Google thinks SpaceX’s Starship needs to launch 1,800 times before space data centers take off

Google’s prototype of its orbital computing satellite took off today aboard a SpaceX rocket launched from California – the first time the tech giant has sent one of its advanced chips into space.

Built by Planet Labs, the satellite will prove that a Tensor processing unit from Google, its competitor to Nvidia’s GPUs, can operate in space. That means providing a kilowatt of continuous power, cooling the chip, and running a series of models through its paces to see if anything is wrong.

“We’ve done some field testing, but you know, no testing is as good as the real thing,” said Travis Beals, a Google chief executive who manages Project Suncatcher, the tech giant’s plan to develop large-scale computing clusters in orbit around Earth.

Once commissioned, the satellite will trigger its TPU in 15-minute bursts to avoid straining the satellite’s energy and thermal management systems. This satellite is based on a standard platform built by Planet Labs, but the two companies are working on a demonstration that is expected to take flight next year and will see two more satellites specifically designed for advanced computing that can run larger workloads. These future versions will attempt to collaborate via a laser communications link.

Suncatcher isn’t the only space AI payload on this SpaceX rocket, which launches more than 100 different payloads, including missions from Satlyt and Cowboy Space Company.

What sets Google’s initiative apart from these startups (and even SpaceX itself) is that it is a long-term project.

The goal of this “long-term project,” as Beals puts it, is to build the space infrastructure and AI workloads that will exist in the future. The company envisions an orbital data center that is a network of 81 satellites flying in close formation, processing in parallel.

“Bandwidth and latency between TPUs are really, really important when you’re trying to run a multi-rack workload…we’re trying to anticipate not only what workloads are out there today, but where they’re going to be five years from now,” Beals said. This is largely because the rockets needed to cost-effectively expand orbital data centers don’t yet exist.

On Thursday, Google also released a peer-reviewed version of its Orbital Data Centers white paper, one of the most rigorous analyzes available of how calculations get to orbit. The article will be published in Joule.

One of the most notable aspects of the document is how Google views access to space. Although the researchers emphasize that their analysis does not constitute an economic feasibility study, it does offer an interesting picture of how the company envisions rockets becoming cheaper over time.

Like all data center companies, Google is turning to SpaceX to get its spacecraft off the ground. (Google is also a major investor in SpaceX.)

Arguing that Elon Musk’s rocket builders have hit a “learning curve” of price reductions of about 20% per year since the launch of the Falcon 1 rocket, the authors say it is reasonable to expect the company to deliver launch prices close to $200 per kilogram by 2035.

What will it take to get there? Based on the amount of payload launched by the Falcon 9, they believe a similar cost-reduction trajectory will require Starship to put 370,000 tons of payload into orbit. That would require about 1,800 launches over the next 10 years, or 180 per year — and that’s if it could carry 200 metric tons on each mission.

That’s a big ask for a vehicle that’s never been stolen more than five times a year. SpaceX predicts the company will fly much more than that – Elon Musk suggested Starship could achieve a hourly flight rate in 2029, for example, but Musk says a lot of things.

The good news, at least, in Google’s updated research is that it seems likely that its chips will survive the radiation from space. The company had to retest blasting the chips in a particle accelerator after realizing that the configuration of the chips provided more protection than they could actually achieve. This produced a few more errors in the chip’s logic circuitry, but the company remains confident that its chips can handle large on-orbit inference workloads over a satellite’s five-year lifespan.

“The error rate is very low if you think about typical inference operations, right? Like one in a million,” Beals said. “On the other hand, it was already problematic to carry out, for example, large-scale training where several thousand chips would work for months.”

Correction: The title of this article originally incorrectly stated that the learning curve estimate for Starship launches was 1,600.; It’s 1,800.