AI Supply Chains: One Copper Mine, Two Systems
It takes roughly seventeen years to bring a new copper mine from discovery to first shipment — and twelve of those years are spent on permitting and legal challenges before a single truckload of ore moves. Meanwhile, a 300 MVA transformer that cost $3 to $5 million in 2019 now runs $6 to $10 million, and the wait time to get one built has stretched from about a year to nearly four. Nobody manufactures grain-oriented electrical steel — the specialized material every transformer core needs — outside of five companies worldwide. These are not numbers about artificial intelligence, or about the energy transition, or about geopolitics. They're numbers about how slowly the physical world moves when everyone suddenly needs the same things at once.
And everyone does need the same things at once. The AI buildout, the renewable energy transition, and the aging fossil fuel system are not three separate stories competing for headlines — they're three claimants on the same finite pool of copper, transformers, skilled labor, and grid interconnection capacity. Hyperscale data center capex is now running above a trillion dollars a year, most of it flowing into power infrastructure. Renewable and grid modernization spending is climbing on a similar trajectory, driven by the same electrification logic. And the legacy energy system — declining in efficiency, geopolitically exposed — needs its own emergency patching just to keep functioning. There is one global supply of specialized steel, and three separate booms all trying to draw from it simultaneously.
The geopolitical backdrop has made this concrete rather than theoretical. Since a war between Iran and a US-Israeli coalition began in February 2026, the Strait of Hormuz — the route for roughly a fifth of the world's oil trade — has been effectively closed to normal traffic for months. Shipping volumes through it remain well below pre-war levels even after a ceasefire and a subsequent memorandum of understanding; fighting resumed this summer, and some energy economists now think the disruption could persist into 2027. The point isn't to dramatize a live conflict still producing real casualties. It's that the old energy system's fragility and the new one's construction costs are compounding at the same moment, for unrelated reasons, and the result lands on the same input markets.
You can watch this tension play out inside a single company's earnings call. Eaton's electrical backlog is up 43% year-over-year, driven by exactly the AI data center and grid buildout this piece is describing — and in the same call, its executives are fielding analyst questions about copper cost inflation eating into margins. The demand and the cost pressure aren't separate storylines. They're the same order book. Quanta Services, the engineering and construction firm doing much of the physical work of grid modernization, just posted a $53 billion backlog and record quarterly revenue, with its CEO describing the company as being in the "early stages of a decade-long infrastructure buildout." That's a bet that the current bottleneck resolves into a long runway, not a stall.
Albemarle, sitting further upstream in lithium and battery materials, is making a visibly more cautious bet: management has described 2026-2027 growth as deliberately capital-efficient rather than expansionary, phasing new investment and explicitly naming Middle East instability as a risk to input costs. Two companies, adjacent in the same supply chain, making different judgments about how long the current squeeze lasts.
The honest complication is that nobody currently knows the answer, and it's not for lack of trying. Economists this year have been re-litigating Robert Solow's 1987 observation that "you can see the computer age everywhere but in the productivity statistics" — because AI is running into the same wall. Federal Reserve surveys this year found the large majority of firms reporting negligible measured productivity gains from AI investment; Goldman Sachs and JPMorgan both cut their productivity forecasts in response. The optimistic case is that this is the same lag electrification and enterprise IT went through before their productivity gains eventually showed up in the data, sometimes a decade later. The pessimistic case is that this time is different. As of this month, that argument is unresolved among people paid to resolve it, which should be a caution against anyone writing a confident date on the other side of this cycle.
That uncertainty, though, points toward a fairly clear positioning conclusion rather than a mushy one. If nobody can reliably time when AI's productivity payoff shows up, or when the Strait of Hormuz situation resolves, the more durable trade isn't a bet on the destination — it's a bet on the toll road. The companies physically building the bottleneck's solution get paid whether the AI productivity thesis takes two years or ten to materialize, because copper, transformers, and grid capacity are needed either way; Quanta's $53 billion backlog and Eaton's 43% order growth aren't contingent on AI proving out, they're contingent on the buildout continuing, which is a lower bar. The more exposed position is anywhere downstream of that bottleneck whose valuation already assumes the productivity gains have arrived — that's the segment the Fed surveys and the Solow-paradox debate should make investors most skeptical of today, independent of the underlying technology's long-run promise. And the more conservative posture upstream, in raw materials, looks like Albemarle's own stated approach: capital-efficient, phased, unwilling to overbuild into a supply chain this exposed to geopolitical shock. None of this is a call on how the AI productivity question ultimately resolves. It's a recognition that the physical infrastructure layer gets paid on a different, more observable timeline than the technology layer above it — and that in an environment this uncertain, the safer bet is the one that doesn't require guessing correctly which decade the payoff arrives in.
