The next bottleneck is power, data centers in the gigawatt range

The short version

• New data centers partly need several gigawatts of capacity, no longer a few megawatts.
• A partnership between a chip maker and an AI company provides for at least ten gigawatts of capacity and up to 100 billion dollars of investment.
• Delivery times for power transformers run to two to four years depending on size.

The bottleneck shifts

After chips and memory the next bottleneck in the AI build out is taking shape, and this time it is not about semiconductors. New data centers no longer need a few megawatts but partly several gigawatts of capacity, so as much as several modern nuclear power units.

The jump in the unit describes the matter more precisely than any growth rate. Between a few megawatts and several gigawatts lies the difference between a large building and a power station park of your own.

How concrete this is is shown by a partnership between a chip maker and an AI company. Data centers with at least ten gigawatts of capacity are planned. The chip maker intends to invest up to 100 billion dollars, tied to each expansion stage of one gigawatt. Alongside this there are negotiations about renting a plant in Ohio with ten gigawatts.

Exactly there lies the difference to other large consumers. A data center runs continuously and cannot be started up when there is wind and sun.

The difficult quantity is therefore not the amount of electricity over the year but the capacity at every point in time. A plant that draws around the clock demands a supply that stands ready every hour, and this requirement helps decide where it is possible to build at all.

The bottleneck stands in the substation

The bottlenecks reach further than the power station. Delivery times for power transformers run to two to four years depending on size.

A commitment of billions can therefore be given faster than the device can be procured that connects the plant to the grid in the first place. Two to four years of delivery time mean that an expansion stage decided today gets its connection only at the end of this period, regardless of how quickly chips and memory are available.

That also shifts the yardstick for progress. An announced capacity of ten gigawatts describes an intention, and a completed grid connection describes a plant that can compute.

What changes about the site

The consequence is a shift in the choice of site. Plants increasingly arise where connected capacity is available, for example at decommissioned power station sites, instead of where the users sit.

A decommissioned power station site brings with it what otherwise takes years, namely an existing link to the high voltage grid and the technology needed for it. The negotiation about renting a plant in Ohio with ten gigawatts follows this logic too, because what is being sought is not in the first place a building but a connection.

The scale of the investments stays high. The large cloud providers are likely to spend more than 600 billion dollars on AI infrastructure together this year.

This sum stands ready, but it does not dissolve any of the bottlenecks described. A transformer is not finished faster because more money is offered for it, and a grid connection does not arise sooner because the capital for it has already been committed.

Assessment

The debate about artificial intelligence turns on chips and models. The actual bottleneck has been somewhere else for months, namely in substations.

What I find notable is the reversal of the usual logic of scarcity. Capital is plentiful, what is missing is manufacturing time. Money can be moved in days, a winding hall for transformers cannot.

Anyone who wants to judge progress should therefore pay less attention to investment commitments and more to two sober quantities, namely delivery times and grid connections actually completed.

Frequently asked questions

How much power does a data center with ten gigawatts need

One gigawatt equals a thousand megawatts, and a large nuclear power unit typically delivers about one to one and a half gigawatts. Ten gigawatts of connected load therefore arithmetically correspond to the output of several power stations, and that around the clock.

Why are transformers the real bottleneck

Without a transformer no large consumer can be connected to a high voltage grid. Large units are one off builds and require special steel, copper and experienced specialists, which is why production cannot be expanded at short notice. Delivery times run to two to four years depending on size.

Why do new plants arise far from the users

Because the connected capacity determines the site. Plants increasingly arise where this capacity is available, for example at decommissioned power station sites, instead of where the users sit.

This text is not investment advice. It reports verifiable figures and puts them in context.

This analysis is for information only and is not investment advice.