Google put AI chips in orbit to see whether a data center can survive up there

Google confirmed on October 1, 2026, that a prototype satellite carrying its tensor processing units reached orbit and is operating as expected, the first hardware step in a research effort called Project Suncatcher. The spacecraft was built with the imaging-satellite company Planet and rode SpaceX’s Transporter-18 rideshare, a launch that put 130 payloads into space. It is a test article, not a data center, and the question it carries is narrow: whether machine-learning chips can live through the radiation and temperature swings of spaceflight.

Travis Beals, the Google senior director who wrote the announcement, framed the logic in one sentence: “Some things can only be tested in space.” Google’s post says the team has confirmed contact with the satellite and that engineers will collect data on how the processors cope with the physical stress of launch and orbit, with results gathered over the coming weeks. The post names neither a TPU generation nor an orbit for the prototype, and it gives no mission length beyond that window.

The launch and what Google has said about the test

Spaceflight Now’s coverage places liftoff on Thursday, October 1, at 11:32 a.m. Pacific time from Space Launch Complex 4 East at Vandenberg Space Force Base in California, aboard a Falcon 9 on its 25th flight. Transporter missions fly to polar orbits and release payloads in sequence, beginning about 54 minutes into the flight, so Google’s satellite shared a ride with a methane-tracking spacecraft, pharmaceutical manufacturing labs and well over a hundred other customers. Spaceflight Now describes the Google item only as an artificial intelligence prototype and does not detail its orbit.

Google has been explicit that the project is a long-term research moonshot, and it has published a paper (peer-reviewed in the journal Joule, per Google). The company’s plan, laid out in November 2025, was a learning mission with Planet slated to launch two prototype satellites by early 2027, testing how the models and TPU hardware behave in space and validating optical links between satellites. Google’s launch-day post says the team will refine its designs with what it learns, and promises further updates as the mission unfolds.

Radiation, heat and the physics of orbital compute

Radiation is the one hazard Google has already measured on the ground. In its 2025 research write-up, engineers put Trillium, the sixth-generation v6e Cloud TPU, in a 67 MeV proton beam. High-bandwidth memory proved the most sensitive part, showing irregularities only after a cumulative dose of 2 krad(Si), nearly three times the 750 rad(Si) expected over a shielded five-year mission, and no hard failures traced to total ionizing dose appeared up to 15 krad(Si) on a single chip. Single-event effects on training runs, Google noted, still require further study.

Heat is the harder problem, and Google lists thermal management among the challenges that remain, along with on-orbit system reliability. In a vacuum there is no air or water to carry heat away, so every watt a processor consumes must eventually leave as infrared radiation from a surface, and radiator area grows in proportion to the power being dissipated. Google’s paper points toward passive heat-transport methods that move heat to radiators, and it floats more integrated designs in which solar collection, compute and cooling share a structure at gigawatt scale. None of that has flown yet.

Mass is the companion constraint. Google’s analysis argues that monolithic designs would need assembly in space by people or robots and that their structural demands “would add mass and complexity,” so the company prefers modular arrays: a conceptual cluster of 81 satellites in a one-kilometer radius at roughly 650 kilometers altitude, linked by free-space optics. Failed chips cannot be swapped by hand in orbit, so Google suggests simply provisioning redundancy, which adds still more mass per unit of useful compute.

Launch cost per kilogram and solar yield in orbit

The appeal is power. In the right orbit, Google says, a solar panel can be up to eight times more productive than on the ground and generate nearly continuously, cutting the battery mass needed. The company’s target is a dawn-dusk sun-synchronous low Earth orbit, where satellites stay in near-constant sunlight. The case for leaving the grid behind rests on that advantage outweighing the cost of lifting everything into orbit. Launch price is where the argument is decided. Datacenter Dynamics reported Google’s figures: launch costs today run between $1,500 and $2,900 per kilogram, and would need to fall to about $200 per kilogram for orbital compute to match terrestrial power costs per kilowatt.

Google’s own analysis suggests that could happen around 2035 if Starship flies frequently. That is a forecast contingent on a rocket cadence that has not yet been demonstrated, and the Datacenter Dynamics piece quotes no independent expert testing the numbers.

Sundar Pichai, Google’s chief executive, has been candid about the gap. Reporting on the November 2025 announcement quoted him saying that like any moonshot, the effort will require the company to solve many complex engineering challenges. Google also lists inter-satellite links running at tens of terabits per second, tight control of large formations, and economic feasibility among the open problems, alongside the thermal and reliability questions above. What the October 1 launch can settle is small but real: whether the chips and their support hardware survive and function in orbit as the ground tests predicted. It cannot settle cooling at scale, the economics, or whether a cluster of satellites can behave as one computer.

The remaining answers depend on later missions, and Google has given no date for anything past the early-2027 learning mission it described last year.

This article was produced with the assistance of AI and reviewed by Morning Overview editors prior to publication.

More from Morning Overview