beals, the Google executive running the effort, told TechCrunch.

Once commissioned, the satellite will run its TPU in 15-minute bursts rather than continuously, to avoid straining the spacecraft's power and thermal management systems.

Project Suncatcher's longer game

The flight is the opening move of Project Suncatcher, Google's plan for large-scale compute clusters in orbit, which Beals calls a "long-term moonshot." The same rocket carried more than 100 payloads, including space-AI missions from Satlyt and Cowboy Space Company. What distinguishes Google's initiative from those startups — and from SpaceX itself — is its timescale: it is designed around space infrastructure and AI workloads the company expects to exist in five years, not today's.

A follow-up demonstration planned for next year will fly two satellites more purpose-built for advanced compute, and the pair will attempt to collaborate over a laser communications link. Farther out, Google envisions a network of 81 satellites flying in close formation and processing in parallel. Beals argues that bandwidth and latency between TPUs are decisive for multi-rack workloads, which is why the company is looking ahead to where workloads will be, not just where they are now.

The launch-cost problem

Alongside the launch, Google released a peer-reviewed version of its white paper on orbital data centers, set to be published in the journal Joule. TechCrunch describes it as one of the most rigorous analyses available of how compute gets to orbit, though the authors stress it is not an economic feasibility study.

The paper's framing of launch prices is its most notable element. The authors argue that SpaceX has been on a price-reducing learning curve of about 20% per year since the Falcon 1, making it reasonable to expect launch prices close to $200 per kilogram by 2035. Reaching that point would require Starship to deliver about 370,000 tons of payload to orbit — roughly 1,800 flights over ten years, or around 180 per year, assuming 200 metric tons per mission. TechCrunch's headline cites 1,600 launches, but the figures given in the article itself work out to about 1,800. That cadence is a tall order for a vehicle that has never flown more than five times in a year, and while Elon Musk has suggested Starship could reach an hourly flight rate by 2029, TechCrunch treats that claim with evident skepticism. Google, notably, is also a major investor in SpaceX.

Radiation: fine for inference, not for training

Google's updated research also revisits whether its chips can survive the radiation of space. The company had to redo particle accelerator tests after realizing its original configuration shielded the chips more than they would actually experience in orbit. The corrected tests produced slightly more errors in the chips' logic circuitry, but Google remains confident the hardware can handle large inference workloads over a satellite's five-year lifetime.

The error rates draw a boundary. Beals puts the rate at roughly one in a million for typical inference operations — but says that was already problematic for a mega-scale training run involving many thousands of chips running for months.

Why it matters

Google has now put one of its own chips in orbit as a proof point, backed by peer-reviewed engineering analysis. But the vision is explicitly conditional: orbital data centers only pencil out if launch costs keep falling, which ties the future of Google's space compute plans to a Starship flight rate that does not yet exist. The radiation results also define the near-term opportunity — inference in space looks plausible, while large-scale orbital training appears out of reach for now.