Google tests orbital AI compute with one 1-kilowatt server riding a SpaceX rocket

Google has already tested the launch half of the experiment on the ground. It says it shook a satellite on all three axes to mimic a launch and put the TPUs through a proton beam, and the chips survived ^[5]. The ride itself lasts about 10 minutes, with components taking 50 to 100 times the force of gravity ^[6]. What the flight adds is the year after: radiation and heat swings over months, and the first in-orbit trial of the heat pipes and radiators Google built to cool the chips ^[7].

Scale is easiest to compare in kilowatts. The Electric Power Research Institute estimates data centers use 4% to 5% of US electricity now and could use as much as 17% by 2030 ^[11]. Eric Schmidt, Google's former chief executive, told a House committee that data centers could need 29 more gigawatts than they do now by 2027 ^[8]. At 1 kilowatt a satellite, meeting that would take 29 million of them ^[9]. SpaceX's entire filing, built to the same size, would add 1 gigawatt, about a twenty-ninth of Schmidt's figure ^[10]. A production satellite need not be a one-server box, and the filings as reported do not state power per satellite.

Valuations and timelines for orbital compute are far apart. Musk announced in February a merger of SpaceX and xAI in a deal valued at $1.25 trillion, and said "space-based AI is obviously the only way to scale" ^[12]. Bezos has called timelines of two or three years a "little ambitious" ^[13]. Google's own analysis, published with its 2025 announcement, put the cost of an orbital data center roughly level with one on Earth by the mid-2030s ^[14]. James Manyika, Google's senior vice president for research, told The New York Times: "We don't expect, to be perfectly frank, that we'll have anything usefully operational in the next few years" ^[15].

If the TPUs run and the cooling works, Google carries the only flight data on its own chips into its next launches. Should the chips survive and the cooling fail, the eight-times solar yield Google cites for low Earth orbit ^[16] buys power the processors cannot use for long. A third outcome makes the orbital data secondary. If the FCC clears SpaceX's application and SpaceX builds at anything near the million satellites it asked for ^[3], launch rate decides the contest, and one server's test settles little.

I think the October flight is a component test of Google's own chips, and the contest it opens is over cooling, where Google is the one company of the three flying test hardware, according to Cryptopolitan ^[21]. Google's effort goes into chips and cooling, with the launch bought as a slot on a rideshare ^[1]. Schmidt bought the launch firm Relativity Space ^[18]. This view is wrong if a year of orbital data lets Google move its cost-parity estimate earlier than the mid-2030s ^[14].