AI Data Centers: Flexible Power Consumption to Speed Up Grid Connection

Data centers for artificial intelligence are springing up faster than power grids can be expanded. The consequence: construction projects could be significantly delayed. Some projects are already building their own power plants and decoupling from the public grid. However, a pilot trial in Great Britain now gives operators hope that they could get their grid connection faster after all. The result refutes the need to plan for peak load.

The trial involved the British grid operator National Grid as well as the companies Nvidia, Nebius, Emerald AI, and EPRI. According to National Grid, they jointly demonstrated that AI data centers can reduce their power consumption significantly and at short notice – without AI calculations having to be interrupted during operation.

Power consumption reduced by a third

In the London data center Nebius AI Factory, the consumption of a 96-GPU cluster (Nvidia Blackwell Ultra) was reduced by up to 40 percent within one minute after a signal from grid control. In total, the situation was tested around 200 times in December 2025. In another test, it was proven that the load could be reduced by a tenth over a period of ten hours.

Previously, data center operators applied for grid capacity based on their theoretical maximum demand. Grid operators then have to provide peak load permanently. The results of the tests could lower this requirement to a lower level. Currently, new construction projects face a waiting time of 10 years for a grid connection. By 2030, data center capacity is planned to increase by 6 gigawatts. National Grid hopes to reduce the waiting time to two years.

What happens next

Flexible data centers could also help absorb surplus electricity from renewable energies. However, the disadvantage for operators is that they have to calculate with fluctuating processing power of their servers. The results are now to be presented to regulators and politicians to enable new connection rules.

(mki)