Google's orbital compute paper implies about 1,800 Starship launches in a decade

A server hall uses air to carry heat off its chips. In orbit there is no air, so conventional cooling systems do not work, and any alternative adds weight that pushes up launch cost, Fast Company reported ^[5]. Google's experimental system lets the TPU work for around 15 minutes before it has to cool down ^[5]. "We've done testing on the ground, but you know, there's no test that's completely as good as the real thing," said Travis Beals, who manages Project Suncatcher at Google ^[6].

Google's stated goal is a network of 81 satellites flying in close formation and processing in parallel ^[7]. What went up is one TPU on a standard Planet Labs satellite ^[2]. Juan A. Fraire, a researcher at France's National Institute for Research in Digital Science and Technology, told Fast Company that laser connections are one of the biggest problems for any space data center ^[9]. Google has shown 800 Gbps links between optical transceivers in a lab, but reaching those speeds in orbit means flying satellites extraordinarily close together ^[9]. Holding that spacing takes maneuvers that burn propellant, and the propellant cannot be replenished ^[9].

The launch figures come from a peer-reviewed white paper Google released Thursday, due to be published in Joule. Its authors stress it is not an economic feasibility study ^[10]. To stay on SpaceX's price curve, they argue, Starship would have to carry 370,000 tons of payload to orbit ^[12]. TechCrunch puts that at about 1,800 launches over ten years, or 180 a year, if each flight carries 200 metric tons ^[13]. Divided out, 370,000 tons at 200 tons a flight is 1,850 launches ^[1]. TechCrunch's headline gives 1,600 ^[14], about 250 fewer than its own figures produce ^[3]. The 180-a-year pace is 36 times Starship's best year so far ^[2]. Elon Musk has suggested Starship could fly hourly in 2029 ^[16].

Beals drew the workload line himself. "The error rate is very low if you're thinking about typical inference operations, right? Like one in a million," Beals said. "On the other hand, it was already problematic for doing, say, some mega-scale training run where you're going to have many thousands of chips running for months." ^[19]

Lifespan is a separate test. Fraire's research suggests an orbital data center would need to run about five to six years to match the emissions of greener terrestrial facilities ^[20]. Google expects its satellites to last five years ^[18], the bottom of that range ^[4]. "I think we need to think before we act," Fraire told Fast Company. "And so far, we see the industry not thinking so much, but just acting." ^[21]

For an operator deciding where orbital compute could ever fit, two tests sort the candidates. One is whether a job can stop when the TPU's 15-minute burst ends without losing work ^[4]. The other is whether it tolerates an error rate near one in a million ^[19]. Short inference jobs pass both ^[18]. A months-long run across thousands of chips fails both ^[19].

On cost, the forcing question for anyone quoting orbital capacity is the launch price per kilogram behind the quote and the yearly flight rate that price assumes. Google's own path to about $200 a kilogram by 2035 needs roughly 180 Starship flights a year ^[11][13]. I'd keep orbital compute in a research budget until some rocket is flying near that rate. The tradeoff is timing. Matthew Weinzierl, author of the space-economics book Space to Grow, told Fast Company: "The costs of data centers on Earth are rising, while the costs of data centers in space will fall, and at some point those curves will cross." He went on: "We don't know when, but the big investments by Google, SpaceX and the like are intended to bring that day closer." ^[22]