The conversation about artificial intelligence tends to centre on software: the models, the training runs, the inference costs, the companies building the interfaces. The physical infrastructure underneath that conversation receives less attention. It probably should receive more, because the physical layer has a constraint that the software layer does not.
That constraint is copper.
What AI actually runs on
A large AI data centre is a copper-intensive structure. Transformers, busbars, cooling systems, power distribution units, server interconnects: copper is the conductive material in each of them. There is no cost-effective substitute at the voltages and current densities data centres require. Aluminium conducts, but it requires larger cross-sections and creates connection reliability problems in high-vibration environments. Copper is what the industry builds with.
J.P. Morgan estimates that a single large AI data centre can require up to 50,000 tonnes of copper. S&P Global projects that data centre copper demand alone will reach 475,000 tonnes annually by 2026. That is roughly 2% of total global copper production, going to one narrow application that barely existed a decade ago.
This is before the grid that powers the data centres is included. Expanding electricity transmission and distribution infrastructure requires copper in quantities that dwarf any single building. The electrification of transport adds a further layer. An electric vehicle uses roughly three times the copper of a conventional car. At scale, the transition from internal combustion engines to electric drivetrains represents a structural step-change in demand that runs alongside the AI buildout rather than instead of it.
S&P Global’s most comprehensive recent analysis projects global copper demand rising from approximately 28 million tonnes in 2025 to 42 million tonnes by 2040. That is a 50% increase over fifteen years. The demand side of the equation is well-understood and, at this point, largely non-discretionary. Governments have committed to grid expansion. Hyperscalers have committed to data centre buildout. Vehicle manufacturers have committed to electrification timelines. The copper demand is locked in by capital expenditure plans that are already underway.
Where the supply comes from, and why it cannot respond quickly
The supply side cannot adjust as fast as the demand side can grow.
The primary constraint is time. A copper mine takes between 15 and 17 years from discovery to production. That timeline is not primarily a function of engineering; it is a function of permitting, environmental assessment, indigenous land consultation, geological appraisal, and financing. None of those processes compress quickly, and regulatory complexity has been increasing rather than decreasing across the main copper-producing regions.
The discovery pipeline has also weakened. Only 5% of the world’s major copper deposits have been found in the last decade, against higher exploration spending in prior decades. Ore grades at existing mines are declining, meaning more rock must be processed to extract the same amount of copper. The mines that exist are getting harder to run profitably at lower copper prices, which discourages the investment needed to maintain capacity.
The resulting shortfall is measurable now. UBS forecasts a refined copper deficit of 230,000 tonnes in 2025, widening to over 400,000 tonnes in 2026. Wood Mackenzie’s estimates run slightly higher. The International Copper Study Group projects a 150,000-tonne shortfall for 2026. The specific numbers vary, but the direction is consistent across forecasters: demand is outpacing supply, the gap is widening, and the structural solution (new mines) operates on a 15-year lag.
What this means for manufactured goods
Copper is not a speculative asset. It is a production input. When its price rises, the cost of everything made with it rises alongside. Energy storage systems, grid-tied inverters, electric motors, wiring, transformers, heat exchangers: all of these become more expensive as copper becomes more expensive.
This creates a straightforward dynamic for the products in the Finite Resources catalogue. The energy storage systems we cover – portable batteries, home battery systems, solar array components – are copper-intensive manufactured goods. The same is true of the electronics and computing hardware covered elsewhere on the site. These products are currently priced as if copper costs what copper costs today. They will not be priced that way when the supply deficit widens further and input costs are passed through to finished goods.
The thesis of this publication is that physical things on the bounded-supply side of the current technological transition will cost more, not less, in five years. Copper is perhaps the clearest single illustration of that thesis. The AI boom that is driving the demand curve cannot be stopped without stopping the AI boom. The supply curve cannot be accelerated without rewinding fifteen years of permitting timelines.
The gap is real. The products that depend on it for their materials are priced today as if the gap will close.
The macro frame that Lyn Alden, Doomberg, and Luke Gromen share is the theoretical foundation. Copper is the practicum.
Finite Resources earns a commission on purchases made through links on this site. This does not affect our editorial independence or the prices you pay.