Google Tests Generative AI Inference from Orbiting Satellite
The Suncatcher satellite successfully generated a haiku using the Gemma 4 model in space.
1Orbital AI Milestone Achieved
Google successfully received a haiku from its experimental Suncatcher satellite in low Earth orbit. Senior director Travis Beals confirmed the Gemma 4 model completed its Hello World milestone by responding to a ground station prompt. This test validates the capability of running machine learning inference directly from space-based hardware platforms.
2Cost Analysis of Space Inference
Generating twenty tokens in space costs roughly $0.000012 via standard API pricing. However, factoring in a 500 kg satellite mass and $7,000 per kg launch costs makes this a significantly expensive operation. Google aims to reduce these costs by amortizing launch expenses over the satellite's lifespan to match terrestrial datacenter electricity spending.
3Future Cost Reduction Projections
Google anticipates launch costs could drop to $200 per kg by 2035, potentially enabled by SpaceX. If achieved, orbital datacenter power costs would align with current US terrestrial rates of $570 to $3,000 per kW per year. The company views prototyping now as a strategic necessity before these lower costs become reality.
4Hardware Resilience and Testing
Beyond poetry, the team is evaluating TPU performance under thermal stress and radiation. Laboratory tests using a 67 MeV proton beam assessed the Trillium Cloud TPU and AMD host server. The high-bandwidth memory showed sensitivity after a cumulative dose of 2 krad, exceeding Google's minimum survival requirement of 750 rad.
5Silent Data Corruption Risks
Research indicates a silent data corruption rate of one event per 17 rad. At an estimated orbital dose of 150 rad per year, this translates to a failure probability of one per 3 million inferences. While acceptable for inference, the team notes that training jobs require further study regarding system-level mitigation efficacy.
6Expanding Orbital Testing Scope
Following the initial haiku, the team successfully tested space-themed limericks and telemetry prompts. These prompts report the technical status of onboard payload systems to ground stations. The ongoing mission focuses on gathering empirical data on how transformer workloads behave in the harsh environment of space compared to controlled laboratory settings.
7Scaling Through Satellite Constellations
Google plans to move beyond single-satellite experiments by testing a pair of interconnected satellites next year. The long-term vision shifts away from large, singular units toward a constellation of many smaller satellites. These units will communicate with one another to create a distributed, resilient orbital computing network for future AI applications.
8Strategic Implications for Infrastructure
This initiative represents a shift in how tech giants approach infrastructure beyond terrestrial boundaries. By validating Gemma 4 in orbit, Google is establishing the technical foundation for space-based AI processing. The project highlights the trade-offs between hardware durability, launch economics, and the necessity of decentralized computing architectures for future global connectivity.