The Next Bottleneck Is Not HBM. It’s Copper.

I spend most of my time underwriting the compute layer of the AI buildout: data centers, power contracts, GPU economics. But every so often someone forces me to look one layer deeper in the stack, and this week it was Dan Dreyfus of Bornite Capital, whose recent talk on critical minerals is one of the cleanest articulations I’ve heard of where this cycle is heading. What follows is his argument, filtered through my own framework, because I think he’s pointing at something most AI investors are not pricing.

His opening line is the thesis in miniature: going forward, human progress will be measured by how much electricity we consume. If you believe that, then everything downstream of electricity, the wires, the transformers, the turbines, the metal in the ground, becomes the binding constraint on progress itself.

The end of the capital-light miracle

From the early 2000s until a few years ago, the United States ran what Dreyfus calls an economic miracle: trillions of dollars of market cap created with almost no capital. Google built a search engine. Meta bought WhatsApp, twelve employees, for nineteen billion. Software ate the world and the world barely had to pour any concrete. And while that was happening, America was actively dismantling its physical industrial base and shipping it to China, doubling down on the capital-light mentality.

Then the bill arrived. COVID, Russia-Ukraine, tariffs, Middle East flare-ups. Each shock sent inflation vertical, and the reason it never fully came back down is that the supply chains had been hollowed out. There was no resiliency left to absorb anything.

Now the country wants to run the movie in reverse: reshore, re-industrialize, electrify, and on top of all that, build an AI compute layer that is orders of magnitude more infrastructure-intensive than any previous generation of computing. That is a demand shock colliding head-on with a supply shock created by twenty years of underinvestment.

Dreyfus counts more simultaneous capital cycles than at any point in his career: aerospace, with a trillion dollars of Boeing and Airbus backlog plus a space economy competing for the same materials; the grid, a trillion dollars per decade for the next thirty years; data centers, now running at roughly a trillion dollars per year; semiconductor fabs; and a global defense buildup from Taiwan to Europe. The common denominator is that none of it works without critical minerals.

The wake-up call

In April of last year, China cut off exports of a list of materials most people cannot pronounce: samarium, gadolinium, terbium, dysprosium. The result, per Dreyfus, is that Ford came within days of shutting down its entire production line for lack of samarium-cobalt magnets. Days.

That put Washington into a genuine panic, and to its credit, the response has been aggressive. The government is now showing up at the door of small resource companies that were left for dead for two decades and handing them three pieces of paper: an equity check, a permit that had been stuck for twenty years, and a take-or-pay offtake agreement with a floor price that guarantees the project a high IRR. Dreyfus, twenty-five years in commodities, says he has never seen anything like it. China’s grip on processing remains nearly absolute, and catching up will take ten to twenty years, but the flywheel has finally started turning.

The copper math

Here is the part worth writing on a whiteboard.

Over the past ten thousand years, going back to the earliest mines of the ancient world, humanity has extracted roughly 700 million tons of copper. If copper demand merely grows in line with GDP, with no extra credit for data centers, robotaxis, or solar, the world will need another 700 million tons over the next eighteen years.

As much copper in eighteen years as in all of recorded history.

Meeting that requires something like five new tier-one mega mines coming online every single year. You can count the tier-one projects actually arriving before 2030 on one hand with fingers to spare. A new copper mine takes seven to twelve years to build. The great Chilean mines are over a century old with depleting grades. The supply response mathematically cannot arrive on time.

And the demand kickers are not small. A one-gigawatt AI factory consumes roughly 50,000 tons of copper. At fifteen gigawatts of new AI capacity per year, that is 750,000 tons of incremental annual demand from data centers alone, against global mined supply that grew by only about 500,000 tons last year. An EV uses five to six times the copper of a combustion car. Solar takes five times the copper per megawatt of a gas turbine, wind seven times. Even war consumes it: artillery is largely copper, and nobody recycles a battlefield.

Today the market is obsessed with HBM and NAND because memory is the visible bottleneck and prices are going vertical. Dreyfus’s point is simple: if you want to see the next bottleneck before it becomes consensus, look at copper. He thinks the price can easily double. Silver tells a similar story, running a structural annual deficit of roughly 200 million ounces against perhaps 600 million ounces of above-ground inventory, with photovoltaics eating the rest.

The grid, and the debasement kicker

None of this even requires the AI thesis to work. The US grid has been quietly rotting since the postwar buildout; parts of it are over a century old, including the transmission line that sparked the Paradise fire in California. Just normal life, heat pumps replacing boilers, modest EV penetration, more devices, already pushes the system toward brownouts and rising prices. The biggest constraint is not even materials. It is craft labor, the electricians and linemen an entire generation was told not to become. The irony Dreyfus highlights is sharp: the same blue-collar middle of the country that globalization gutted is now seeing trade school graduates start at $150,000 while entry-level white-collar work gets automated.

Layer on the macro. Forty trillion of federal debt growing at two and a half trillion per year, a similar present value of unfunded social liabilities growing just as fast, against five and a half trillion of tax receipts. The next recession gets resolved the way the 1970s were resolved: print, debase, inflate. And in the 1970s, the decade the dollar lost most of its purchasing power, the best performing asset class by a mile was commodities. Hard assets are how you defend purchasing power in that regime.

Where this lands in my framework

Regular readers know my portfolio is concentrated in the power-to-compute conversion layer: companies turning megawatts into intelligence. The Dreyfus thesis is the same trade, one layer down the stack. Energy is becoming the binding constraint on progress, and copper is the binding constraint on energy.

His own warning is worth keeping front of mind, though: great end markets do not protect you from owning the wrong part of the supply chain. China can dump supply, technology can route around a pinch point, and a miner with bad jurisdiction or bad grades can lose you money in the middle of a supercycle. The work is in finding where the pinch is genuinely unsolvable on the relevant timeline, and copper’s seven-to-twelve-year mine cycle is about as unsolvable as it gets.

Commodity cycles tend to run fifteen years and produce returns measured in hundreds of percent. By Dreyfus’s count we are a few years in. The capital-light era built the software. The capital-heavy era has to build everything else.

Not investment advice. I may hold or initiate positions in companies exposed to these themes. The ideas summarized here are Dan Dreyfus’s; the errors of interpretation are mine.