Testing AI in orbit: Google prepares 4 processors for Project Suncatcher

First orbital test. Google prepared the satellite with Planet for the Transporter-18 mission. The TPU processors will run Gemini and undergo testing after launch and in space conditions. After every 15 minutes of computation, they will need to pause to cool down.

Before the orbital experiment, Google tested the hardware on Earth. In updated radiation tests, Trillium TPUs withstood a dose equivalent to 5 years of the mission without permanent failures. The team included the results in the 18-page version of its technical preprint dated Jun 17, 2026.

Why orbit. According to Google's Nov 4, 2025 estimate, solar panels in a suitable orbit could receive up to 8 times more energy than on Earth and operate almost continuously. The company is exploring this approach as AI's electricity consumption grows.

Google aims to distribute computation across thousands of satellites and transmit responses to Earth. In the June preprint, the team examined a smaller cluster of 81 satellites with a radius of 1 km and high-speed optical links.

Google ties the economics of such a data center to launch costs falling below $200 per kilogram by the mid-2030s. According to the company's November 2025 calculations, at that price, launching and operating orbital hardware could cost as much as the energy used by a ground-based data center per kilowatt per year.

In 2027, Google plans to test laser communication between 2 Project Suncatcher satellites.