Hajra Naz
Google’s prototype orbital compute satellite took off today. It launched aboard a SpaceX rocket from California. This marks the first time the tech giant has sent one of its advanced chips into space.
Built by Planet Labs, this satellite will test something important. It will prove whether a Google Tensor Processing Unit (TPU) can function in space. That’s Google’s competitor to Nvidia’s GPUs. This test involves several challenges. The satellite needs to supply a kilowatt of continuous power. It must cool the chip properly. It also needs to run a series of models to check for any issues.
“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. He’s the Google executive managing Project Suncatcher. That’s the company’s plan to develop large-scale compute clusters in orbit around Earth.
Once commissioned, this satellite will fire up its TPU in 15-minute bursts. This approach avoids straining the satellite’s power and thermal management systems. This particular satellite runs on a standard platform built by Planet Labs. Still, both companies are working on something more advanced.
A demo satellite expected to launch next year will feature two satellites. These will be more purpose-built for advanced compute work. They’ll be capable of running more substantial workloads too. These future versions will attempt to collaborate using a laser communications link.
Suncatcher isn’t the only space AI payload on this particular SpaceX rocket. The rocket is launching more than 100 different payloads total. That includes missions from Satlyt and Cowboy Space Company.
Read More: SpaceX Plans to Send Starship Into Orbit for the First Time on September 22
What sets Google’s initiative apart from these startups, and from SpaceX itself, is its long-term scope. This is genuinely a long-term project.
Beals describes this effort as a “long-term moonshot.” The focus centers on building infrastructure for future space-based AI workloads. Google envisions an orbital data center eventually. This would function as a network of 81 satellites, flying in close formation. They’d process workloads in parallel together.
“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. This forward-looking approach matters for a specific reason. The rockets required to scale up orbital data centers cost-effectively don’t yet exist.
On Thursday, Google also released a peer-reviewed version of its white paper. This paper covers orbital data centers specifically. It represents one of the most rigorous analyses available on how compute gets to orbit. The paper will be published in Joule.
One notable aspect of this paper involves access to space. Researchers stress their analysis isn’t meant as an economic feasibility study. Still, it offers an interesting picture. It shows how Google envisions rockets becoming cheaper over time.
Like all data center companies, Google relies on SpaceX to get its spacecraft into orbit. Notably, Google is also a major investor in SpaceX.
The paper’s authors make a specific argument. They say Elon Musk’s rocket builders have achieved a price-reducing “learning curve.” That’s roughly 20% per year since the company launched the Falcon 1 rocket. Based on this trend, the authors believe something is reasonable to expect. SpaceX could deliver launch prices close to $200 per kilogram by 2035.
What would it take to achieve that goal? Based on payload amounts launched by the Falcon 9, researchers calculated something specific. A similar cost-reduction trajectory would require Starship to fly 370,000 tons of payload into orbit. That would take roughly 1,800 launches over the next 10 years. That breaks down to 180 launches per year. This assumes Starship can carry 200 metric tons on each individual mission.
That’s a significant ask for a vehicle that has never flown more than five times in a single year. SpaceX predicts it will fly far more frequently than that going forward. Elon Musk has suggested Starship could eventually achieve an hourly flight rate by 2029. Still, Musk makes many ambitious predictions.
Read More: OpenAI Cuts Cursor Off From Its Models After SpaceX Buys the Company
There’s good news in Google’s updated research, at least regarding chip durability. It seems likely that Google’s chips will survive the radiation found in space. The company had to redo certain tests recently. These tests involved blasting chips in a particle accelerator. Researchers realized their original chip configuration provided more shielding than it would actually experience in space. This produced slightly more errors within the chip’s logic circuitry. Still, the company remains confident. It believes its chips can handle large inference workloads in orbit. This confidence extends across a satellite’s full 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.”
