Copper’s Record High Isn’t About AI Chips. It’s About the Grid AI Runs On.

Copper closed at $14,371 a ton on London Metal Exchange on September 4, an all-time high for the exchange's daily settlement record. The explanation making the rounds is simple: artificial intelligence data centers are buying up the world's copper, and a shortage is coming.

The same week's data tells a more complicated story. LME copper stocks climbed to 234,175 tons on September 4 from 207,825 tons on August 17 — inventories rose as the price hit its peak. Over roughly the same stretch, the premium buyers were paying for immediate delivery over three-month futures collapsed from $535 a ton to almost nothing. Both are signs that the rush for physical metal was cooling, not accelerating, even as the headline price climbed. Something is tightening the copper market. It just isn't a shortage of copper on shelves, and it isn't, in the most literal sense, AI.

The mine cuts are real — and are a rounding error

Two Chilean producers did trim their outlooks this year, and their reasoning had nothing to do with computing demand. A second winter storm tore through Chile's Atacama region in mid-August, re-damaging a transmission tower that had already knocked out power to Lundin Mining's Caserones mine once in July. Lundin cut its 2026 guidance for the mine to 120,000–130,000 tons, from 130,000–140,000 tons, and raised its cash-cost estimate. Antofagasta, hit by the same weather system at its Los Pelambres operation, had already lowered its full-year copper forecast in mid-August to a range centered several tons lower than the guidance it gave back in February.

Add the two cuts together and the midpoint of combined guidance drops by roughly 45,000 tons for the year. Set against the roughly 25.9 million tons of copper the world used in 2023, by the International Energy Agency's count, that's about 0.17% of annual demand. It is a real loss, and in a market with almost no spare capacity, even a sliver like that can move a price. But it is not the arithmetic of an actual shortage — it's closer to what happens when a market with very little slack in either supply or demand gets a small shock and the price does most of the adjusting.

The storms also aren't a story about ore running out. The IEA's own minerals outlook estimates that projects already announced could fall roughly 31% short of the mining capacity a straightforward policy path would require by 2035, and that just over half of the world's mines sit in places under high water stress. A power line in the Atacama going down twice in a month isn't a freak accident so much as a preview of how thin the operating margin already is.

What AI actually buys is electricity — and copper follows the wires, not the chips

Here is where the popular version of the story breaks down. Nobody — not the International Copper Study Group, not the IEA, not the mining companies filing their own results — has published a widely accepted figure for how many tons of copper a data center actually consumes per unit of computing power. The number sometimes cited for how much copper sits inside a server hall is not something this analysis could trace to an authoritative source, so it isn't used here.

What the IEA does track with more confidence is where the copper is actually going: into the electrification of everything else. Copper demand tied to the energy transition — grid buildout, renewables, electric vehicles — is projected to rise from about 6.3 million tons in 2023 to roughly 12 million tons in 2030 and 16.3 million tons by 2040. Demand from copper's older, more traditional uses is essentially flat over the same period, holding around 19–20 million tons. Almost the entire growth curve in copper demand, in other words, sits inside the grid, not inside the server.

That framing lines up with how the miners themselves talk about it. In its half-year results, Antofagasta listed the forces behind long-term demand as energy security, electrification, digital infrastructure, and then — last on the list — AI. The company's near-term production growth story, meanwhile, is about the Centinela and Los Pelambres expansions finishing commissioning in 2027 and a plan to extend the Zaldívar mine's life to 2051 by weaning it off freshwater, not about signing supply deals with hyperscalers. AI is real in this picture. It just isn't the customer at the counter. It's the thing accelerating a grid-investment wave that was already underway, and pulling forward electricity demand that utilities otherwise would have spread across a decade.

The tariff story is about scope, not existence

A lot of coverage frames the Trump administration's copper tariffs as something the market is still waiting on. They're not — a 50% Section 232 duty on semi-finished copper and copper-intensive derivative products has been in place since August 2025, tightened twice since through 2026. What's actually undecided, and what's moving the price, is how much further the list of covered products grows. On August 6, the Commerce Department opened a comment period on adding roughly a dozen more derivative categories, including certain electric conductor cables and heat-exchanger components — grid and industrial hardware, not chips or servers. That is the detail worth sitting with: American buyers aren't nervous about whether a tariff exists. They're stocking up ahead of the next category that might get added to it, and cable and heat-exchanger parts are squarely electrification equipment.

That dynamic helps explain the price action better than a chips narrative does. U.S. buyers pulling metal in ahead of a possible tariff expansion pushes up the premium domestic buyers pay over the London price and pulls deliverable metal out of the seaborne market — a policy-driven repositioning of supply, layered on top of a real but small mining shortfall.

What to watch instead of the AI headline

Four things say more about where copper goes next than any AI narrative: whether the spot-to-three-month price spread widens again, whether LME inventories fall back below the roughly 208,000-ton low hit in mid-August, whether Chile or Peru announce a second round of guidance cuts, and whether the August tariff proposal is finalized to actually include cable and heat-exchanger products.

A few things are worth flagging as open questions rather than settled fact. Reported figures for a first-half decline in global mined copper output, and a bank forecast that 2026 could be the first down year for output since 2017, could not be traced back to a document this analysis could independently read, so they aren't treated as load-bearing here. Separately, how much substitution toward aluminum is actually happening — and how fast it could scale — doesn't have a quantified, authoritative source either, so this piece doesn't take a position on it.

Copper's climb has real support underneath it. But the case for higher prices holds up better as a story about grid investment, mining fragility and tariff uncertainty than as a story about AI buying up the world's wire. The AI narrative is the more dramatic one — but the boring version, an old industrial metal getting squeezed by an old industrial problem (the electric grid), and a fresh policy one (tariffs), is the one the data actually backs.