Google Targets Orbital AI Compute with Project Suncatcher

Google is moving toward orbital AI infrastructure with the launch of a satellite named MVP, scheduled for October 1 aboard a SpaceX Falcon 9 rocket. As Engadget first reported, the mission is part of "Project Suncatcher," an initiative aimed at eventually deploying large AI installations in low-Earth orbit.

The MVP satellite serves as a hardware testbed, carrying four specialized tensor processing units (TPUs) that provide the computing power of a single server. Powered by solar panels delivering one kilowatt of energy, the system is designed to answer simple AI queries for one year, though the satellite will remain in orbit for six years. To combat radiation-induced "bit flips," Google intends to restart the chips when errors occur. The hardware also utilizes a proprietary cooling system using conductive material, though Engadget notes this system only functions for 15 minutes before the chips must be powered down to cool.

While James Manyika, Senior Vice President for Research, Technology & Society at Google, stated the company does not expect anything "usefully operational" in the next few years, the scaling roadmap is aggressive. Google plans to launch two additional similar satellites next year, with a long-term goal of deploying 80 satellites in close formation. The company is also exploring the construction of a satellite the size of a football field.

Significant engineering and financial hurdles remain. Brandon Lucia, a Carnegie Mellon University professor of electrical and computer engineering, noted to The New York Times that expanding the project to a larger scale introduces substantial engineering difficulties. These include managing cosmic rays that can melt semiconductor junctions and the high cost of deployment; Engadget notes a single Falcon 9 launch costs approximately $74 million. Google has not disclosed the total funding allocated to the project.