Google has sent one of its orbital compute satellites aboard a SpaceX rocket to test whether its advanced Tensor Processing Unit, or TPU, can actually function in space. The launch marks a step forward for Project Suncatcher, a long-term research moonshot exploring whether outer space could one day host scalable machine-learning infrastructure.
Testing Power, Heat and Survival in Orbit
Google announced the first test flight under Project Suncatcher, aimed at checking whether several thousand satellites in low-Earth orbit could receive constant sunlight and generate up to eight times more solar power than they would on Earth. Sending TPUs into space is the next step, meant to test whether the chips can function there and eventually compete with Nvidia’s GPUs.
The satellite itself was built by Planet Labs. The experiment is designed to test whether a kilowatt of continuous power can be delivered to the chips, and whether the chips can be cooled using the conditions available in outer space. If those conditions hold, Google plans to run AI models through the chips and see how they actually perform.
We’ve done testing on the ground, but you know, there’s no test that’s completely as good as the real thing, wrote Travis Beals, the Google executive leading Project Suncatcher, in a blog post. Once commissioned, the satellite will run the TPU in short, 15-minute bursts to avoid draining its available power.
Not the Only AI Experiment Headed to Orbit
From SpaceX’s side, Suncatcher isn’t the only payload on this launch. The rocket is carrying 100 different payloads in total, including one from Satlyt, a company founded by former Google and SpaceX managers that recently raised $8 million specifically to run AI on satellites.
The Bigger Goal: An Orbital Data Centre
Google describes this as part of a long-term project focused on building space infrastructure and AI workloads for the future. The company’s eventual goal is an orbital data centre built from a network of 81 satellites, flying in close formation and processing data in parallel with one another.
It also released a peer-reviewed version of its white paper on orbital data centres this week, offering a detailed analysis of how computing could shift into Earth’s orbit over time. A version aimed at a wider general audience is expected soon.
The Economics of Getting There
The research notes that rocket launches are likely to get cheaper over time, a trend Google is counting on heavily. The paper’s authors point out that SpaceX appears to be on a learning curve that could cut launch prices by 20% within a year, with hopes of eventually reaching launch costs close to $200 per kilogram within the next decade. Even then, achieving the payload capacity needed to support a full orbital data centre could require roughly 1,800 separate launches over that period.
Currently, SpaceX’s rocket flies no more than five times a year. The company is hoping that number grows substantially, with Elon Musk claiming Starship could reach an hourly flight rate within the next three years. As with many of Musk’s public predictions, that claim is being met with considerable skepticism.
The Real Question: Can the Chips Survive Radiation?
From Google’s perspective, the more immediate concern isn’t Musk’s launch rate predictions, but whether its own TPUs can actually survive space radiation. The company has been testing ways to improve the chips shielding, but found that added shielding came with a tradeoff: the chip’s logical circuitry experienced a higher rate of errors as a result.
Even so, Beals believes Google is on the right track, estimating the resulting error rate at roughly one in a million, low enough, in his view, that it shouldn’t meaningfully disrupt inference operations. Several other experiments have already flown aboard SpaceX rockets testing similar questions, but Google considers its own experiment the most significant test of the group so far.
