Copper: the perfect squeeze?

LME cash copper held around $13,500 a tonne in mid-July, up roughly 10% year-to-date, after touching records near $14,500/t earlier in 2026. That resilience sits on top of two forces: a mine-supply base that is visibly struggling to grow, and a demand picture, supported by an "AI supercycle". Our reading is that copper is structurally tight, and could be much tighter in the near future.

‍

AI is copper's new marginal buyer — even as China cools

The demand case has acquired a second engine alongside electrification. The largest cloud and platform companies are on course to spend roughly $725–750 billion of capital in 2026, up around 70% on the year, and increasing more by 2027. The overwhelming majority of it on data-centre capacity and the chips inside it. Whatever the return on that spend sometimes questioned as a buble, its physical footprint is real, and copper is one of its hardest inputs to design out.

AI-optimised facilities are markedly more copper-intensive than the cloud data centres that preceded them. High-density GPU racks distribute power through heavy copper busbars rather than conventional cabling, liquid-cooling manifolds and bus-work add further copper, and the grounding and interconnect requirements of multi-billion-dollar clusters are larger again. A useful real-world anchor is Microsoft's Chicago campus, which embeds roughly 2,177 tonnes of copper across about 81 MW — close to 27 tonnes per megawatt of capacity. That ~27 t/MW figure is a reasonable facility-level benchmark for a modern build; AI-dense halls run higher, in the 30–45 t/MW range, once heavier power delivery and cooling are included. Crucially, none of this counts the grid connection — substations, transformers, redundant feeds and transmission — which can add anywhere from the same amount again to several multiples per megawatt when a campus forces new network build.

The global operational data-centre base is on the order of 100 GW and expanding at a mid-teens-to-20% pace as AI build-out accelerates, implying something like 15–20 GW of new capacity commissioned this year. Applying a facility-level intensity of 27–40 t/MW to that range gives roughly 300,000 to 700,000 tonnes of copper embodied in 2026 data-centre construction, with a central figure near half a million tonnes. Of that, the genuinely AI-specific uplift — the extra copper from higher-intensity designs plus the capacity AI is pulling forward that would not otherwise have been built — is on the order of 100,000–200,000 tonnes. The grid reinforcement needed to actually power these sites is a further 300,000–500,000 tonnes, though that copper is shared with the wider electrification build and should not be double-counted as "data-center demand."

Put in context, the central ~500,000-tonne facility figure is under 2% of the roughly 28.7 million tonnes of refined copper the world will consume this year — small against the stock. Global copper demand is growing only about 1.6% in 2026, or roughly 450,000 tonnes. On that basis, data-centre construction alone is absorbing an amount of copper equivalent to essentially the entire year's growth in global consumption — before the grid build-out on top. AI is not yet a large share of copper demand; it is a dominant share of the marginal tonne. And in a tight market, the marginal tonne sets the price.

Two features make this demand unusually price-insensitive. First, copper is a trivial share of a hyperscale project's cost: a $50,000-per-megawatt copper bill is a rounding error against a campus that can run past $10 billion, so builders are effectively indifferent to whether copper trades at $10,000 or $20,000. Second, the electrical and safety requirements of high-density power delivery leave limited room to substitute inside the rack. That combination — large marginal demand, near-zero price elasticity — is what makes the theme a genuine structural support rather than a cyclical one.

Against that structural pull runs a cyclical drag from China, still around 58% of global refined demand. Property remains in contraction, passenger-vehicle sales fell about 19% year-on-year and home-appliance sales grew just 1% in the first five months of the year. Yet copper is behaving as the exception among base metals: Shanghai exchange inventories have drawn down in 2026 even as other metals accumulate, supported by a surge in State Grid investment — power-grid spending up an estimated 35–40% year-on-year — that is offsetting the property and solar downturn. Refined-copper imports rose about 3% year-on-year in June to 478,000 tonnes, the strongest month since September 2025, while copper-concentrate imports slid to a 19-month low and smelter treatment charges have gone negative. The signal is not a demand collapse but a rotation: China's copper pull is narrowing onto the grid and electrification, precisely the copper-intensive end of the economy, while construction and consumer channels fade.

‍

Supply is genuinely constrained

The second-quarter reporting season was a run of disappointments so far on the production side. Rio Tinto's consolidated copper output fell 7% year-on-year (and 7% quarter-on-quarter) to 213,000 tonnes; BHP's FY26 copper slipped 3% to 1.95 million tonnes and it guided FY27 lower still, to 1,650–1,800 kt, flagging a grade-driven fall of as much as 15.5% next year at Escondida, the world's largest copper mine. Antofagasta edged down 0.7% quarter-on-quarter and is banking on a stronger second half to hit guidance. Codelco — carrying roughly $25 billion of debt with output near a 28-year low — now expects production merely to hold near current levels, against the backdrop of an external audit and a scandal over some 20,000 tonnes allegedly misreported in a 2025 production report.

