Can Orbital Data Centers Solve AI’s Power Crisis?

What is the difference between a stupid idea and a brilliant idea? Sometimes it just depends on resources. Virtually unlimited funds, like unlimited push, can make even a crazy idea take off.

The same could be true for the concept of putting AI data centers into orbit. In a rare moment of unadulterated agreement, some of the richest and most powerful men in the tech world strongly support the idea. The group includes Elon Musk, Jeff Bezos, Jensen Huang, Sam Altman and Google CEO Sundar Pichai. In all likelihood, hundreds of people are now working on the concept of space data centers at the companies controlled directly or indirectly by these men: SpaceX, Starlink, Tesla, Amazon, Blue Origin, Nvidia, OpenAI and Google, among others.

The likely costs of designing, building and launching a 1 GW orbital data center, based on a network of some 4,400 satellites and including operating costs over a five-year period, would exceed $50 billion. That’s about three times the cost of a 1 GW data center on Earth, including five years of operation.John MacNeil

So how much would it cost to start training large language models in space? The best accounting is probably the one created by aerospace engineer Andrew McCalip. McCalip’s comprehensive and detailed analysis includes interactive sliders that allow you to compare space and terrestrial data center costs in the range of 1 to 100 gigawatts. Gigawatt data centers currently under construction dry landand Meta has announced plans to build a 5 GW facility, scheduled for completion sometime after 2030.

In an interview, McCalip says his first rough calculations a few years ago suggested that data centers in space would cost between 7 and 10 times more, per gigawatt of capacity, than their terrestrial counterparts. “It just wasn’t practical,” he says. “Not even close.” But when Elon Musk began publicly supporting the idea, McCalip revisited the numbers using publicly available information about Starlink and Tesla’s technologies and capabilities.

This has changed the situation considerably. The numbers in his online analysis assume an orbital network of data center satellites that borrows heavily from Musk’s tech treasure chest: “Essentially…you just start installing radiation-resistant ASIC chips on the Starlink fleet and you start increasing peak capacity organically on the Starlink fleet,” McCalip says. The network would rely on the type of watt-efficient GPU architecture used in Teslas for autonomous driving, he adds. “You start dropping them on the back of Starlinks. You can slowly build that up, and that would be about the performance you’d get.”

Ultimately, with sound but not necessarily heroic engineering, the cost of an orbital data center could be as low as three times that of a comparable terrestrial data center. This gap, although still high, at least takes the concept out of the immediately rejectable category. “I have my particular opinions, but I want the data to speak for itself,” McCalip says.

For this illustration, we have chosen a configuration with a total capacity of 1 GW. The network would include some 4,300 satellites, each equipped with a one-square-kilometer solar array generating 250 kilowatts. The data center of this satellite, powered by the bay, could have at least 175 GPUs; McCalip notes that a popular GPU rack, Nvidia’s NVL72, has 72 GPUs and requires 120 to 140 kW.

The total cost of the satellite network would be approximately US$51 billion, including launch expenses and five years of operational expenses; a comparable terrestrial system would cost around $16 billion over the same period.

Stupid? Not stupid? You decide.