Gigawatt Data Center

A data center is a building full of computers that work around the clock for others. When you make a search query or stream a video, a device there computes on your behalf. Such facilities require a great deal of electricity, which is why their size is often measured not in square meters but in watts. Watt is the unit of power, meaning the electricity consumed per second. A gigawatt is a billion watts — roughly the output of a large nuclear power plant. A gigawatt data center, then, is a facility whose computers together draw as much electricity as a city with a million inhabitants.

Why electricity became the most important ingredient in the AI industry

In the past, the question about a data center used to be: How many servers fit into the building? Today the question is: How much electricity does the grid deliver to this location? That’s because the specialized chips used for AI consume many times more power per unit than a normal server. Space is no longer the bottleneck — the power line is.

That’s why companies like Microsoft, Google, Amazon, or OpenAI now speak openly about power plants. Some sign long-term contracts with nuclear power plant operators. Others build their own gas power plants right next to the server halls. For investors, this is one reason why utilities, grid operators, and turbine manufacturers are suddenly considered AI winners.

At the same time, conflicts are emerging. A new gigawatt data center can drive up electricity prices in a region. Water is also an issue, since many facilities rely on it for cooling. In the US, there are already communities that have blocked such projects through votes.

What’s inside such a hall

The centerpiece consists of tens of thousands of graphics processing units, or GPUs for short. These are chips that carry out huge numbers of simple computational steps simultaneously — exactly what training AI models requires. They sit in servers, and the servers sit in cabinets called racks. A single modern rack can draw over 100 kilowatts, as much as several dozen households.

To let the chips work together on a single task, they’re connected via extremely fast networks. A large AI model doesn’t fit onto a single chip. It’s distributed across thousands of chips that must constantly exchange intermediate results. If the network is too slow, expensive chips just sit idle.

The second major factor is cooling. Nearly all the electricity consumed is ultimately converted into heat. At gigawatt scale, fans are no longer sufficient. That’s why facilities rely on liquid cooling: water or a cooling fluid runs through pipes directly past the chip. On top of that come transformers, backup diesel generators, and batteries, so that a power outage doesn’t immediately halt operations.

These are the numbers that show up in the news

In business reports, the size of AI projects is now usually given in gigawatts. OpenAI and Oracle's Stargate project, for example, is described in such units. When a company announces a data center of five gigawatts, what’s meant is an entire site with multiple buildings that grows over the course of years. No single building reaches that figure.

The key distinction is between announced and connected. An announcement costs nothing, while a grid connection often takes five to ten years. Many projects will therefore never operate at the planned scale. Anyone reading such reports should pay attention to the word that precedes the gigawatt figure: planned, under construction, or in operation.

A common misconception is that all this electricity goes into training new models. The larger share now goes into ongoing operation — that is, into answering user queries. Each chatbot response consumes little power, but it happens billions of times over. That’s why these facilities run continuously, not just during individual training runs.