The ore grades have been declining for decades deep in the Atacama Desert, where Chilean copper mines have been in operation for more than a century. The majority of the surface rock that made the nation the largest producer of copper in the world is now gone. Compared to twenty years ago, a lot more earth must be moved and processed for every ton of copper. The mines continue to operate. The employees are still present.
However, the cost of extracting the metal that comes out the other end is rising, and the demand for it is growing faster than the mining industry’s lengthy development timeframes can keep up. Copper reached a record price of $14,527 per metric ton on the London Metal Exchange in February 2026. Santiago and the Atacama were on the path to that price, but a data center tale rather than a mining story served as its catalyst.
The focus on software in the tech sector tends to mask the physical phenomenon of the AI infrastructure build-out. Up to 50,000 metric tons of copper are needed for a hyperscale AI data center, the kind that Google, Microsoft, and Amazon are racing to construct spanning several continents at once. Compared to conventional computing loads, the power distribution systems, cooling infrastructure, and cable connecting servers consume three to five times as much electricity per rack. Approximately 5,000 tons are used in a typical data center.
The businesses constructing these facilities are not constructing one, and the scale difference is an order of magnitude. They are attempting to construct dozens of these in about the same amount of time. Together, Google, Microsoft, Meta, and Amazon have declared plans for capital expenditures totaling over $725 billion in 2026 alone, a 77 percent increase from the previous year. A large amount of that money is being used to pursue the same limited supply of steel, copper, and electrical equipment.
The current pricing levels feel less speculative than they might be because the supply side of this story has been developing for a longer period of time than the AI story. A supply pipeline that is unable to react to current demand signals on any timeline that matters for the data centers being built today is a result of the structural underinvestment in new copper mine development during the 2010s, a decade when prices were low and miners reduced exploration budgets. In ideal conditions, the development of new mines takes seven to ten years.
The timeline is longer in areas with more complicated permitting, such as Southeast Asia and parts of North America. The second-largest copper facility in the world, the Grasberg mine in Indonesia, experienced a mud intrusion in 2025 that affected production. Earthquake-related events inundated Kamoa-Kakula in the Democratic Republic of the Congo. In contrast to demand, which is expanding at multiples of that rate, mine supply growth for 2026 is predicted to be about 1.4 percent, or practically unchanged.
Although there are a number of ways for investors to position themselves for this dynamic, pure-play mining businesses are the most straightforward. At present pricing levels, Freeport-McMoRan, the largest producer of copper in the world, and BHP Group are producing record cash flows due to their significant exposure to copper and other commodity categories.
Given that mining companies carry site-specific risks, labor disputes, and political exposure that are challenging to assess from the outside, the Sprott Copper Miners ETF provides diversified exposure across the mining industry without the operational risk of choosing individual companies. If the demand story keeps going in the same direction and mine supply stays as limited as it seems, the Citigroup price prediction of $15,000 per metric ton, which was released in late 2025, represents the bull case.
It’s likely that demand destruction occurs before prices hit the Citi objective because purchasers in the construction and appliance manufacturing industries have already begun substituting where they can, and extremely expensive copper lowers its use in several traditional sectors. However, the demand for AI data centers is noticeably price-inelastic; a $5 billion data center won’t be postponed by a corporation like Microsoft or Amazon just because copper costs 20% more.
Commodity cost sensitivity has little place in the economics of the AI infrastructure race. There’s a feeling that the copper market is functioning on a different timescale from the quarterly earnings cycles that usually control investor attention, and that the story is most likely not done at $14,000 when you watch the price charts and the mine development pipelines at the same time.
