Google's orbital computing satellite launched today on a SpaceX rocket from California, marking the tech giant's first time sending its advanced chips into space.
This satellite, manufactured by Planet Labs, will be used to test whether Google’s Tensor Processing Unit (TPU)—its chip designed to compete with NVIDIA’s GPUs—can function properly in space. This requires providing a continuous 1 kilowatt of power, cooling the chip, and running a series of models to observe if any issues arise.
“We’ve done tests on the ground, but you know, no test can fully replicate the real environment,” said Travis Beals, the Google executive in charge of Project Suncatcher, the tech giant’s initiative to develop a large-scale orbital computing cluster orbiting Earth.
Once operational, the satellite will activate TPUs in groups of 15 minutes to avoid overwhelming its power and thermal management systems. The satellite is built on Planet Labs’ standard platform, but the two companies are collaborating on a demonstration project scheduled for launch next year, featuring two satellites specifically designed for advanced computing. Future versions will also attempt to coordinate work via laser communication links.
Suncatcher is not the only AI payload on this SpaceX rocket. This launch carried over 100 different payloads, including missions from Satlyt and Cowboy Space Company.
However, Google's project differs most significantly from these startups and even SpaceX's own initiatives in that it is a long-term plan.
Beals said the focus of this "long-term moonshot" project is to prepare for future space infrastructure and AI workloads. The company envisions a network of 81 satellites flying in close formation and processing data in parallel.
“When you’re trying to run multi-rack workloads, the bandwidth and latency between TPUs really matter… We’re trying to look beyond today’s workloads to what they’ll look like five years from now,” Beals said. He noted that this is largely because the rockets needed to scale data centers cost-effectively don’t yet exist.
On Thursday, Google also released the peer-reviewed version of its white paper on orbital data centers, one of the most rigorous analyses to date on how computing power can enter orbit. The paper will be published in Joule.
One of the most notable aspects of this paper is how Google views the cost of accessing space. Although the researchers emphasize that their analysis is not an economic feasibility study, it still illustrates how the company perceives the decline in rocket costs over time.
Like all data center companies, Google is counting on SpaceX to help launch its spacecraft into orbit. (Google is also a major investor in SpaceX.)
The paper's authors argue that since the launch of the Falcon 1 rocket, Elon Musk's rocket team has achieved an annual learning curve cost reduction of approximately 20%, making it reasonable to expect that by 2035, the company could reduce launch costs to nearly $200 per kilogram.
To achieve this, what is required? Based on the payload capacity of Falcon 9 launches, the author estimates that a similar cost-reduction path would require Starship to deliver 3.7 million metric tons into orbit. This implies approximately 1,800 launches over the next decade—180 launches per year—assuming each mission can carry 200 metric tons.
For a vehicle that has never flown more than five times in a year, this requirement is quite high. SpaceX expects its vehicles to fly far more frequently—for example, Elon Musk has said that Starship could achieve a flight rate of once per hour by 2029, though Musk often makes many statements.
At least in the post-Google update research, the good news is that its chips appear likely to withstand space radiation. The company had to retest the chips in a particle accelerator because it realized the chip’s configuration provided more shielding than in actual environments. This led to a slight increase in errors in the chip’s logic circuits, but the company still believes its chips can handle large-scale inference workloads over a satellite’s five-year lifespan.
“If you’re talking about typical inference operations, the error rate is extremely low, right? Like one in a million,” Beals said. “On the other hand, if it’s some massive training workload, then problems already arise, because you’re running tens of thousands of chips continuously for months.”