Google Orbits 4 TPUs to Run Gemini at 8x Solar Power

For AI researchers and infrastructure engineers, the MVP mission turns an infrastructure debate into a measurable experiment: can the same TPU and Gemini stack function in orbit with a single server's worth of compute and 8x solar energy availability? If the one-year test succeeds, it will challenge assumptions about where training and inference workloads must live.

Google took its first concrete step toward moving AI data centers off Earth on October 1, 2026, when a refrigerator-sized satellite nicknamed MVP lifted off from Vandenberg Space Force Base in California at 11:32 a.m. PT. The launch occurred aboard a SpaceX Falcon 9 rocket as part of the Transporter-18 rideshare mission, placing MVP among 130 payloads ranging from cubesats to orbital transfer vehicles. For Google, the launch was not about delivering a product but about answering a research question: Can the specialized AI hardware that powers Gemini operate reliably in orbit?

“

The launch occurred aboard a SpaceX Falcon 9 rocket as part of the Transporter-18 rideshare mission, placing MVP among 130 payloads ranging from cubesats to orbital transfer vehicles.

MVP is the initial in-orbit test for Project Suncatcher, a research effort Google first announced in November 2025. The satellite carries four of Google's tensor processing units, or TPUs, which have roughly the combined computing power of one server in a terrestrial data center. During its approximately one-year operational mission, MVP will answer simple AI queries using Gemini, effectively running a small slice of Google's AI stack in low Earth orbit. Even after operations end, the satellite could continue circling Earth for up to six years before gravity pulls it down and it burns up in the atmosphere. Google emphasized that it is not the first organization to test AI hardware in space, but MVP represents the first step by a major technology company toward building orbital data centers.

The pitch for space-based AI data centers is fundamentally about energy. In low Earth orbit, satellites receive near-constant sunlight, and Google estimates they can generate up to eight times more solar power than panels on the ground. That contrast matters because AI's electricity appetite is growing quickly and terrestrial data center development is running into social and environmental friction. Rural communities across the United States are fighting to keep AI data centers out of their backyards, while many existing facilities strain local water supplies by using cooling towers that shed heat through evaporation. Orbital data centers could, in theory, sidestep both bottlenecks by using abundant space-based solar power and avoiding local water and land conflicts.

The MVP launch also matters because of how it was delivered. Rideshare missions like SpaceX's Transporter-18 have lowered the cost of reaching low Earth orbit, allowing a research satellite with a single server's worth of AI compute to fly as one of 130 payloads rather than requiring a dedicated launch. That launch economics shift lowers the barrier for experimentation, even if orbital AI data centers are still far from commercially viable. The mission is deliberately small: one refrigerator-sized satellite, four TPUs, a one-year operational window, and a modest workload of simple Gemini queries. Google framed the mission around a basic question—"Can our AI hardware operate in space?"—and the first hurdle is likely to be hardware reliability in the harsh environment of orbit.