The bottleneck stack

Point at the headline technology and you will usually miss where the value is. The value sits lower down, in something unglamorous, and it moves.

Every industrial system runs on a stack of constraints. Above the visible technology sits nothing. Below it sits a series of layers that each have to hold for the system to work: physics, capital, supply chain, integration, regulation, workforce, procurement, customer trust. The technology at the top gets the coverage. One of the layers below it is the constraint that actually decides how fast the system can grow. That layer is the binding constraint, and reading an industrial system means finding it.

The stack matters because value pools at the binding constraint. When a system is short of one thing, whoever controls that thing captures the margin, regardless of who owns the headline technology. The engineer’s instinct is to look at the most advanced layer. The analyst’s job is to look at the layer that binds, which is almost never the advanced one. And the binding constraint moves. Relieve it, and the pressure jumps to the next layer up or down the stack. A framework that names today’s bottleneck without tracking its next move is a snapshot, and snapshots go stale.

The worked example: the AI data-centre build-out

Read the AI build-out through the stack and it stops being a story about chips.

The headline layer is compute. This is where the coverage sits: the newest accelerator, the frontier model, the benchmark. It is real, and chip supply is tight and improving. The binding constraint has already moved below it.

Drop one layer to power. Data centres used roughly 415 terawatt-hours of electricity in 2024, and the IEA projects that reaching about 945 terawatt-hours by 2030, close to the current annual consumption of Japan. In the United States, data centres are on course to drive nearly half of all electricity demand growth this decade. The compute exists. The power to run it is now the question, and value has begun to pool wherever firm, dispatchable electricity can be secured. That is why compute companies are signing power deals and reopening nuclear plants. They are buying the binding constraint.

Drop again, because power alone is not the floor. Electricity has to reach the site through a grid connection, and the connection is where the queue is. Berkeley Lab’s 2025 study found over two thousand gigawatts of generation and storage waiting in the US interconnection queue. The median wait from request to operation now runs beyond five years. Securing power on paper does not secure it on the grid. The connection binds.

Drop once more, to the transformer. A grid connection needs large power transformers to step voltage up and down, and these are now the hard floor of the entire build-out. Wood Mackenzie’s 2025 survey put lead times for large power transformers at well beyond two years, with some units approaching four. A transformer is copper, grain-oriented electrical steel, and a skilled winding workforce that the industry spent two decades not training. You cannot download it, and you cannot rush it. For a large share of data-centre projects, the transformer is the item that dominates the schedule. Everything above it in the stack, the connection, the power, the compute, waits on a piece of heavy electrical equipment with a four-year queue.

Below the transformer sit the constraints that decide the transformer: grain-oriented electrical steel capacity, and the skilled labour to wind and assemble the units. The bottleneck runs all the way down to a material and a trade.

Reading the stack

The value in the AI build-out is therefore not distributed the way the coverage suggests. It pools at whichever layer binds, and over the past two years that layer has walked down the stack. From chips, to power, to grid connection, to transformers, to the steel and labour beneath them. An analyst who fixed their attention on the top layer missed every move. An analyst reading the stack watched the constraint travel and watched the margin travel with it.

This is the general method, and it works on any industrial system. Name the headline technology, then ignore it. Walk down the layers and find the one that actually binds, the one that is short and cannot be relieved quickly. That is where the value is pooling now. Then ask the second question, the one that separates a framework from a snapshot: if that constraint is relieved, where does the pressure go next. Build the transformer capacity and the bottleneck jumps to electrical steel and skilled winders. Relieve those and it jumps back up to grid permitting. The stack is a map of where value has been, is, and is going, read in the order the constraints actually bind.

The unglamorous layers are the story. The headline technology is the part everyone can see, which is exactly why the value has already moved somewhere else.

This is the lens the diagnostic applies to find where the value is actually pooling.