Google's Project Suncatcher puts AI chips in orbit

Google is putting its AI hardware on a rocket. The company has confirmed that Project Suncatcher, the research moonshot it announced last year, will fly a prototype satellite on SpaceX's Transporter-18 rideshare mission, developed in partnership with Planet, to find out whether its Tensor Processing Units can survive space. The post is careful about the framing: this is a survival test first and a compute test second.

Engineering problems before computing ones

The flight is designed to gather in-orbit data on how TPUs handle the physical stress of launch and the radiation and thermal extremes outside the atmosphere. Vibration testing shook the satellite on all three axes to mimic launch frequencies, and the chips themselves are expected to see loads of 50 to 100 times the force of gravity, well above the roughly 10 g a rocket ride delivers over about ten minutes. Radiation is the harder problem to rehearse: Google ran its TPUs through a proton beam facility at UC Davis's Crocker Nuclear Laboratory while they worked through AI workloads, watching for bit flips and their effect on the output.

No airflow, so radiators instead

Cooling is the constraint that most data-centre thinking simply does not prepare you for. A TPU produces a lot of heat in a very small area, and in a vacuum there is no air to carry it away, so the heat has to leave through radiators. Google says it is working on a combination of heat pipes and radiators for the chips, a design it has already exercised in a thermal vacuum chamber that simulates both the temperature and the vacuum of orbit. The company's own summary of the engineering gaps is unusually blunt: some of this can only be tested in space.

The sunlight is the whole point

The reason to bother is power. In low Earth orbit a satellite sits in near-constant sunlight, which Google calculates could deliver up to eight times the solar power available from the same panel on the ground, and that is the input a terrestrial data centre normally has to buy and cool. The long-term plan is clusters of satellites linked by lasers rather than fibre, and that is the part worth being sceptical about: the links have to carry very high bandwidth across very short distances between two moving points, which Google compares to hitting a coin-sized target from miles away while both ends move. The next milestone is scheduled for 2027.

Our opinion

Every orbital data-centre pitch eventually hits the same wall, which is that the accelerator is rarely the difficult part. Getting heat out of a sealed box and getting answers back to Earth are the problems that decide whether a constellation is an infrastructure or a demo, and Google's post spends most of its word count on exactly those two. That is a better sign than a benchmark chart. The honest reading of this launch is that it is an experiment with a real chance of failing in a publicly instructive way, and the 2027 laser milestone is the number to watch. Solar power in orbit is genuinely cheap and abundant; everything else about running AI up there is still expensive and unproven.