Heat limits Google's four orbiting TPUs to 15-minute runs

Space makes it hard to shed heat with water or fans, so Google's researchers built the satellite around heat pipes and radiators, with solar panels for power ^[5]. Even with that system, NPR reported, the chips will run for 15 minutes and then shut down to cool off ^[6]. NPR's account does not say how long each rest lasts, so the share of each hour the four chips spend computing cannot yet be calculated.

On the ground, data centres run into limits on power, water and land ^[20]. In orbit the added cost is mass. ARK Invest research puts SpaceX's launch cost at about $1,000 a kilogram after a 95% cut since 2008 ^[11], which puts the 2008 price at about $20,000 ^[16]. Jeff Thornburg, chief executive of Portal Space Systems, told Fortune that the large, heavy solar arrays an orbital data centre would need are expensive to transport ^[10]. A larger radiator to keep chips running past 15 minutes would be paid for at the same rate per kilogram ^[11].

The 15 minutes can be read in more than one way. It may be a property of a refrigerator-sized test craft ^[1] built to study launch vibration and what Google called the "radiation and thermal extremes of space" ^[3]. If so, a production design with more radiator lifts the cap. It may instead be an early look at the real trade-off: the radiator needed to run continuously weighs enough that launch price sets the cost of orbital compute. Or heat may matter less than failures. Google said ground tests showed its chips "can survive a radiation total ionizing dose greater than what they would receive during a five-year space mission," while acknowledging that solar events and cosmic rays pose a risk ^[12]. Repairing equipment is also much harder in space ^[14].

We think the second reading is the one Google's year of experiments in low Earth orbit ^[4] has to settle, because it decides price. Elon Musk said at the World Economic Forum in January: "The lowest-cost place to put AI will be space, and that will be true within two years, maybe three, three at the latest" ^[18]. If Google's experiment runs its full year from the Oct. 1 launch, it ends around October 2027. That is about three months before the early end of Musk's window and some 15 months before the late end ^[19]. SpaceX has filed with the FCC for up to 1 million data-center satellites ^[7], and Blue Origin has filed for 51,600 ^[8], about one-nineteenth as many ^[17].

The strongest case against our view is Starcloud, which ran Google's Gemma model on a single Nvidia H100 in orbit last November ^[13]. If useful jobs fit inside short windows, the cooling cycle matters less than we think. We would also be wrong if Google reports continuous operation, or a radiator light enough that the $1,000-a-kilogram launch price ^[11] stops being the larger cost.

Meanwhile, according to Fortune, companies are still putting billions of dollars into data centres on the ground ^[15]. Google's own orbital programme, started in November 2025 ^[9], has one satellite carrying four chips ^[1].