Elon Musk’s talk about maintaining a million-strong AI data center satellite megaconstellation may not exactly be practical or economical, but it might be unique. It’s about the closest we’ve come to confronting a scheme that would export a considerable amount of valuable materials into space. Humans aren’t doing a great job of material sustainability, but normally we’re talking about stuff escaping a recycling pipeline rather than escaping Earth’s gravitational pull.

The commercial space sector likes talking about the allure of mining asteroids for precious metals to bring back to Earth. Under what circumstances are we going to be willing to do that in reverse? Starlink alone has doubled the mass of objects in low-Earth orbit, and this orbital data center constellation would dwarf that—and dispose of at least some satellites by pushing them away from Earth.

Given the roughly five-year expected lifetime for data center GPUs, about 200,000 of the 1 million proposed SpaceX AI1 satellites would be decommissioned each year. Based on their May 29 FCC filing, about 40,000 would definitely deorbit and burn up in the atmosphere. (Those materials would largely be dispersed throughout the atmosphere, turning a resource into a diffuse contaminant that slowly settles over the globe. One related issue: the aluminum would cause an unknown amount of ozone depletion over a period of decades.) Some or all of the remaining 160,000 satellites would be moved outward into a distant “disposal” orbit, instead. Either way, they’re lost from a “material life cycle” point of view.

Without full, detailed specifications for these satellites, there’s no way to properly tally the amount of material we’re talking about. Focusing on just the GPUs themselves—ignoring solar panels, cooling systems, and the rest of the server and networking devices—can at least provide a starting point.

Musk has described these satellites as using a modified Nvidia Vera Rubin NVL72 rack, which contains 72 GPUs. Though it references a slightly older card, a May study on the material footprint of LLMs provided a full chemical analysis of an A100, covering 32 elements. The massive air-cooled heatsink on that card accounted for 88 percent of its mass, which we’ll simply have to exclude, since the satellite will obviously require another type of cooling that has not been defined.

But using the extremely conservative assumption that each AI1 satellite was simply composed of 72 naked A100 GPUs taped together, we can estimate the material exported to space (or vaporized so thoroughly that it might as well have been) each year.

That includes 1,000 tons of copper, 170 kilograms of gold, almost 2 tons of silver, over 20 tons each of bismuth and titanium, over 2 tons of palladium, and 76 kilograms of thallium.

Some of these elements are, unsurprisingly, rounding errors compared to the amount we mine each year. But that’s around 1 percent of global annual palladium and thallium—a remarkable amount to eject into space.

Earth’s neighborhood inconvenience store

Another way to think about this is to calculate the size of asteroid you would have to mine to recover the amounts of these elements being lost. As a 2023 study notes, there are only a few elements that can be found at a higher concentration in asteroids compared to ores on Earth, like the platinum group metals.

Using average chemistry and densities for a couple different types of asteroids—common CM-group carbonaceous chondrites and rarer iron-rich M-type asteroids—some of the elements lost could be found in modestly sized bodies. The platinum, for example, equates to the contents of an asteroid 16 to 43 meters in diameter. The 180 kilograms of cobalt could be recovered from an asteroid about 3 to 6 meters across.