Google is about to find out whether its AI chips can survive in space. The company's Project Suncatcher will launch its first prototype satellite, carrying four TPUs, on October 1 aboard SpaceX's Transporter-18 rideshare mission. It is a modest experiment, roughly one ground server's worth of compute, but it is the first real-world test of an idea that has quickly moved from science fiction to boardroom slide decks: AI data centers in orbit.
What Is Launching
The satellite is called MVP, short for Minimum Viable Product. It was built by Planet, the San Francisco Earth-imaging company partnering with Google on Suncatcher, and will ride a Falcon 9 from Vandenberg Space Force Base in California.
Reported specifications include:
- Four Google TPUs, delivering about the computing power of one data-center server.
- Roughly 1 kilowatt of solar power.
- A dawn-dusk, sun-synchronous low Earth orbit, which keeps the panels in near-constant sunlight.
- A planned one-year operating trial, according to The New York Times, with the spacecraft potentially staying in orbit for up to six years.
Sundar Pichai marked the announcement by publicly asking whether Google's TPUs could work in space and promising the company would find out.
The Engineering Questions
Suncatcher is fundamentally a research mission. Google says it wants real data on three hazards that ground testing can only approximate.
Launch stress
The ride to orbit lasts about ten minutes but subjects hardware to heavy vibration and sustained acceleration up to 10 g. Individual components such as the TPU chips can experience local loads of 50 to 100 g.
Radiation
Space radiation can flip bits and degrade silicon over time. Google says its Trillium-generation TPUs came through particle-accelerator tests simulating low-Earth-orbit radiation without damage. Orbit is the real exam.
Heat
With no air to carry heat away, a satellite can only shed energy by radiating it. Google is using heat pipes and radiators, validated in a thermal vacuum chamber, and will now see how the design holds up under real orbital temperature swings.
The choice of partner matters too. Planet has launched and operated hundreds of small Earth-imaging satellites, giving Google access to a proven spacecraft platform and operations team rather than building that capability from scratch. The dawn-dusk orbit adds another advantage, keeping the solar panels almost continuously lit and avoiding repeated plunges into Earth's shadow that would stress both power and thermal systems.
The Long-Term Vision
The single MVP satellite is the first step in a much larger plan. Google intends to test laser links between two satellites in 2027, a prerequisite for making orbiting chips behave like a cluster. The full concept envisions groups of 81 satellites flying within a one-kilometre radius at about 650 km altitude, linked by optical interconnects.
The economic case rests on launch costs. Google has argued that if the price of reaching low Earth orbit falls to around $200 per kilogram, launch costs spread over a satellite's lifetime could become roughly comparable to the energy costs of a ground data center per kilowatt. Unlimited solar power and no need for land, grid connections or cooling water are the attraction.
Google is not alone. Starcloud flew an Nvidia H100 GPU to orbit in November 2025, and SpaceX has made space-based AI infrastructure a prominent part of its own pitch to investors.
Why It Matters
AI's appetite for electricity is now the binding constraint on the industry. Companies are signing multi-gigawatt power deals, restarting nuclear plants and pre-buying memory to keep up with demand. Space-based compute is one of the most radical proposed answers, and Suncatcher is the most concrete test of it by a major AI lab.
Expectations should stay grounded. Experts cited by several outlets say orbital data centers remain years from commercial viability, held back by launch costs, engineering limits and satellite manufacturing capacity. A single four-chip satellite will not change the compute market.
What the mission can do is replace speculation with measurements:
- Hardware reliability data on how accelerators age in radiation and thermal cycling.
- Thermal design validation for radiator-only cooling at data-center power densities.
- Operational lessons on running and updating AI workloads on hardware that cannot be serviced.
If the TPUs perform well over the coming year, the debate over orbital AI compute shifts from "whether" to "when and at what cost." If they do not, Google will have learned that cheaply, one satellite at a time.
