Everyone knew AI needed power. Nobody was paid to follow the wire to the transformer maker.
Everyone has heard that AI needs electricity. Ask a language model what follows and it will draw the chain in one pass: more compute means more data centers, which need power, which needs grid connections, transformers, switchgear, gas turbines, cooling and construction crews, several of which cannot be built quickly. If a chatbot can draw that map in seconds, shouldn’t the market have priced it already?
The efficient-market answer is yes: anything a chatbot can infer is in the price. Yet theses like this one often take quarters or years to show up in prices, and the reason is not that investors are stupid. A market is not a single intelligence. It is a bureaucracy of capital.
One map, many desks
Imagine a fire spreading through a city where each department sees one street. The chip analyst sees NVIDIA demand, the utility analyst sees load growth, and the industrial analyst sees transformer backlog. Each is correct locally, but nobody is paid to own the whole causal map. The LLM draws the map immediately. The market funds it slowly.
That delay is market latency: the time between when a structural implication becomes inferable and when enough capital has the permission, conviction and mandate fit to price it. The power thesis has to pass a series of gates, each rational on its own. Someone must translate “AI needs power” into quantities an analyst can model, such as backlog, lead times and how long margins stay high. A portfolio manager must be allowed to buy the beneficiary, and an AI fund may not be permitted to own switchgear. The trade must stop sounding weird, which happens only after the sell-side has a name for it and management teams repeat it on earnings calls. Only then do revisions, momentum screens and ETFs pull it into the price. Like an inadequate equilibrium, the system fails to act on the whole while competent people see pieces of it.
Where the edge lives
The headline that AI needs power was priced long ago. The slower question is whether the market has priced the duration and margin structure of the bottleneck. AI demand may have turned boring physical components into scarce, standardized, mission-critical parts, and a customer fighting for a delivery slot cares more about getting a transformer than about its price. If that lasts for years, the old normalized margins are the wrong base rate, because scarcity can change industry structure before consensus changes its margin model.
LLMs are unusually good at this gap between a popular narrative and a change in an industry’s economics, because they can assemble a causal graph across domains before the institutions holding capital reorganize around it. So the truth lies between “markets are dumb” and “it’s already priced in”: markets are efficient at processing information that already fits their institutional machinery, and inefficient at processing implications that cross mandate boundaries.
Alpha here takes the form of a permission gap: the implication is inferable, but the beneficiary sits outside the obvious sector bucket and the old valuation model still uses stale margins. That is why cross-domain trades feel strange at first. Once the trade has an acronym, a sell-side basket and an ETF, the edge has been domesticated.
Seeing early is not getting paid
The LLM can see the road, but the market knows how crowded the road already is, and how painful it is to be early. The power thesis fails if the stocks already discount years of perfect execution, if double ordering inflates backlogs, or if the bottleneck migrates to companies you cannot buy.
A cross-domain thesis therefore needs three separate answers. Is it true? Is the market late to it? And is there a liquid instrument where the risk still pays you for being early? Most intelligent commentary answers only the first. The question to ask of any idea that seems obvious is which layer of the market knows it, and whether that knowledge has become capital yet.
References:
- Current AI power-equipment thesis discussion, May 2026