Google Launches First Orbital TPU Satellite, Testing Feasibility of Space-Based AI Computing
nashnova research
Google plans to send an experimental satellite carrying four of its own TPU chips into low Earth orbit in October 2026, running Gemini in space for the first time — a proof-of-concept for moving AI compute off the ground, with a long-term vision of orbital data centers.
What is this satellite actually supposed to do?
Google and Planet co-developed the satellite, which will ride SpaceX's Transporter-18 rideshare mission. It carries four Trillium TPUs — Google's custom AI accelerator chips — and will run Google's Gemini model in orbit while receiving commands from the ground.
The satellite is roughly the size of a household refrigerator. Its solar panels deliver about 1 kilowatt of power, and its compute is roughly equal to one ground-based server. Design life: one year, then atmospheric re-entry and burn-up.
This means → this is not a commercial deployment. It is an engineering validation: can the chips survive space and run AI inference? That is the core question.
What are the biggest threats to chips in space?
Launch shock: the satellite endures about 10 g during ascent; internal chip components may experience 50 to 100 g of shock loading. In plain terms = during the few minutes the rocket hurls you upward, the chips must survive being violently shaken.
Space radiation: beyond the atmosphere, cosmic rays penetrate electronics and cause compute errors. Google tested the Trillium TPU at UC's Crocker Nuclear Laboratory; results showed the chip can withstand radiation and recover from errors via a system reboot.
Thermal management (the hardest unsolved problem): in vacuum there is no air convection. Heat can only leave the chip through conduction and radiation. Google designed a passive heat-path structure, but the satellite can run at full power for only about 15 minutes before it must shut down and cool off. This reflects how far space-based compute still is from "always on."
How ambitious is the long-term goal?
Google's long-range vision: deploy 81 satellites flying in formation at roughly 650 km altitude, sharing compute across ultra-high-bandwidth inter-satellite laser links within a 1 km radius — an orbital data center.
The hard part is laser communication. Traditional inter-satellite lasers transmit data over long distances. This scheme requires precise alignment at just a few kilometers, while satellites move at high relative speeds. Google's own analogy: "hitting a moving coin while driving at full speed."
This means → the 81 satellites are not computing independently. They must work together like servers in a single rack — communication precision is the make-or-break factor.
Why not just keep building data centers on the ground?
The core driver is AI infrastructure's energy pressure. Low Earth orbit theoretically receives about 8× the solar energy available on the ground, potentially easing the power, land, and grid-capacity constraints facing terrestrial data centers.
The project, dubbed "Project Suncatcher," was first proposed by Google VP Blaise Agüera y Arcas, and later backed by CEO Sundar Pichai and co-founder Sergey Brin.
Google estimates that as rocket-launch costs keep falling, total operating costs for a space data center could approach ground-facility levels by the mid-2030s. In plain terms = it is far more expensive today, but if rockets get cheap enough and the thermal problem is solved, costs could converge within a decade.
Where does this stand right now?
The first in-orbit test launches on October 1, moving Project Suncatcher into its orbital-validation phase.
The current bottleneck is clear: whether the thermal problem can be solved at commercial scale is the key gate for the entire roadmap.
This means → radiation hardness has passed initial tests, launch shock can be engineered around, but "running AI continuously in space" is stuck on the most basic physics question — how to get rid of the heat.
市场有风险，内容仅供研究参考，不构成投资建议。