Data Center Capacity

A data center is a building full of computers that work around the clock for others. Video streams, online banking, corporate databases, and for some years now especially AI programs run there. Data center capacity indicates how much work is possible in total in such buildings. In practice, it is given in megawatts, meaning electrical power. One megawatt corresponds roughly to the continuous consumption of a thousand households. This measure is common because electricity and cooling set the actual limits, not the space on the shelf.

The power connection as the bottleneck of the AI industry

AI models require enormous computing power, both during training and in daily operation. Anyone wanting to offer such a model therefore needs access to free capacity. If none is available, even the best program is of no use. That is why capacities are today a scarce commodity that corporations compete over.

Building a large data center takes several years. Connecting it to the power grid often takes even longer, because lines and substations have to be built. Demand for AI, by contrast, grows within months. This gap between fast-growing demand and slow construction explains many of the recent headlines.

For the stock market, this topic is therefore highly relevant. When a corporation pours billions into new server halls, this changes its balance sheet for years. At the same time, chip manufacturers, construction companies, cooling technology providers, and power utilities benefit. Some analysts are already warning of overcapacity, meaning halls that will later hardly be used.

What is actually inside such a hall

Inside stand shelves, so-called racks, filled with servers. Servers are computers without a screen and keyboard that are controlled over the network. For AI, they contain special chips, mostly graphics processors, which carry out many computing steps simultaneously. A single such rack can draw as much power as a small residential building.

Every watt of electricity that flows into the chips ultimately turns into heat. This heat has to be removed, otherwise the computers shut down. Traditionally, this is done with air and large fans. With modern AI chips, air is no longer sufficient, so operators route water directly past the chips. This is called liquid cooling.

An important metric is the PUE value. It compares the total power consumption with the consumption of the computers alone. A value of 1.2 means: for every 100 watts of computing, an additional 20 watts are needed for cooling and technical equipment. Good modern facilities fall within this range, older ones considerably worse. In addition, there is redundancy through backup diesel generators and batteries, since a power outage would immediately stop all services.

From cloud bills to local debates

In everyday life, almost everyone uses this capacity without seeing it. A chatbot does not respond on the phone itself, but in a hall somewhere in the world. Anyone booking cloud services is renting nothing other than a share of it. Cloud simply means: computing power as a service over the internet.

In the news, the term often appears in investment announcements. Companies like Microsoft, Amazon, Google, or Meta then name sums in the billions and locations. Operators who build halls and rent them out to others also play a major role. In Germany, Frankfurt am Main is the most important location, because a central internet hub is located there.

Locally, the topic regularly leads to disputes. Data centers consume a lot of electricity, take up space, and need water or grid connections that are lacking elsewhere. Some cities therefore require that the waste heat be fed into district heating networks. A common misconception, by the way, is confusing capacity with storage space. What is meant is the ability to compute, not the amount of stored data.