The last inelastic thing

The big five #hyperscalers have guided analysts toward something in the range of $700 billion in capital expenditure for 2026, roughly 2% of US GDP and more than the combined military budgets of Germany, France, the UK, Japan, Italy and Canada. The market has spent three years learning to price this: the model layer, the #inference layer, the accelerator, the power contract. It has priced them enthusiastically and, in a few cases, absurdly. What it has not priced with any seriousness is the fact that every dollar of that number eventually terminates in an arrangement of metal that somebody has to remove from the ground. There is no version of this buildout that routes around the pit. This is the largest mispricing currently on offer, and it is hiding in plain sight because it is boring.

Start with the arithmetic, because the arithmetic is not in dispute. A modern data centre consumes something on the order of 20 to 30 tonnes of copper per megawatt once you account for busbars, distribution and thermal management. @BloombergNEF expects cumulative #copper embedded in data centres to pass 4.3 million tonnes by 2035. Against that, mine supply is growing at roughly 1.4% a year, average ore grades have fallen about 40% since 1991, and the @IEA has said that existing and planned mines can cover only around 70% of projected 2035 demand. In the United States, the average time from discovery to first production runs close to 29 years. That last figure is the whole thesis. A software company facing demand can ship capacity in a quarter. A model can be distilled, quantised, or moved to cheaper silicon. An orebody cannot be prompted into existence. Price signals that would ordinarily summon supply instead summon a permitting process that outlives the executives who started it.

Then there is the leverage question, which is where this stops being a #commodities argument and becomes a strategic one. @China controls roughly 70% of rare earth production and close to 90% of refining and processing. Between January and June of this year, neodymium-praseodymium oxide prices rose sixfold, #tungsten concentrate tripled, and #antimony doubled. The export-control suspensions negotiated in late 2025 lapse in November. The IEA has warned that full enforcement of China's regime could put something like $6.5 trillion of downstream production at risk. Consider what that means for the companies everyone is actually buying. $Nvidia's moat is real. It also sits downstream of a licensing decision made in Beijing by people who understand exactly where it sits. When the binding constraint on a $700 billion buildout is a gallium licence and a permitting queue, the correct question is not who captures the margin but who controls the gate.

Mining is a historically miserable business. It is capital-intensive, cyclical, and populated by price-takers who have reliably destroyed shareholder value at the top of every cycle by overpaying for assets. Producers frequently fail to capture spot moves at all. Jurisdiction risk is genuine and occasionally total. And "this time the buyer is different" is the oldest sentence in commodities. What I think has actually changed is the counterparty. The marginal buyer is no longer a Chinese construction cycle of uncertain durability. It is a handful of the most creditworthy balance sheets on earth, alongside 54 governments that convened in Washington in February and left with 21 new bilateral minerals agreements, and defence procurement arms writing cheques at multiples of Chinese spot to stand up domestic processing. Offtake is starting to look like a contract rather than a hope. That does not make every junior explorer a good investment; most of them remain lottery tickets with a corporate structure. It does mean the sector is being repriced from cyclical to strategic, and that repricing is early. Everyone is buying the model. Almost nobody is buying the hole.