Google will launch its first Project Suncatcher satellite on October 1, sending four Trillium Tensor Processing Units into low Earth orbit aboard SpaceX’s Transporter-18 rideshare mission. The refrigerator-sized spacecraft, built with satellite operator Planet, is the first hardware test for Google’s plan to run AI compute in space.
The company announced the launch date in a September 24 update titled “Behind Project Suncatcher, our moonshot to put AI in space.” The mission has one question to answer: whether Google’s AI hardware can survive and operate in orbit. Radiation, launch stress and vacuum are the immediate tests. The longer-term question is whether orbital data centers make economic sense at all.
Why put a data center in space
The pitch is power. In low Earth orbit, a satellite sees sunlight almost continuously, and Google says solar panels there can generate up to eight times as much energy per year as comparable panels on Earth. Terrestrial data centers face grid queues, permitting fights and cooling limits as AI demand grows. Space offers unlimited solar with no utility bill, at the cost of everything that makes orbit hostile: radiation, thermal swings and the difficulty of repair.
Project Suncatcher, announced in November 2025, envisions constellations of solar-powered satellites equipped with Google TPUs and linked through high-speed optical connections. The original plan called for two prototype satellites with Planet in early 2027. The October 1 mission moves the first hardware test forward by several months by mounting the TPUs on an existing satellite platform instead of waiting for custom hardware.
The choice of platform matters for the timeline. Planet’s satellite bus is flight-proven and available on commercial rideshare schedules, which meant Google could skip years of spacecraft development and put its chips on a vehicle that already knows how to operate in orbit. The tradeoff is that the test satellite, named MVP, was not designed around thermal management for dense compute hardware, so some of what Google learns will be about the gaps between a standard bus and a purpose-built compute platform.
What is flying
The satellite carries four Trillium-generation TPUs, the same chip family that powers Google’s current AI infrastructure on the ground. Before launch, the chips went through ground-based radiation testing and withstood doses of 2 to 15 kilorads without failures, exposure levels that exceed what the hardware is expected to see over a five-year mission. Thermal management is the harder problem in vacuum, where there is no air to carry heat away, and Google says it is testing cooling approaches combining heat pipes and radiators.
The mission also plans to test laser crosslinks, the optical links that would let a future constellation of compute satellites talk to each other at high bandwidth. That capability is central to the idea of a distributed orbital data center, where individual satellites function as nodes in a cluster rather than standalone machines. Google’s own consumer products already use laser links between satellites, but linking TPUs running coordinated workloads is a different requirement, since distributed training jobs are sensitive to link latency and drops in ways video relay is not.
Power is the other constraint on the test. Futurum Research notes the Trillium TPUs on this flight will run on roughly 1 kilowatt of solar energy, a fraction of what a single ground-based TPU pod draws. The October mission proves survivability, not scale. Scaling to useful compute levels would require satellite designs with far larger solar arrays and power systems than anything Planet flies today.
The race it joins
Google is not alone in looking up. SpaceX has outlined plans for orbital AI data centers and has filed with the Federal Communications Commission for a satellite network that could eventually reach enormous scale, with an initial AI1 satellite architecture discussed at roughly 120 kilowatts of sustained compute. Google’s first experiment is far smaller in ambition, a learning mission rather than a production deployment.
Futurum Research estimates orbital compute could address a $1 trillion market by 2030 if launch costs keep falling, a big if that depends on reusable rocket economics continuing to improve. The October 1 launch rides a Falcon 9 that has already flown multiple times, on a rideshare slot that costs a fraction of a dedicated mission, which is exactly the cost curve the orbital compute thesis relies on.
Google’s next major Suncatcher milestone remains planned for 2027, when two prototype satellites built for the mission are scheduled to launch and test the hardware in orbit under dedicated designs. Engineers will use data from the October flight to inform that build, particularly around thermal control and radiation effects on the TPU packages.
The hard problems ahead
Surviving launch and radiation is the first hurdle, and the ground tests suggest Google expects to clear it. The harder issues come after. Radiation-induced bit flips in memory over a multi-year mission, thermal cycling as the satellite moves in and out of sunlight every 90 minutes, and the sheer difficulty of servicing hardware that cannot be opened and repaired. Terrestrial data centers replace failed boards weekly. In orbit, a failed TPU is a failed satellite.
Bandwidth back to Earth is another constraint. An orbital cluster can compute all it wants, but results still have to come down through laser links to ground stations, and the total data volume a constellation can return is bounded by physics and spectrum licensing. Workloads that never need to leave orbit, such as processing satellite imagery in place, fit the model best. Serving cloud customers on the ground is a much taller order, because moving petabytes down a laser link is nothing like moving them across a fiber network.
There is also the regulatory layer. Constellations of hundreds or thousands of compute satellites would need spectrum allocations, debris mitigation plans and launch approvals that no space agency has processed at that scale. Google’s moonshot framing glosses over how much of the path to orbital data centers runs through filing cabinets rather than engineering labs.
The October 1 launch will not settle any of these questions. It is a single satellite with four chips, flying to answer whether the hardware works at all. But it moves the orbital data center idea from whitepapers to a hardware test with a launch date, and it puts Google and SpaceX in the same race with very different starting points.