Nvidia's chief executive, Jensen Huang, describes the AI ecosystem as a "five-layer cake".
Layers four and five, the applications people actually use such as OpenAI and Claude, are the most visible part of that stack.
Layers one through three, energy, chips and infrastructure, are the foundational "AI factories" required to build the models in the first place.
What a hyperscaler actually is
Hyperscalers are companies that operate computing infrastructure at a massive scale as their core business.
Currently, Google is the only hyperscaler that owns all five layers of the ecosystem in-house.
How an AI factory works
An AI factory functions like an assembly line.
Raw fuel is converted into energy, which is fed into a data centre, where it powers the chips that produce compute, the raw processing power AI models depend on.
Energy has emerged as the primary limiting factor for AI data centres, more so than the chips themselves.
A grid built in three pieces
The United States power grid is a complex, interconnected system made up of three main parts: the Western Interconnection, the Eastern Interconnection, and the Texas Interconnection, known as ERCOT.
It is managed by 66 separate balancing authorities, each operating under its own state regulations.
The grid works by matching supply and demand in real time, a process often visualised as a "generation stack", where power plants are ordered from cheapest to most expensive to run.
That stack includes nuclear power, coal, which is difficult to ramp up or down quickly, renewables such as solar and wind, and natural gas plants, which typically meet the bulk of overall demand.
Why your electricity bill moves
Electricity pricing is set by the cost of the "marginal plant", the least efficient and most expensive plant needed to meet total demand at any given moment.
Residential demand is dynamic, typically peaking between 2pm and 9pm as people get home and switch on appliances.
Data centre demand, by contrast, is static, running at a constant level 24 hours a day.
Because data centres never switch off, they push up electricity prices and add extra strain to grids that are frequently already running close to capacity.
A grid under pressure
That strain became visible in July, when a severe heat wave pushed electricity demand toward a record high across the Mid-Atlantic region, prompting the Department of Energy to declare a grid emergency for PJM Interconnection, the operator serving the area.
Partly in response to pressures like these, a growing number of states are introducing laws that require data centre developers to provide their own power sources, rather than drawing entirely on the shared grid.
The picture that emerges is one where the industry's next constraint may not be how many chips it can buy, but how much electricity it can actually get its hands on, a question with implications for the tax breaks and financial structures the next part of this series will examine.