Opportunity assessment: high ceiling, long runway. Google just sent one of its own AI chips into orbit for the first time. According to TechCrunch AI, a prototype satellite for Project Suncatcher launched today on a SpaceX rocket from California. It carries a Google Tensor Processing Unit (TPU), the company’s in-house rival to Nvidia’s GPUs. Alongside the launch, Google released a peer-reviewed white paper with a blunt conclusion: space data centers won’t scale until Starship flies about 1,800 times.
🛰️ The Mission
Planet Labs built the satellite on its standard platform. The job is simple to describe and hard to pull off: show that a TPU can work in space.
- Power. Supply a kilowatt of continuous power to the chip.
- Cooling. Keep the chip from overheating with no air around it.
- Workloads. Run a series of AI models and see what breaks.
- Duty cycle. Fire the TPU in 15-minute bursts so the satellite’s power and thermal systems don’t get overloaded.
“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, the Google executive running Suncatcher.
🎯 The Bigger Plan
This prototype is step one. Next year, Google and Planet plan to fly two satellites built specifically for heavy compute. They’ll try to work together over a laser link.
The end goal is an orbital data center made of 81 satellites flying in tight formation and processing in parallel. Beals stressed why the formation matters: “The bandwidth and the latency between TPUs really, really matters when you’re trying to run a multi-rack workload.”
Google isn’t alone on this flight. The rocket carried more than 100 payloads, including space AI missions from startups Satlyt and Cowboy Space Company. Google’s edge is patience. Beals calls Suncatcher a “long-term moonshot” aimed at the workloads of five years from now, not today’s.
🚀 The Launch Math
This is the part that stands out. Google’s paper, set to appear in the journal Joule, is one of the most rigorous public looks at what it costs to put compute in orbit. The authors say it isn’t an economic feasibility study, but the numbers are telling:
- The learning curve. SpaceX has cut launch costs by about 20% a year since Falcon 1.
- The target. If that holds, launch prices near $200 per kilogram by 2035 are reasonable.
- The requirement. Getting there means Starship hauling roughly 370,000 tons to orbit.
- The flight count. That’s about 1,800 launches over 10 years, or 180 a year, and only if each flight carries 200 metric tons.
Starship has never flown more than five times in a single year. Elon Musk has floated hourly flights by 2029, but Musk’s timelines have a long history of slipping. Worth noting: Google is also a major SpaceX investor, so it has a stake in that curve bending.
☢️ Radiation Check
The good news is on the chips. Google redid its particle accelerator tests after realizing the original setup shielded the chips more than space would. The corrected tests showed slightly more errors in the logic circuitry. Google still expects the TPUs to handle large inference workloads for a satellite’s 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. Training is a different story. He called it “already problematic” for a mega-scale run with thousands of chips working for months.
📋 Why It Matters
- Energy is the bottleneck. AI’s hunger for power and land is pushing hyperscalers to look for any way out, orbit included. Constant sunlight in space is the pitch.
- Inference first. Expect space compute to serve models, not train them. Training stays on the ground for now.
- SpaceX is the gatekeeper. Every orbital data center plan runs through Starship’s flight rate. Watch that number more closely than any chip spec.
- Timeline reality. This is a 2030s story. Today’s launch proves the hardware can survive. It doesn’t prove the business works.
Outlook
Google has taken the first real step from slide deck to hardware in orbit. The next checkpoint is next year’s two-satellite laser-link demo. After that, it’s up to SpaceX. If Starship can’t hit triple-digit annual launches, orbital AI stays a moonshot. More details are available in TechCrunch AI’s original report.