Grid connection & energy supply

A big AI data centre needs a large, steady supply of electricity, and getting connected to the grid can take years, so power, not chips, often decides where and when one gets built.

Layer 1 · Physical & Energy · Build and run time · reviewed 2026-10-30

Read first: Power & cooling for AI data centres.

D1 · StoryFor young children and total beginners

Guess first: which takes longer, building the data centre or connecting its power?

A giant AI building needs as much electricity as a town.

The wires, power stations and big transformers that bring it there take years to build.

So companies often have to wait for power, or build where spare power already exists.

Getting power to a new AI building: the waiting is the slow part

Ask for power: from the power company
Wait in line: others asked first
Build new wires: and power equipment
Power on: the computers can start

Like: Building a new house is quick. Waiting for the water main to reach it is slow.

D2 · PictureFor curious beginners

A data centre asks the local grid operator for a connection of a certain size. The operator studies whether the lines and substations can carry it, and often finds that new transmission lines, substations or transformers are needed first.

Requests join a connection queue. New power stations queue too, so both the supply and the demand side can wait years. Grid parts such as large transformers have their own long lead times.

That is why builders hunt for sites with spare grid capacity, add on-site generation, or agree to cut their load when the grid is stressed. A building can be finished, and chips delivered, long before the power is ready.

From request to power on

Request: ask for capacity
Queue: wait your turn
Studies: can the grid cope?
Upgrades: lines, transformers
Power on: racks can run

Like: Building a new house is quick. Waiting for the water main to reach it is slow.
Where it breaks: A water main serves one street. A grid is one shared, balanced system: a big new load can affect prices and reliability for everyone on it, which is why operators study each request so carefully.

D5 · FrontierFor experts

D2 treats a grid connection as a queue to wait in. At scale the queue itself is unreliable, demand forecasts disagree, and the rules on who pays are being rewritten.

Speculative queues. Of the generation capacity that asked to connect in the US from 2000 to 2020, only 13% was operating by the end of 2025 and about 75% had withdrawn. Queue totals say little about what will actually be built (Berkeley Lab, 2026).

Forecast spread. One national estimate puts US data centres at 6.7% to 12% of electricity use in 2028. A range that wide makes it hard to size new lines and power stations (Berkeley Lab, 2024).

Who pays, and who can be curtailed. US regulators are asking whether large loads should pay the full cost of the grid upgrades they trigger, and whether loads that agree to cut power at peak times could be studied faster. The rulemaking is still open (FERC, docket RM26-4).

Examples as of 2026-10-30

PJM and ERCOT, US grid operators with long data-centre connection queues
EirGrid, Ireland's grid operator, whose 2026 rules ask new data centres to bring their own generation
Microsoft's deal to restart the Three Mile Island Unit 1 nuclear plant (now the Crane Clean Energy Center)
On-site gas turbines at large AI campuses

Sources

How it connects

Grid connection & energy supply delivers megawatts to Power & cooling for AI data centres.
Energy per token stretches each megawatt of Grid connection & energy supply.