SpaceX's orbital AI data centers would create a new e-waste

SpaceX's pitch for a million-satellite AI data center constellation, floated by Elon Musk, would create a category of electronic waste nobody has had to plan for before, Ars Technica reported. Data center GPUs run about five years before they're retired, which means roughly 200,000 of the proposed 1 million AI1 satellites would need decommissioning every single year once the constellation reached full scale.

SpaceX's own May 29 FCC filing lays out where those satellites go. About 40,000 a year would deorbit and burn up in the atmosphere, scattering their materials as a diffuse contaminant that settles across the globe over time, with the aluminum content adding an unknown amount of ozone depletion over the following decades. The remaining roughly 160,000 would get pushed outward into a distant disposal orbit instead. Either path removes the materials from any recycling pipeline for good.

Musk has described the satellites as built around a modified Nvidia Vera Rubin NVL72 rack, packing 72 GPUs each. Ars Technica built its estimate on the conservative assumption that each satellite carries the equivalent of 72 older A100 GPUs, whose full material composition a separate May study had already mapped across 32 elements. Excluding the heavy air-cooled heatsink, which a satellite would replace with a different cooling system entirely, that math puts the annual material loss at roughly 1,000 tons of copper, 170 kilograms of gold, close to 2 tons of silver, more than 20 tons each of bismuth and titanium, over 2 tons of palladium, and 76 kilograms of thallium.

• AI1 satellites proposed: up to 1 million
• Satellites needing decommissioning per year at full scale: roughly 200,000
• Estimated annual copper lost to orbit: about 1,000 tons
• Estimated annual palladium lost to orbit: over 2 tons, about 1% of global annual mining output
• Estimated annual gold lost to orbit: 170 kilograms

Ars Technica framed the scale by running the numbers in reverse, calculating what size asteroid you'd need to mine to replace what's lost each year. The platinum alone would take an asteroid 16 to 43 meters across. Recovering the equivalent copper would need a 140-to-190-meter asteroid, the silver and barium a 225-to-300-meter one, and the tin something closer to 530 meters, a mining operation that would have to repeat annually to break even against the constellation's losses.

SpaceX has floated one way around the problem: building the satellites somewhere the materials wouldn't need to travel from Earth in the first place.

“We intend to establish lunar-based manufacturing capabilities, including factories to produce large-scale AI compute satellites. We expect to use raw materials from the Moon to construct most of the mass of the satellites and ship chips and other lower-mass elements from Earth.”

— SpaceX, SEC filing

Turning that plan into more than a paper proposal means launching a full moonbase, mining operation, and satellite factory before a single AI1 unit gets built there, and it still leaves the question Ars Technica raises open: whether anyone will require an environmental review that accounts for the mass of material a project like this proposes to remove from Earth's material cycle permanently. The compute crunch pushing companies toward ideas this extreme is the same one reshaping how AI providers price and route ordinary compute demand here on the ground.

This piece is informational, not a recommendation to buy, sell, or hold any asset.