Google to test AI chips in orbit aboard a Falcon 9

Google puts an AI rack in orbit: physics says the problem is heat

Next week, if the launch schedule holds, a SpaceX Falcon 9 will put a Google satellite with artificial intelligence processors on board into orbit. It is the first test of Project Suncatcher: finding out whether the hardware can survive the vacuum, radiation and thermal cycles of the space environment. Google plans two more satellites in 2027 to test laser links between them. The underlying promise is huge: in orbit, up to eight times more solar energy could be generated than on Earth. The problem is the usual one. Heat.

Why the vacuum doesn't cool the chips

The instinct is to think that at −270 °C the processors cool themselves. The opposite happens. In a vacuum there is no air or water to carry heat away, so the vacuum works like a thermos: whatever the chip generates stays inside. The only way out is to radiate infrared from huge panels, and the colder the radiator is, the less it radiates. The Stefan-Boltzmann law doesn't negotiate. The radiators of the International Space Station, those grey metal surfaces, give the real scale of the undertaking, and on top of that they require maintenance: one small hole in the circuit and the coolant escapes. On the ISS there are people to fix it. In an uncrewed data center, there aren't.

The real scale: one satellite, one rack

The numbers SpaceX handles for its own AI1 satellite make it clear, according to a calculation circulating in the thread. 150 kW peak and 120 kW sustained to power the equivalent of a single high-performance server rack. An NVIDIA GB300 rack consumes about 140 kW. In other words: a satellite the size of a Boeing 747, 70 meters long, for the power of one cabinet. Scaling that up to a full data center requires a fleet that doesn't exist today and maintenance that nobody has solved. In terrestrial data centers, drives in the cabinets are replaced every three years due to degradation; in orbit, that routine has no answer.

The cost of putting a kilo up there

The energy needed to place a kilogram in orbit can be as low as 0.1 MWh, according to the calculations presented in the debate. The expensive part isn't the physics, it's everything surrounding the launch: the ratio of non-energy to energy costs for most rockets is 1% or worse. That margin for reduction is the industry's serious argument. If the price per kilo falls, the equation changes. Meanwhile, the Falcon 9's success rate of 99.6% is the figure that supports the credible part of the story.

Publicity, stock prices and a precedent at sea

The skeptical reading has its own logic. A small test serves to announce a big project, and a big announcement moves stock prices. It already peine with Microsoft's underwater data center: it was built to save on cooling, people talked about it for a while, and then nothing more was heard. The difference is that here three of the biggest companies in the business are looking at the same place. It could be smoke. It could also be the first step of something that, if it works, reshapes the energy map of the industry.

The analysis gets stuck right here: the physics of heat gives no quarter, but launch costs have an enormous margin to fall. Both things are true at the same time.