The physical infrastructure of artificial intelligence is being constructed at a speed and scale that the electrical grid was never intended to support, in places like the flat industrial zones outside of Phoenix, the rural corridors of Virginia’s data center alley, and the periphery of towns in Iowa and Nebraska that the majority of people have never been to. The buildings are massive, low, windowless complexes that span several of acres.
The industrial cooling systems that are connected to them give the buildings the appearance of tiny power plants. which is becoming a fairly true description. Electricity consumption in modern AI data centers is at levels previously only found in large cities and heavy manufacturing. Between 50 and 150 megawatts can be continuously consumed by a single AI supercluster. 500 megawatts to over a gigawatt of sustained power demand is the goal of next-generation facilities that are being developed and announced in 2025 and 2026. One gigawatt. consistently. from a data center.
| Category | Details |
|---|---|
| Topic | Energy as the Primary Constraint on AI Scaling |
| Key Metric | Megawatts (MW) — replacing parameters as the binding constraint |
| Typical AI Supercluster | 50–150 MW (current); 500 MW–1+ GW (next-generation) |
| 1 MW Equivalent | Powers ~800 homes simultaneously |
| 100 MW Equivalent | Grid demand of a city of ~50,000 people |
| Grid Interconnection Delay | 5–10 years in some markets |
| Training Energy Scale | Thousands of megawatt-hours per large model training run |
| Industry Response | Tech companies becoming energy developers (nuclear, gas, on-site) |
| Key Companies Involved | Microsoft, Google, Amazon, Meta (data center buildout) |
| Reference Website | energy.gov/AI |
For the most of AI’s history, journalists, researchers, and tech executives used parameters—the internal numerical values that a neural network modifies during training and that dictate how it reacts to inputs—to describe model capabilities. There were 175 billion parameters in GPT-3. According to reports, GPT-4 had more than a trillion. Larger models are typically thought to be more capable models, at least initially, hence the term “parameter race” evolved to refer to the AI capability race.
When the primary limitation on what models could be constructed was mainly architectural and computational in the abstract sense—that is, the question of how to create training algorithms and model structures that scaled effectively—that framing made sense. The restriction has changed. How to create a better model is not the question at hand. Where to plug it in is the question.
Approximately 800 households may be powered simultaneously by one megawatt, which is equivalent to 1,000 kilowatts. A 50,000-person city’s worth of electricity is consumed by a 100 megawatt data center. These are descriptions of facilities that are currently in operation; they are not forecasts for the future. Thousands of megawatt-hours of energy are used during training runs for the biggest models spanning weeks or months; even ten years ago, this amount of energy would have been unimaginable for a technology business.
Furthermore, training is just one aspect of the issue. Inference adds a constant, daily energy requirement that grows with the user base rather than decreasing after training is finished. Inference is the process of running queries through a deployed model, which occurs billions of times every day across users of ChatGPT, Gemini, Claude, and similar products. AI’s megawatt demand is not a one-time capital expenditure. It’s a continuous operational reality.
This demand is being met by the electrical grid in ways that its planning frameworks were not intended to handle. The grid was constructed and expanded over the course of more than a century to serve residential, commercial, and industrial loads that followed somewhat predictable patterns. In certain markets, grid interconnection queues—the procedure by which a new facility requests and is granted permission to connect to the grid—now take five to ten years.
Even though a data center can be conceived and constructed in 18 to 24 months, it takes years longer to get the power connection it requires to function. This limitation, which is more directly related to the rate of AI scaling than any other single element, is a physical infrastructure issue rather than a computer science one.
The big tech firms have responded by approaching energy as a resource they create rather than a utility they buy. Nuclear power companies have relationships with Microsoft. Advanced geothermal initiatives are being funded by Google. Amazon has been purchasing natural gas production capacity and renewable energy assets.
The recurring theme is that hyperscalers are turning into energy firms because the energy isn’t available at the scale they want through traditional procurement channels, not because of their preferences or ideological commitments. If the grid can supply you with 200 megawatts in five years but you require a gigawatt, you can either create the remaining capacity yourself or forego building the data center.
Observing this change in the infrastructure choices being made in boardrooms and planning offices gives the impression that the AI industry is about to enter a phase where its growth is more limited by the physical world’s ability to supply electrons at the rate the models require than by intelligence, or the algorithmic and architectural questions that the research community is still working on. Mathematics and software can be used to boost parameters. Transmission lines, substations, generators, and the years needed to construct and obtain permits are all necessary for megawatts.
