Google launches Suncatcher TPU satellite, says Starship needs about 1,800 launches

Google's prototype orbital compute satellite took off on a SpaceX rocket launched from California on the day of the article. It is the first time the company has sent one of its advanced chips into space. Built by Planet Labs, the satellite is meant to show that a Google Tensor Processing Unit, its competitor to Nvidia's GPUs, can work in orbit. That means supplying a kilowatt of continuous power, cooling the chip, and running a series of models through their paces to see if anything goes wrong. Travis Beals, the Google executive managing Project Suncatcher, the company's plan to build large-scale compute clusters in orbit, said: "We’ve done testing on the ground, but you know, there’s no test that’s completely as good as the real thing."

Once commissioned, the satellite will fire up its TPU in 15-minute bursts, to avoid straining its power and thermal management systems. It is based on a standard Planet Labs platform. The two companies are working on a demo expected to fly next year with two satellites more purpose-built for advanced compute, able to run more substantial workloads. Those future satellites will try to collaborate over a laser communications link.

Suncatcher is not the only space AI payload on this rocket, which carries more than 100 payloads, including missions from Satlyt and Cowboy Space Company. What sets Google apart from those startups, and from SpaceX itself, is that this is a long-term project. Beals calls it a "long-term moonshot", aimed at the space infrastructure and AI workloads that will exist in the future. The company envisions an orbital data center made of a network of 81 satellites flying in close formation and processing in parallel. Beals said bandwidth and latency between TPUs matter a great deal for multi-rack workloads, and that Google is looking ahead to where workloads will be in five years. That matters in part because, per the article, the rockets needed to scale orbital data centers cost-effectively do not yet exist.

On Thursday Google also released a peer-reviewed version of its white paper on orbital data centers, to be published in Joule. The researchers stress that the analysis is not an economic feasibility study, but it shows how Google sees rockets getting cheaper. Google looks to SpaceX to get its spacecraft up, and is also a major investor in SpaceX, the article notes. The authors argue SpaceX has achieved a price-reducing learning curve of about 20% a year since Falcon 1, and believe it is reasonable to expect launch prices close to $200 per kilogram by 2035. Using the payload flown by Falcon 9 as a guide, they think a similar cost-reduction trajectory requires Starship to fly 370,000 tons of payload into orbit. That would take about 1,800 launches over the next 10 years, or 180 a year, and only if each mission carries 200 metric tons.

The article calls that a big ask for a vehicle that has never flown more than five times in a year. SpaceX predicts far more: Elon Musk has suggested Starship could reach an hourly flight rate in 2029, though the author adds that Musk says a lot of things.

The paper also updates the radiation findings. It now seems likely that Google's chips will survive the radiation of space. The company had to redo particle-accelerator tests after realising the chip configuration gave more shielding than the chips would get in orbit. The redo produced slightly more errors in the logic circuitry, but Google remains confident its chips can handle large inference workloads in orbit for a satellite's five-year lifespan. Beals said the error rate is very low for typical inference operations, "Like one in a million", but that radiation was already problematic for mega-scale training runs with many thousands of chips running for months.

A correction on the article notes that its headline originally gave the Starship launch figure as 1,600; the correct figure is 1,800.

Key facts

Google's prototype Planet Labs satellite carrying a TPU launched on a SpaceX rocket from California as part of Project Suncatcher; it must supply a kilowatt of continuous power, cool the chip and run models, firing the TPU in 15-minute bursts.
Google's peer-reviewed white paper (to be published in Joule) says reaching launch prices close to $200 per kilogram by 2035 would take Starship about 1,800 launches over 10 years, or 180 a year, at 200 metric tons per mission.
Starship has never flown more than five times in a year; the researchers stress their analysis is not an economic feasibility study.
Updated radiation tests in a particle accelerator produced slightly more logic errors, but Google is still confident its chips can handle large inference workloads for a five-year satellite lifespan.
A demo with two more purpose-built satellites is expected next year and will try a laser communications link; the long-term vision is 81 satellites flying in close formation.

Why it matters

This is the first time Google has put one of its advanced chips into space, and it is a first hardware test of Project Suncatcher, the company's plan for large-scale compute clusters in orbit. The paper adds a concrete figure to the debate: about 1,800 Starship launches over 10 years, or 180 a year, if each flight carries 200 metric tons. It shows how much has to go right with launch cadence before orbital data centers could make cost sense. Beals describes the effort as a long-term moonshot.

Who it affects

Google and Planet Labs are directly involved, since Planet Labs built the satellite. SpaceX is both the launch provider and, per the article, a company Google invests in heavily. Other startups with space AI payloads on the same rocket, Satlyt and Cowboy Space Company, work in the same area. Anyone following AI compute supply, including Nvidia's GPU business as the TPU's competitor, has reason to watch whether chips run reliably in orbit.

How to use it

There is nothing to use yet; this is a prototype test. The satellite will be commissioned and then run its TPU in 15-minute bursts while Google checks power, cooling and model behaviour. The next step is the two-satellite demo expected next year, which will try to collaborate over a laser link. Readers who want the full analysis can look for the white paper, which is due to be published in Joule.

How solid is it

The launch and the satellite's goals are reported directly. The white paper is peer-reviewed, and the article calls it one of the most rigorous analyses available of how compute gets to orbit. The 1,800-launch figure is the authors' estimate of what a Falcon-9-like cost trajectory would require, not a forecast. The article's headline was corrected from 1,600 to 1,800. Beals himself says no ground test is as good as the real thing, which is why the flight matters.

Risks and caveats

The researchers say their analysis is not an economic feasibility study. The required launch rate is far above anything Starship has done: it has never flown more than five times in a year, against 180 a year in the estimate. Musk has suggested an hourly flight rate in 2029, but the article notes that Musk says a lot of things. The rockets needed for cost-effective scaling do not yet exist. On radiation, the redone tests gave slightly more logic errors, and Beals says mega-scale training runs, with many thousands of chips running for months, were already problematic. The confidence Google expresses applies to inference workloads over a five-year satellite lifespan.

“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 managing Project Suncatcher