AINEWS
Orbital hardware cannot be recycled, so SpaceX's planned AI satellite constellation would lift 1,000 tons of copper and 1% of the world's palladium and thallium into permanent loss each year.
In low Earth orbit, around 400 kilometers up where the ISS and most Starlink satellites operate, SpaceX's planned AI constellation would lift each year roughly 1% of the world's annual palladium and thallium production, alongside 1,000 tons of copper, 170 kilograms of gold, and about 2 tons of silver, according to an Ars Technica analysis of the proposed architecture. None of it would come back. That is what makes the loss a new e-waste category, and it is the category the Ars Technica analysis argues the space and AI industries need to adopt.
The analysis starts with two earlier studies, one on a GPU's chemical composition and one on asteroid mineral concentrations. Musk has described SpaceX's AI1 satellites as a modified Nvidia Vera Rubin NVL72 rack, a 72-GPU unit used in terrestrial AI data centers, packed into each spacecraft. To convert that into a material footprint, the analysis borrows a full chemical breakdown of an Nvidia A100 GPU from a May 2026 study on LLM material footprints. The A100 is a stand-in, not a match, and the Ars piece flags the substitution, but the A100 analysis is the only public per-chip chemical baseline that covers the 32 elements the calculation needs.
A single satellite, with 72 naked A100s and no cooling system, exports roughly 14 kilograms of copper, 4 grams of gold, 47 grams of silver, 500 grams of bismuth, 470 grams of palladium, and 17 grams of thallium into orbit. At a constellation-scale launch cadence, the per-year figure lands at 1,000 tons of copper, 170 kilograms of gold, about 2 tons of silver, more than 20 tons each of bismuth and titanium, 2+ tons of palladium, and 76 kilograms of thallium. The palladium and thallium totals work out to roughly 1% of global annual production of each, a fraction that is small in percentage terms and large in the context of two metals that already face tight supply.
Both metals come from concentrated supply chains, and both are widely treated as critical inputs by the industries that depend on them. Losing 1% of annual production to unrecoverable orbit each year is not, on its own, a supply crisis, but it is a supply subtraction no terrestrial industry has had to absorb before.
The analysis flags three ways the true per-year figure could be higher. The naked-A100 assumption strips away the A100's air-cooled heatsink, which accounts for 88% of the card's mass. Satellite cooling has not been defined, so it sits outside the math. AI1 satellite counts and launch cadence are also unconfirmed, so the per-year totals are extrapolations rather than reported fleet figures. The A100 baseline is a proxy for the Vera Rubin NVL72 silicon that would actually fly, and the source flags the slight mismatch.
In the orbital e-waste category, every kilogram of palladium, copper, or thallium lifted into space is lost from the recycling stream, because orbital velocities (around 7.8 kilometers per second at that altitude) make recovery uneconomic at today's prices and probably for the foreseeable future. A ground data center that retires 1,000 tons of servers sends the metals back to the smelter. A satellite that retires the same mass adds to a debris shell that no one has built a recycling supply chain for.
By the benchmark of asteroid mineral concentrations, the palladium SpaceX's AI1 plan would put in orbit each year is roughly equivalent to the palladium content of an asteroid 16 to 43 meters across, a small but not microscopic body. The cobalt the same fleet would export each year is comparable to the cobalt recoverable from an asteroid only 3 to 6 meters across. The 2023 study that supplies the comparison found that only a few elements, mostly platinum-group metals, are more concentrated in asteroids than in Earth ores. The asteroid numbers are a thought experiment, but they give the loss a physical scale: the plan trades terrestrial mining for asteroid-scale mining, except the asteroid is on Earth first.
One question sits outside the analysis: whether SpaceX's AI1 architecture, as Musk has described it, will actually fly in the form the numbers assume. If the satellites ship with a different GPU, a different cooling system, or a smaller constellation, every figure in this article shifts. SpaceX has not yet published a per-satellite material manifest, an actual launch cadence, or a satellite-count target for the AI1 architecture. Until those numbers exist, every figure in the Ars Technica calculation is conditional on the design Musk described. The direction of the material flow does not depend on the totals. It is one way.