The country-level picture is starker than any single company's. Chile's copper output fell 13% year-on-year in May to 423,623 tonnes, a tenth consecutive month of annual decline, with Escondida, Collahuasi and Codelco all down double digits. That simultaneity across three unrelated operations is the tell: this is not a one-mine outage but the maturation of an entire ore base — falling grades, ageing infrastructure and years of under-investment in grade management showing up at once. Layered on top is the overhang from Indonesia, where Freeport's Grasberg — the world's second-largest mine, around 3% of global supply — remains under force majeure after the September 2025 mud-inflow disaster, with close to 600,000 tonnes of output lost between the accident and the end of 2026 and a phased restart only reaching an estimated 85% of capacity in the second half of this year. Codelco's El Teniente also suffered a fatal tunnel collapse in mid-2025.

The structural read-through is the real story. Ore grades at the major Chilean mines have drifted below 0.7%, meaning miners move more rock for less metal, at higher energy and water cost. The capital intensity of new greenfield projects now routinely runs $20,000–30,000 per tonne of annual capacity, and above $30,000 is no longer unusual — so even at today's prices, sanctioning new supply is a marginal decision, and lead times from discovery to production stretch 15–20 years. Concentrate is scarce enough that treatment and refining charges have turned negative, an extraordinary state of affairs in which smelters rely on by product credits (such as gold and molybdenum) to remain profitable. World mine-production growth this year is running at the best by 1.6% with a strong H2 2026, and refined-production growth barely positive.

What to expect

The official industry balance has swung from an expected shortfall to a modest surplus of roughly 100,000 tonnes this year, with a larger surplus pencilled in for 2027. The reason is not any recovery in mine supply — it is that demand has softened, particularly in China. Visible exchange inventories are also high, around 660,000 tonnes across the major venues, but that figure is badly distorted: it is heavily concentrated on the US COMEX exchange after a year of tariff-driven front-running. Stocks outside America, on the London and Shanghai exchanges, are by contrast tight, with a rising share of metal earmarked for withdrawal. So the correct summary is nuanced: mine supply is genuinely constrained and structurally so; that constraint is currently buffered by weak consumption on the old China economy and a scrap surge.

The bull case for copper looks genuine and multi-year: AI, grid modernisation and electrification pulling in the same direction against a supply base that responds in decades, not quarters. On the potential downside risk, a stronger dollar is the cleanest headwin. With the market pricing at least one further Fed move this year and oil-led inflation risk emanating from the Middle East, dollar strength weighs directly on a dollar-priced metal regardless of the physical balance. Weaker Chinese demand is the second: if the property drag deepens and cooling consumption settles in as the structural norm, surplus Chinese metal could spill into the global market and keep inventories elevated. Third is an AI-capex air-pocket: the build-out is increasingly debt-funded — the largest hyperscalers have roughly doubled their combined debt in five years, one recently saw free cash flow turn negative, and bond investors have started to push back on the volume of new issuance. Because AI is copper's marginal buyer, any slowdown in build-out plans would remove precisely the tonne that is currently setting the price.

However, any kind of normalization in US-Iran relations or Chinese stimulus sufficient to sustain GDP growth at around 4.5% could rapidly reshape the copper outlook. While these developments would reinforce demand expectations, they could also accelerate the market's bullish narrative, with a growing number of investors now targeting $20,000/t over the medium term—around 50% above current prices.

Finally, the AI build-out is not a copper-only trade. The same capex wave lifts a basket of adjacent materials. Aluminium benefits twice over, as a data-centre structural and busbar metal and as copper's substitute-of-choice when prices bite. The minor and critical metals embedded in AI servers — tungsten, tantalum, indium, gallium, germanium — face structurally tight balances on inelastic supply and concentrated processing.

But the purest adjacency is power itself. Every megawatt of compute needs a megawatt of generation and the network to deliver it, which turns the AI theme into a demand story for natural gas, uranium (and the small-modular-reactor pipeline), electricity as an asset, and the hardware of the grid: transformers, grain-oriented electrical steel, and high-voltage cable.

Main commodities Performance Ytd

Source: Marketview

Market insights you can trust

Kpler delivers unbiased, expert-driven intelligence that helps you stay ahead of supply, demand, and market shifts.
Trade smarter. Request access to Kpler today.