An emerging challenge for electricity system operators is “the increasing speed at which electricity demand can vary within modern data centres”, according to a new International Energy Agency (IEA) report, ‘Key questions on energy and AI’.
The report said that although data centres are often viewed as stable, near-constant baseload electricity consumers, in fact “AI and other compute-intensive workloads can trigger steep and sudden load ramps as clusters of GPUs transition synchronously between tasks”. This can cause demand to swing within fractions of a second, and the report highlights a case in which load in a data centre block focused on AI training increased from 6MW to 30MW in 0.25 seconds. This is far faster than onsite technologies such as diesel generators or gas turbines can respond, while for grid-connected centres buffering would be required to stop the swing propagating to the grid and affect power quality.
The report notes that data centre operators may not have visibility of demand swings if they are caused by customer usage.
The report said grid operators have begun publishing ramp‑rate requirements for large loads and operators are increasingly deploying hybrid control systems that combine uniterruptible power system (UPS) units, battery energy storage systems, battery backup units on rack and capacitors to absorb rapid load swings. The report said such hybrid systems “are likely to play a crucial role inenabling onsite systems to overcome the operational challenges of meeting data centre
loads”. There are also emerging solutions at the GPU and rack levels, although they may involve wasteful “compute burn”.
The report said onsite batteries provide fast‑acting reserve capacity and have many roles during operation, including managing controlled shutdowns, absorbing short‑duration fluctuations, or enhancing power quality by providing frequency support and harmonic mitigation. As a result, the report projects that UPS capacity will rise by more than 100GW by 2030, with longer-duration battery energy storage in data centres rising from around 5GW today to 20-25GW by 2030. It says, “data centre battery storage could become an important grid asset if the incentives are right, helping both to smooth internal load swings and support the flexibility of the grid”. It says data centres can be flexible in three ways.
First, onsite power assets such as batteries or gas turbines can be used to reduce data centre loads at times of peak demand. Second, auxiliary power consumption, notably for cooling, can be reduced. Third, data centre operations (also known as workloads) could be re-routed geographically or deferred in time. However, it says “numerous challenges must be overcome to make data centre operations more flexible, particularly for the third option.”
The report also projects installed onsite gas‑fired generation for data centres could expand rapidly to up to 27GW globally by 2030.