Who benefits from the AI power bottleneck?

Mathew’s exercise starts from two things he treats as settled: AI is a foundational technology that will persist, and far more energy infrastructure is needed. What is unsettled is how much and for how long, and underneath that gap sits a structural problem he calls a duration mismatch. Technology plans on one-to-two-year cycles while energy assets are underwritten over many years or decades. Because the uncertainty is genuinely large, the stories told about it carry weight, and the way to read them is to follow the capital: from the hyperscalers spending it, through chips and data center equipment, out to power equipment, utilities, engineering firms and labor. Mathew is explicit that this is not an accusation of dishonesty. The point is that economic incentives shape views more than the market appreciates.

The hyperscalers are the hardest case. They can build capacity or buy it from neo clouds and colocators, and they prefer to build, having the teams and wanting the control. But committing to energy assets underwritten over ten or twenty years is a different risk than they want to carry, so the incentive is to talk the constraint up: that recruits developers, speculators and neo clouds to build instead, and if the market overbuilds, leases get cheap and the hyperscaler buys them. Mathew points to Satya Nadella saying something close to this in a podcast interview. A second convenience follows, which is that if power is the binding constraint then demand for AI services is not. Kann then puts the opposite incentive: if the world does not solve the bottleneck, hyperscalers are best placed to win a scarce resource, because they can write long-lead turbine orders a neo cloud cannot finance, which makes a lasting shortage a moat against exactly the class they are leaning on. His guess is that wanting the world to solve the problem wins out, but he says it is not clean. Mathew adds that the behavior looks like optionality rather than conviction: deposits on long-lead turbines and switchgear, which he sizes hypothetically at a few million dollars, are cheap against the opportunity cost of an idle data center, and can be walked away from.

Most of the rest is simpler. Equipment makers and engineering contractors get record backlogs, longer backlogs, and pricing power many have not had in generations; scarce electricians and plumbers put contractors on the same side of the trade. Powered-land owners gain directly, since anything that shortens time to power rises in value while the constraint holds. Chipmakers look conflicted and are not: NVIDIA’s pitch is tokens per watt, meaning useful output per unit of energy, so the scarcer and costlier watts become, the more its efficiency advantage is worth and the weaker the case for cheaper purpose-built custom silicon. Utilities are the real exception, not because their upside is smaller but because they are also the ones blamed. They earn a regulated return on capital spending, so a generational build should suit them, but they answer to public utility commissions, face rate backlash and local opposition, and cannot blame slow AI growth on a scarcity they are supposed to be solving. Mathew describes their tone shifting over recent quarters from advertising gigawatt pipelines toward proving they can execute quickly and without wasting capital.

The other side of the trade is short but instructive. Independent power producers own existing fleets and sell into power markets, so scarcity drops almost straight through to margin without new capital, and new supply would dampen the prices they earn on. Mathew reads Constellation’s chief executive arguing that power prices do not support building new gas plants, and pointing instead to flexibility options such as throttling workloads at peak, as consistent with that incentive, while saying plainly the claim may or may not be true on the merits. Gas producers sit similarly, since a drilling rush would cap their own prices. Liquefied natural gas exporters sit opposite: they buy cheap domestic gas and sell into premium markets abroad, so they want production high but domestic demand low, which makes a wave of new gas plants bidding for the same molecules a threat to the arbitrage. All of which leaves the question no incentive map can answer, and the one Mathew says the market has moved onto. Roughly half a trillion dollars a year is going in, so is the return positive, and if power were suddenly unconstrained, would AI’s trajectory still be this steep? The AI infrastructure trade is now priced on the durability of the constraint rather than on its existence.