A conventional data center draws almost constant power. An AI training hall does not. Thousands of GPUs compute, then pause together to exchange results, and start again, several times a second, all day long. The power the hall draws rises and falls at the same pace.
This behaviour is new for power grids. It now matters to grid operators as much as to data center designers, and it is decided very early in a project.
WHAT HAPPENS IN THE HALL
A widely shared worked example gives the order of magnitude: a hall of 2,304 GPUs, in 32 racks of about 130 kW, can see nearly half its load, about 2.3 MW, appear and disappear in under a second, up to three times a second. The size of the swing is not the real problem: its repetition is. Each swing travels through the installation, from racks to UPS, transformers and substation, up to the grid.
Disconnection on voltage dips: a brief grid disturbance makes many sites switch to UPS or generators at once, and the grid suddenly loses a large load
WHY GRID OPERATORS ARE PAYING ATTENTION
Two incidents are often cited. In July 2024, in Northern Virginia, a voltage fluctuation caused about sixty data centers to disconnect simultaneously, around 1,500 MW of load lost at once. In May 2025, in Ireland, a remote transient fault caused 387 MW of data center load to drop instantly, and the system operator had to activate emergency stability measures.
In the United States, the body responsible for grid reliability (NERC) has since issued an alert asking grid operators to characterise and manage these behaviours, citing significant oscillations that occur within seconds.
WHAT THE FRENCH FRAMEWORK PROVIDES
In France, grid-connected consumption installations must meet technical requirements set by EU Regulation 2016/1388 (the demand connection network code) and by the arrêté of 9 June 2020 on technical requirements for design and operation for connection to electricity networks. These texts require installations not to disturb the grid or other users: harmonics, flicker, voltage jolts. The connection agreement then sets, site by site, the expected characteristics and performance of the installation.
These rules were written for conventional industrial consumers. How they apply to AI loads, which oscillate constantly and react together to disturbances, is being studied. Since October 2025, RTE, Data4 and Schneider Electric have been running a demonstrator on Data4's Marcoussis campus, presented as the first in Europe, analysing how a data center behaves under voltage variations and disturbances, partly to inform future technical connection requirements.
For a developer, the practical consequence is simple: expect the site's electrical behaviour to be examined more closely in the connection request, and requirements to become more precise in the coming years.
FOUR WAYS TO TAME THE OSCILLATIONS
Battery storage sized in power rather than energy, able to respond within milliseconds
Storage at rack level: capacitors and new power shelves that absorb the fastest swings
Software smoothing of GPU load, at the cost of some wasted computing
Add to this voltage-dip ride-through: UPS, transfer switches and protections must be set so the whole site does not disconnect at the slightest disturbance. The solutions combine, and none is free: floor space, fire safety and cost for batteries, computing performance for smoothing, negotiation for the contractual limit.
WHAT TO DECIDE, AND WHEN
These choices are made before design freeze. Afterwards they become expensive: a battery room nobody drew, UPS to re-specify, transformers sized too tight.
What smoothing strategy, and at what level: grid, hall, rack or software?
Are UPS, transformers and protections specified for these swings and for voltage dips?
Was the load profile described to RTE or Enedis in the connection request?
For an investor, lender or insurer, the same questions show whether an AI project is designed to run without incidents. See also our risk surveys for insurers and battery storage.
Typical situation: our 4 MW AI hall keeps tripping our UPS, is it the grid or our installation?
SOURCES
Regulation (EU) 2016/1388 establishing a network code on demand connection; arrêté of 9 June 2020 on technical requirements for design and operation for connection to electricity networks; CRE, "Les prescriptions techniques"; Data4, announcement of the Marcoussis demonstrator with RTE and Schneider Electric, 29 October 2025; Schneider Electric, "Making data centers grid friendly", February 2026; Energy-Storage.news, on NERC's Level 3 alert.
FREQUENTLY ASKED QUESTIONS
Why does an AI hall's consumption oscillate?
During model training, GPUs alternate between computing and exchanging results, in sync. The whole hall speeds up and slows down together, several times a second.
Does RTE limit a data center's load variations?
Technical connection requirements forbid disturbing the grid, and the connection agreement sets the expected characteristics of each installation. Requirements specific to AI loads are being studied, notably through the Marcoussis demonstrator.
Do batteries solve the problem?
They help if sized in power and able to respond within milliseconds. They are usually combined with rack-level storage, software smoothing and voltage-dip ride-through settings.
NEXT STEP
Holding an exploratory study, a PTF or a connection agreement? We review it independently before you sign or invest, or prepare the PTF request with you. The expert: Xavier Watrelos.
THE GRIDREADINESS MONTHLY LETTER
Grid connection, equipment and data center projects in France: what matters, once a month.
Your email is used only for this letter. Legal notice.
Thank you, you are subscribed.
Sending failed. Email fred@gridreadiness.com.