In almost every global conversation about artificial intelligence, we hear the same concepts: algorithms, models, chips, silicon, compute. We talk about parameters, scaling laws, GPUs, latency, hallucinations, benchmarks. But there’s something essential that’s almost never mentioned: AI is not just software. AI is matter.
Behind every model we train, behind every chip we manufacture, and behind every ever-larger data center that goes up, there is a gigantic physical infrastructure: copper extracted from mines, lithium pumped from salt flats, rare earths refined through complex industrial chains, gallium produced in highly concentrated supply networks, millions of liters of water, and enormous amounts of energy.
AI does not float in the cloud. AI has weight. AI takes up space. AI consumes critical minerals. It is, in essence, a material system that depends on territories, ecosystems, and supply chains that are already under heavy strain.
This creates the central paradox of our time: the most digital revolution in history is also one of the most physical. The data economy rests on the minerals economy. The future of computing depends on the future of extraction. Artificial intelligence literally needs more mineral intelligence.
That is the conversation missing from the global debate on AI, and it’s the conversation we should open today to better prepare ourselves. Intelligence is anticipation.
1. The material base of AI: critical minerals, energy, water, and land
The computational demand of frontier models is growing exponentially: training compute has doubled every five to six months since 2010 — a pace far faster than Moore’s Law once was for computing. Each doubling requires more copper, lithium, rare earths, and gallium.
Copper. AI data centers require between 27 and 33 tons of copper per megawatt, rising to as much as 47 tons per megawatt in intensive training clusters. Including transmission lines and substations, total demand reaches between 100 and 150 tons per megawatt.
Gallium. High-performance chips increasingly depend on gallium nitride. Data center demand could exceed ten percent of global supply by 2030.
Rare earths. Cooling systems use neodymium-iron-boron magnets. That demand will grow by around 7.5% annually through 2040.
Lithium. Energy storage associated with AI could reach around 300 gigawatt-hours by 2030.
Water and land. AI data centers consumed close to one trillion liters of water in 2025 — nearly 1 cubic kilometer — roughly 10 times more than Escondida, the world’s largest copper mine, consumes.
Training GPT-3 in Microsoft’s U.S. data centers consumed around 700,000 liters; larger models, such as GPT-4, consume even more. Today’s “hyperscale campuses” (the largest data centers) span anywhere from dozens to thousands of hectares: Google’s in Maryland covers around 364 hectares, while Meta’s in Louisiana already exceeds 1,480. In built floor area, the Stargate campus in Texas is approaching 372,000 square meters.
2. A relationship that can be symbiotic
AI and mining also complement each other. The same AI that demands so much mining can also help strengthen it, enabling faster, more efficient exploration and greater operational productivity and safety — for example, by combining autonomy with digital twins, as Escondida in Chile (operated by BHP) and Quellaveco in Peru (operated by Anglo American) are already doing.
When it comes to AI-assisted exploration, the most striking case is Mingomba, in Zambia: mining company KoBold went from discovery to development — the time between discovering a copper deposit and bringing it into production — in just five years, compared with a global average of nearly 18. AI cut more than a decade off that cycle.
This isn’t an isolated case. In general, the most advanced predictive models reduce the exploration area needed by up to 80%, without sacrificing — and sometimes even improving — the discovery rate. They process decades of historical data in weeks, versus months for traditional analysis, and cut exploration capital spending by 40% to 80% through optimized drilling.
If mining is, in part, the material base of AI, then AI can in turn be the lever that accelerates the supply of those very minerals that AI itself — and the energy transition — need. That is the symbiotic relationship: each needs the other, and each can help the other grow
But they can also be financially competitive, since both are highly capital-intensive activities that draw on the same sovereign wealth funds, infrastructure vehicles, and debt markets. AI requires USD 2–5 billion for each major data center, with training cycles renewed every few years, while mining needs USD 2–10 billion to develop long-lived physical assets, with complex energy, water, and logistics infrastructure.
This similarity in scale means both industries draw on the same sovereign wealth funds, infrastructure vehicles, and debt markets — though AI attracts fast-return capital while mining attracts long-term strategic capital. Even so, AI depends on critical minerals such as copper, lithium, nickel, and rare earths, so although the two compete for financing, they end up being complementary sectors within the global economy.
3. The falling cost of open models: opportunity and risk in a two-speed mining industry
There’s another data point that’s changing the game: open-source AI models are, on average, only about four months behind closed frontier models — and are enormously cheaper. That gap, which was nearly twelve months just a year ago, keeps narrowing.
This is a huge opportunity for junior mining companies. Firms that could never afford the full proprietary stack of a major tech provider can now access AI-assisted exploration, blast optimization, mine planning, predictive maintenance, and autonomous systems at a fraction of the cost of just a few years ago. It is potentially the biggest productivity leap these smaller miners can seize to close the gap with the majors.
But that drop in cost doesn’t come without risk. Artisanal and small-scale mining — which employs between 40 and 45 million people worldwide — risks falling even further behind if it doesn’t gain access to these same tools. Today, adoption of autonomous mining is concentrated in large projects, while the rest of the sector continues to face limited access to technology, digital infrastructure, and training opportunities. Without deliberate intervention, AI will not help close the sector’s productivity gap — it will tend to widen it instead.
Conclusion: the missing public policy
AI and mining are intertwined in the same material and technological cycle. AI needs minerals, and mining needs AI to explore, operate, and compete. And both need water, energy, and land. That relationship can be symbiotic — but it won’t be by default.
An explicit public policy is needed, with at least three tasks: assessing how the two industries affect the same territory (competition, or positive synergies based on enabling infrastructure if coordinated); monitoring how AI adoption is distributed across the sector, to help close the gap facing junior and small and medium-sized mining companies; and facilitating the technological leap that open, more affordable models already make possible — before the window of opportunity closes.
Without such policies, AI and mining could end up competing for the same land, and the symbiosis between them could stall — even turning, in practice, into a two-speed race, where the most advanced players already walk with a flashlight while the rest try to move forward in the dark.
The question is not whether AI will transform mining — because it already is — but rather how to help that relationship become synergistic, so that its symbiotic transformation reaches everyone, not just a few.
This article was written by Eduardo Barrera, SFAI Global Consultant.
