How Compute Demand Outran the Grid

Start with the arithmetic, because the arithmetic is what makes this course necessary. A single modern AI training campus can draw hundreds of megawatts continuously, and the largest announced projects reach into the gigawatts — the scale of a mid-sized city or a nuclear station. A gigawatt is a billion watts of continuous draw. Ten years ago a large data center was 30 to 50 megawatts and utilities could absorb that inside normal load growth. Today a single operator may ask one utility for more capacity than that utility added in the prior decade. The grid was engineered for demand that grew one or two percent a year. AI compute is asking it to grow in step changes.

The grid cannot answer on that timescale, and it is worth being precise about why. Three things have to happen before a large new load can be served reliably. First, generation has to exist — actual power plants producing electrons. Second, transmission has to move that power from where it is generated to where the load sits, which means high-voltage lines and substations. Third, the interconnection itself — the physical and contractual tie between the customer and the grid — has to be studied, approved, and built. Each of these has its own queue, and the queues have lengthened as everyone tries to plug in at once. In several U.S. regions the interconnection study process alone runs four to seven years before a shovel touches dirt on the actual connection. That is Faraday's read on why the phrase 'just connect to the grid' has quietly become the hardest sentence in the industry.

The mismatch is structural, not a passing bottleneck. Transmission lines require rights-of-way, environmental review, and multi-state coordination; a single major line can take a decade from proposal to energization. New central-station generation — whether a gas plant or a nuclear unit — carries its own multi-year permitting and construction timeline. Meanwhile the compute buyer is working to an entirely different clock. A frontier model lab wants its next training cluster energized in twelve to eighteen months, because the competitive value of that cluster decays fast. When a buyer operating on an eighteen-month horizon meets a grid operating on a seven-year horizon, something has to give, and what gives is the assumption that the grid will be the power source at all.

This is the pressure that produced everything you will study in the rest of this course. Behind-the-fence generation — building your own power on your own land, next to your own load — is not a clever optimization. It is a response to a hard constraint. If you cannot get 500 megawatts from the utility until 2032, and you need it in 2027, you build it yourself. That single trade, repeated across dozens of projects and hundreds of billions of dollars, is reorganizing who supplies power to the digital economy.

It helps to separate two terms that get used loosely. 'Behind the meter' means the generation sits on the customer's side of the utility revenue meter, so the power never transacts across that meter — the customer produces and consumes it on the same site. 'Behind the fence' is the physical version of the same idea: the generating equipment is inside the data center's own fence line, colocated with the racks it feeds. In practice the two phrases point at the same strategy — self-supply — and this course uses them interchangeably, as the industry does. What matters is the economic consequence: the operator is no longer waiting in a utility's interconnection queue for its primary power. It has become, functionally, its own utility.

That reframing has consequences that ripple far beyond engineering. If the data center is now a power producer, then the skills that matter shift. Fuel procurement, emissions permitting, turbine maintenance, and gas pipeline access become core competencies for companies that a few years ago thought only about servers and cooling. Capital allocation changes too, because generation is a different kind of asset than IT equipment — longer-lived, differently financed, and subject to environmental regulation that never touched a server rack. Over the coming modules we will walk through the specific technologies being deployed, the oil-and-gas majors moving into the space, and the permitting and grid trade-offs that decide which projects actually get built. But hold onto the thesis, because it explains all of it: power is the first-order constraint on AI infrastructure, and the entire behind-the-fence movement is the industry's answer to a grid that cannot keep up.

One more framing before we go deeper. The constraint is not merely about total energy over a year; it is about firm, deliverable capacity available on demand, every hour, right where the racks sit. A training cluster does not tolerate the power wandering off when the wind drops. That distinction — between energy and firm capacity — is why so much of the behind-the-fence response leans on dispatchable sources like gas engines, turbines, and fuel cells rather than on intermittent renewables alone. Keep that in mind; it is the hinge on which the next module turns.