Lead Times

There's a fact about the AI buildout that doesn't fit the narrative around it. Large power transformers – the substation-grade units that step grid voltage down for industrial customers like hyperscale data centers – have lead times of 3 to 5 years. Five years ago, before anyone was building gigawatt data centers for training runs, lead times were 24 to 30 months. The current backlog is AI demand piled on top of capacity that was already constrained.

The companies making them are Eaton, GE Vernova, Quanta, Hitachi Energy, and a handful of others. The base technology is roughly a century old. Tesla and Westinghouse fought their currents war in the 1880s; the equipment evolved through the 1920s and was mostly architecturally settled by the 1960s. What's been added since is automation, monitoring, and incremental efficiency. The transformers themselves are the same kind of object: a stack of iron-core laminations, copper windings, oil insulation, a tank, bushings. A 1960s electrical engineer could walk into a modern substation and recognize most of what's there.

This is the substrate the AI buildout sits on. The most advanced computing infrastructure ever built – racks of GPUs running training workloads that didn't exist three years ago – is fed by a grid whose limiting component evolved before silicon.

The pattern across the physical layer

The constraint isn't unique to electrical infrastructure. It shows up everywhere the AI buildout meets the physical world.

Copper is the same story. Demand is being driven by electrification generally and AI specifically (each data center needs miles of conductor; each EV needs ~80kg vs ~20kg for an internal-combustion car). But supply expansion runs on geology – finding a deposit, getting permits, building a mine, ramping production – and the timescale for that is 10 to 15 years. The deposits being brought online now were identified before the iPhone. The pipeline of deposits being identified now will produce metal for the 2040s.

Cooling has its own physical-history layer. Liquid cooling for high-density racks is a real innovation, but the underlying chemistry – heat-transfer fluids, the metallurgy of cold plates, the hydraulics of closed-loop systems – is mostly post-WWII industrial chemistry adapted from refrigeration and process cooling. Vertiv's revenue is growing 78% year-over-year by selling configurations of components whose individual building blocks are decades old.

Water rights, transmission rights-of-way, port capacity, semiconductor-grade chemical supply chains, the limited number of cranes that can lift a 400-ton transformer – every layer the AI era touches turns out to have been built (or not built) by someone making decisions for a different future.

Why the rhetoric ignores this

The rhetoric of AI is the rhetoric of acceleration. "Exponential improvement." "Faster than Moore's law." "Compute doubling every six months." The framing comes from the part of the industry whose product is software – code, models, services that can be copied for the marginal cost of electricity. From inside that part, the world does feel like it accelerates exponentially, because the constraints look like they're somewhere else.

Software is the top layer. Underneath: silicon fabs with 3-year construction timelines, data centers with multi-year permitting, transformers with multi-year lead times, copper mines with decade-plus development cycles, electricians and welders and crane operators that take years to train. The exponential accelerates against a substrate that's basically linear. Sometimes sublinear. Often constrained.

The actors loudest about AI's pace are systematically the ones least exposed to the physical layer. OpenAI doesn't manufacture transformers. Anthropic doesn't pour concrete. The hyperscalers (Meta, Microsoft, Google, Amazon) do operate at the physical layer – and they're the ones who actually talk about lead times and grid connections and water permits in their earnings calls, because they're hitting the constraints directly. When Sam Altman says capability will double again in six months, he's reporting what's true inside his domain. When the CEO of Eaton talks about needing 18-month lead times even for emergency orders, he's also reporting what's true. Both can be right; they describe different layers.

The pattern across technology eras

Every technological era leaps forward on top of slow infrastructure that didn't anticipate it.

Railroads in the 1840s outran the canal system that had been the dominant freight network. Canals were dug for 1-2 mph mule-towed barges; the early railroads ran at 30 mph and ignored the canal geography entirely. Container shipping in the 1950s and 60s did the same to the port-warehouse-truck infrastructure of the previous era – Malcolm McLean's standardized 20-foot box made most existing ports obsolete because their cranes couldn't handle the loads. The internet ran for decades on copper telephone wires laid for human voice; the speeds we now consider broadband only became possible after Cold War-era fiber-optic research filtered into commercial deployment, and even then the last-mile wires often stayed copper for another generation.

Each leap inherits the substrate that didn't expect it. The substrate becomes the bottleneck. The bottleneck becomes where the actual money is, even though it's not where the visible glamour is.

In the AI era specifically: it's not the model companies that get the durable value from the buildout. It's the companies that own the bottleneck infrastructure. Eaton's market cap has roughly doubled since 2023, while it makes 1960s-style transformers. Vertiv's gone up 5x, selling cooling configurations. The lithium and copper miners are the ones building the resource base for the 2030s. The dollars flow to the physical layer because the physical layer is what's actually scarce.

How to read the present

The bottleneck IS the story. Whoever controls the bottleneck inherits more of the value than whoever invented the thing on top. The AI revolution will, in the historical reading, look much like the railroad revolution: a software layer (timetables, scheduling, ticketing) that organized something powerful, on top of a hardware layer (rails, locomotives, signaling) where the actual durable wealth accumulated. The model companies are railways' Pullman and Western Union; the durable infrastructure plays are railways' steel mills and coal miners.

Reading the present requires distrust of the dominant narrative. The dominant narrative is written by the layer with the loudest voice – the software layer, because its content travels for free. The constraining layer is mostly silent because it's busy doing the constrained thing. To understand a technology era, listen to the people complaining about lead times. They're not behind. They're upstream.

And the "AI is unprecedented" framing is wrong in a specific way. It's not unprecedented; it's the latest era doing what every era does. The leap is real. The unprecedented-ness is rhetoric. What's actually unprecedented is the speed of the visible layer, not the speed of the substrate it sits on. The substrate moves like substrates have always moved. The era's specific shape comes from the gap between those two speeds.

Coda

This essay's substrate is the research section of this garden. Three thesis pieces from April – on transformers, copper, cooling – plus two follow-up reviews laid out the specific bottlenecks as investment theses. The argument here generalizes from that work: the specific bottlenecks are instances of the pattern, and the pattern is older than the era it's currently visible in.

The proof-of-pattern test is whether the next decade rewards the bottleneck operators more than the model companies. The research bets say yes. The argument here just names what those bets are an instance of.