Google is launching AI chips into orbit on October 1
Google is putting AI hardware in space before its rivals even finish their regulatory paperwork. On October 1, a SpaceX Falcon 9 rocket lifts off from Vandenberg Space Force Base carrying Google's first Project Suncatcher satellite — a prototype loaded with four of the company's Trillium TPUs (Tensor Processing Units), the same chips that power its AI workloads on the ground. The goal is straightforward: find out whether AI chips can actually work up there.
The hardware
The satellite, called MVP, draws roughly 1 kW from solar panels — enough to run the four TPUs at the compute level of about one standard data center server. That's modest, but the point is the proof of concept. Per the Google Research Blog, the TPUs were tested in a 67 MeV proton beam and handled radiation doses three times higher than a five-year mission would realistically deliver. Bit flips — random memory errors caused by cosmic radiation — are the primary risk, and Google's plan for handling them is refreshingly low-tech: restart the chip.
We’re taking AI to space with @Google. ?️ Project Suncatcher explores putting AI compute in orbit, powered by the sun. Soon, we’ll be sending our first satellite into space to start testing that idea. It’s a new approach to AI infrastructure. Check out a behind-the-scenes look with the Planet team and our Chief Space Officer, James Mason, featured in @nytimes: https://t.co/d6ko08ZpgJ
— Planet (@planet) September 24, 2026
Cooling is the harder problem. In a vacuum, heat can't dissipate the way it does in air or water. Google engineered a multi-layer thermal system, but it can only keep the chips running for about 15 minutes before they need to shut down and cool off. The satellite will spend six years in orbit, but Google expects to run active AI tasks for only about one year of that time.
The bigger race
SpaceX is seeking FCC approval for up to one million AI satellites; Blue Origin has filed for 52,000. Google isn't waiting — it's shipping hardware first and letting the results speak. According to Cryptopolitan, this hardware-first strategy is a deliberate end-run around the approval backlog that's slowing rivals.
The commercial case hinges on launch costs. Google estimates it needs prices below $200 per kilogram to make orbital compute viable against ground-based alternatives. SpaceX's Transporter rideshare program makes this first test affordable, but scaling to the full target — around 81 satellites by the mid-2030s — will require costs to keep falling.
AI data centers are already straining power grids from Virginia to the UK. Orbital solar has an inherent edge: satellites in low Earth orbit receive roughly eight times more solar energy than ground-based panels. Suncatcher's pitch is that space-based compute could eventually operate entirely off-grid.
What's next
Google plans two more satellites in 2027, including a test of laser inter-satellite links — the communication backbone any real constellation would need. A full orbital data center remains a long-term ambition, not an imminent product. But October 1 is where the experiment begins, and The Decoder notes that even reaching the 81-satellite target would require solving radiation hardening, thermal management, and power delivery simultaneously. For now, four TPUs and a 15-minute cooling window will tell Google whether any of that is worth pursuing at scale.
