Google's Suncatcher Orbital Data Center Test: 4 TPUs, 15-Minute Bursts
Google is about to find out whether AI data centers belong in orbit. On October 1, its first Suncatcher satellite — a fridge-sized craft called MVP — will hitch a ride on a SpaceX Falcon 9 rideshare. Inside: four of Google's custom TPU accelerators, the same silicon that trains and runs models on the ground. The catch? It can only run for about 15 minutes at a time before the cooling system needs to catch up. This is a test, not a product, and the results will shape whether orbital AI compute is a real path or just a moonshot.
What's Actually Launching
MVP is small by data center standards — roughly refrigerator-sized — and carries four Google TPUs, the custom accelerators used for AI training and inference. Its solar panels supply about one kilowatt, enough to run a microwave or hair dryer. Google didn't build the spacecraft itself; the bus comes from satellite imagery firm Planet Labs, with Google's chips integrated into a satellite Planet had already built for an early test. The original plan called for two custom satellites in 2027, but Google accelerated the timeline. It launches via Transporter-18 aboard a SpaceX Falcon 9, and the mission is expected to operate for just a few months.
Why Space Is So Hard on AI Chips
Radiation can damage chips or flip bits, corrupting calculations. Launch vibration and high g-forces stress hardware too. Google says it has done ground testing, but only orbit will show the full picture. Cooling may be the biggest problem: AI accelerators generate far more heat than the small radiator systems typically used in space are designed to remove. Google's solution uses a malleable thermal interface material connecting chips to aluminum and copper heat pipes that conduct heat into a radiator projecting it into space. That system only supports bursts of about 15 minutes before the TPUs must shut down so radiators can catch up.
The Bigger Bet — and the Reality Check
Suncatcher's long-term vision is a constellation of orbiting AI satellites linked by high-speed laser connections, likely requiring dedicated launches. Backers like Elon Musk and Jeff Bezos have pitched orbital data centers as an alternative to controversial terrestrial facilities, and solar power in space is genuinely attractive given how much energy AI chips consume on Earth. But Google's own team expects years before Suncatcher moves from project to product. The 2027 launches remain on the agenda. This first mission is about learning what works — how Tensor chips handle space, whether off-the-shelf hardware holds up, and whether the cooling math can ever scale.
Key Takeaways
- Google's MVP satellite launches October 1 on a SpaceX Falcon 9 rideshare, carrying four custom TPUs.
- The satellite runs Gemini models but only in ~15-minute bursts because its cooling system can't keep up continuously.
- Cooling, radiation, and launch stress are the core unknowns Google wants to measure.
- The spacecraft bus comes from Planet Labs; Google supplied only the AI hardware.
- Google still targets 2027 launches but says Suncatcher is years from becoming a product.
Source: Ars Technica • 🇺🇸 San Francisco
Keep Reading
