Gigawatt scale (data centers)
We speak of gigawatt scale when a single data center site needs as much electricity as a major city – roughly one billion watts. Such facilities are currently being built for training large AI systems and are turning electricity into the industry's most critical bottleneck.
A data center is a building full of computers that compute for others around the clock. Each of these computers consumes electricity, and this consumption is measured in watts. An electric kettle draws about 2,000 watts, or two kilowatts. A gigawatt is one billion watts, or as much as half a million kettles running simultaneously. We speak of gigawatt scale when a single site draws power in that order of magnitude. That corresponds roughly to the consumption of a city with around one million residents and roughly the output of a large nuclear power plant unit.
Electricity as the AI industry’s new bottleneck
Until just a few years ago, a large data center with 30 or 50 megawatts was already remarkable. A megawatt is one thousandth of a gigawatt. Today, companies like Microsoft, Amazon, Google, Meta, and OpenAI are announcing facilities that individually are meant to require several gigawatts. The reason is the training of modern AI models: it can be distributed across very many chips, and more chips mean better results. The competition for more powerful models is thus becoming a competition for electricity.
As a result, the bottleneck is shifting. For a long time, the scarce resource was the AI chip itself, especially Nvidia's graphics processors. Now, operators report that they have chips but no location with a sufficient grid connection. A new connection to the high-voltage grid can take five to ten years to get approved in the US and Europe. Anyone wanting to build faster has to plan for their own power plants or gas turbines as well.
For investors, this is the reason why utilities, grid operators, turbine manufacturers, and cable producers are suddenly being traded as AI stocks. A gigawatt site costs roughly 30 to 50 billion dollars, with the largest share going to chips. Sums of this magnitude change electricity prices and investment plans for entire regions.
What actually stands in such a campus
A gigawatt site is rarely a single building. What’s typical is a campus made up of several halls that are completed in stages. Inside the halls stand shelves, so-called racks, densely packed with servers. A modern AI rack can draw 100 kilowatts or more, whereas a classic server rack draws only five to ten. This densification is the real leap.
All that electricity ultimately turns into heat. Air cooling is no longer sufficient at such densities, so liquid cooling is used: water or a coolant flows directly over the chips. In addition, there are substations, transformers, backup diesel generators, and huge battery banks that bridge second-long gaps in the grid. How efficiently a site operates is described by the PUE metric: it indicates how much electricity flows overall compared to the electricity that actually reaches the servers.
A common misconception is that a gigawatt data center constantly consumes a gigawatt. The figure usually describes the maximum connected capacity. However, when training large models, utilization is in fact very high and very steady, unlike classic cloud services, which fluctuate between day and night.
Where gigawatt reports turn up
In business news, you encounter this term almost daily. Projects like Stargate in Texas, Meta’s facility in Louisiana, or xAI’s site in Memphis are consistently described in gigawatts, not in square meters or server counts. Power has become the currency in which data centers are compared.
The topic is also politically present. In Ireland and parts of the Netherlands, there are construction halts on new data centers because the grid is reaching its limits. In the US, residents are arguing over who bears the costs of grid expansion. At the same time, operators are signing long-term contracts for nuclear power in order not to miss their climate targets.
In everyday life, you notice this indirectly: a chatbot answers you from such a facility, and the operating costs are baked into subscription prices for AI services. To be clear: gigawatt scale describes the infrastructure, not the model size. A model is measured in parameters, a data center in watts.