How AI Satellites Could Send Thousands of Tons of Critical Materials Into Space

Elon Musk explained that these satellites would use modified versions of high-performance data-center hardware. For comparison, the NVIDIA Vera Rubin NVL72 rack contains 72 GPUs. Although the estimate below uses the slightly older NVIDIA A100, a recent study of the material footprint of large language models provided a detailed chemical analysis of the chip, covering 32 elements.

The A100’s large, air-cooled heatsink accounts for approximately 88% of its total mass. That component would likely be unnecessary in a satellite, which would require a different cooling system—one that has not yet been clearly defined. Even so, using an extremely conservative assumption that each AI satellite contains 72 bare A100 GPUs, it is possible to estimate the quantity of materials launched into space each year.

Under this simplified model, the satellites could carry approximately 1,000 tonnes of copper, 170 kilograms of gold, nearly two tonnes of silver, more than 20 tonnes each of bismuth and titanium, over two tonnes of palladium, and about 76 kilograms of thallium.

To put those figures into perspective, compare them with global annual mining and production totals. The estimated material loss would equal roughly 1% of the world’s annual palladium and thallium production—an extraordinary amount of valuable material potentially placed in orbit.

Could AI Satellites Become a Space-Based Material Waste Problem?

Another way to understand the scale is to estimate the size of an asteroid that would need to be mined to recover the same amount of material. According to a 2023 survey, only certain elements—particularly platinum-group metals—are typically found at higher concentrations in asteroids than in terrestrial ores.

Using the average composition and density of several asteroid types, including CM-group carbonaceous chondrites and the much rarer iron-rich M-type asteroids, researchers can estimate the size of celestial bodies that might contain equivalent quantities of these materials.

For example, the estimated amount of platinum would be comparable to the contents of an asteroid measuring between 16 and 43 meters in diameter. Recovering 180 kilograms of cobalt could require mining an asteroid approximately 3 to 6 meters across.

These calculations are only theoretical and depend heavily on the satellite design, GPU configuration, cooling technology, launch cadence, and recycling strategy. Nevertheless, they illustrate the potentially significant material footprint of deploying large-scale AI computing infrastructure in space.