On October 1, 2026, a refrigerator-sized satellite left Vandenberg Space Force Base on SpaceX's Transporter-18 rideshare and carried four of Google's Trillium-generation Tensor Processing Units (TPUs) into low Earth orbit.
It was the first time Google had put one of its advanced AI chips in space, and the first flight hardware for Project Suncatcher, the research effort the company disclosed in late 2025.
Built on a Planet Labs bus and nicknamed MVP, the craft is not a data center.
It is a probe of whether the idea can survive contact with vacuum, radiation, and heat that has nowhere to go.
Google later confirmed contact and said the satellite was operating as expected. Travis Beals, the executive running Suncatcher, put the point plainly: ground tests are useful, but nothing is as good as the real thing.
The bet behind that launch is that putting data centers in space is the next frontier of computing infrastructure, not a science-fair stunt.
Terrestrial AI has run into a physical wall. Training and inference now demand power plants, substations, land, and cooling water on a schedule that grids and permitting offices cannot match.
Communities are pushing back on the footprint.
This is why putting those massive hardware in space is being sold as a way around that bottleneck: continuous sunlight with no clouds, no night, and no atmosphere in the way, so the same panel area can collect several times the energy it would on the ground, sometimes cited at up to eight times, and can do it without waiting years for a new gas plant or transmission line.
There is no local grid to overload and no neighborhood to persuade.
Heat from the chips can, in principle, be thrown into the cold of space by radiators rather than dumped into rivers.
A compute node up there is also isolated from earthquakes, floods, and a certain class of geopolitical interruption. The electricity is not beamed home. It is used where it is collected, to run models.
That is the attractive version.
The flight itself shows how far the attractive version still is from a rack.
The satellite's solar arrays are sized around a kilowatt, about a microwave, and the TPUs are allowed to run only in bursts of roughly 15 minutes so the power and thermal systems are not overwhelmed. After that they shut down and wait for the radiators to catch up.
Space is cold, but it is also an insulator.
Without air, heat leaves only by radiation, which means surface area, mass, and a duty cycle that would be unacceptable in a cloud region.
And speaking of radiation, radiation requires extra shielding, and this attempt changed the error profile enough that Google retested the chips in a particle accelerator.
The error rate Beals described for ordinary inference is on the order of one in a million, tolerable for a five-year satellite life if the job is answering queries. It is already a problem for a mega-scale training run with thousands of chips working for months.
Google's near-term plan is a pair of purpose-built satellites next year, linked by laser, then eventually a formation of about eighty-one craft flying close enough that bandwidth and latency between chips can support a multi-rack workload.
Beals has said the interesting question is not today's jobs but the jobs five years out, when the link between processors matters as much as the processors.
The cost case is a launch-rate case.
In a paper headed for the journal Joule, Google walks a learning curve from Falcon 1 forward and argues that prices near $200 a kilogram by 2035 would require Starship to put something like 370,000 tons into orbit.
At an assumed 200 tons a flight, that is about 1,800 launches over ten years, or 180 a year.
Starship has never flown more than five times in a year. Elon Musk has talked about hourly flights by 2029. Both numbers can be true as statements and still leave a canyon between them.
Google is a major SpaceX investor and, for now, a customer. Suncatcher does not get off the ground as a business unless that cadence becomes ordinary.
The October test is aimed at inference-class work, including short runs of models in the Gemini and Gemma family, and at finding the failure points before anyone draws a constellation.
Debris, station-keeping in a tight formation, and laser links that have to hold while both ends move at orbital speed sit on the same list.
So the frontier is real, and it is still mostly a frontier. The energy and land argument is not marketing fog.
AI's hunger for electricity and real estate is the constraint that made orbital compute worth a rideshare slot.
What flew on October 1 is a kilowatt, 15 minutes at a time, 4 chips, and a paper that says the economic version needs a launch cadence the industry has never demonstrated.
Google calls Suncatcher a long-term moonshot, and that is the accurate name for it. The next step is not a data center in the sky. It is whether those chips still behave after the radiators, the radiation, and the next two satellites have had their say.
