Material Needs of Artificial Intelligence Eclipsed by Energy Debates

President Trump attended an important summit on Artificial Intelligence (AI) and concomitant energy infrastructure at Carnegie Mellon University in Pittsburgh on July 15th, 2025. Major investments were announced by Google and Blackstone capital to build data centers and hydroelectric dams in Pennsylvania to power the AI economy for the region. Missing from the conversations was the material needs for AI technologies that could be just as serious resource constraint to upscaling the use of these technologies. While there is much talk of critical minerals for defense and for clean energy infrastructure, the material needs of AI are not as prominently discussed. Concerns about energy intensity dominate debates about AI. Carbon emissions estimates have also been researched by Google researchers using Life Cycle Analysis techniques, but the materiality of AI infrastructure has not been well-researched.

In her prominent book Atlas of AI, renowned Microsoft researcher and academic Kate Crawford documents the vast extractive needs of AI but notes that detailed analysis of materials that will be needed has been sparse. Part of the challenge is the secrecy around AI hardware material needs. At a recent colloquium on critical materials for AI hosted by Professor Alondra Nelson at the Institute for Advanced Study in Princeton New Jersey, researchers lamented that confidentiality concerns often prevented forensic accounting of material needs for AI. Where estimates are available, they are often focused on the material needs for the electricity infrastructure needed for data centers. For example, the Wall Street Journal did a story earlier this year on the copper metal needs for AI but focused on the energy infrastructure needs.

The article cited JP Morgan forecasts which suggested that the copper needed for AI energy supply would require another 2.6 million tons of copper adding to the projected 4-million-ton projected metal deficit by 2030. There was a reference to a Bank of America study in the article which differentiated the material needs of the data centers themselves at around 200,000 metric tons a year compared with 500,000 tons annually for energy infrastructure. Yet these are only estimates for one metal. Gallium has gained some interest in recent years because of its high-performance potential in AI chips but systematic estimates of upscaling supply are limited. Journalists and academics alike also conflate material needs due to a lack of understanding of the chemistry of these technologies. For example, lithium is largely needed for batteries that are not necessarily an AI infrastructure issue but can get conflated in discussions on any novel technological entity.

Indium and arsenic will also be needed for refined chip technologies but there are no clear estimates of projected demand. Similarly, germanium demand is expected to increase with AI infrastructure but only rough estimates of 60% growth by 2034 are available via consulting firms with no peer-reviewed research on demand growth linked to particular targets for AI penetration in various technologies. High purity alumina is another key material for AI technologies which also presents important opportunities for innovation in deriving the material from a range of existing material stocks. Australia is going to be a key provider of this material and has recently started construction of the world’s largest factory for its production in Gladstone, Queensland.

Quantum computing may also take on many roles of conventional AI processors in coming years. While many of the metals needed for this infrastructure may be similar, there are some notable additions of materials such as boron and ytterbium which are also specifically more well-suited for quantum computing technologies. Superconductors with a range of exotic material needs are essential for quantum computing. These materials also operate only at lower temperature ranges and hence additional infrastructure for cooling would be needed above and beyond what is already needed for data centers.

As momentum builds towards international governance of AI and the findings of the U.N. Secretary General’s High Level Advisory Board on Artificial Intelligence get implemented, a sharp focus on material forecasts is needed. Scenarios for mineral demand which are linked to specific upscaling targets of countries for AI infrastructure should be developed. Based on those scenarios, a prioritization of those tasks for which AI has most societal benefit should be developed. In some cases, AI could itself assist with material efficiency. In coming years researchers have a highly consequential area for inquiry set before them on figuring out optimal material usage profile for AI technologies and how they might transform both our physical and social reality.