FEATURE
A more robust power strategy should connect procurement with physical delivery; in different cases this may include dedicated renewable generation, battery storage, grid supply and, where appropriate, on-site generation. The aim is not to make each campus completely electrically self-sufficient, but to align supply more closely with the facility’ s operating profile and lessen strain on overloaded networks.
Flexibility must reflect the workload
The Government’ s proposed approach may require large data centres to provide greater demand flexibility, raising an important technical question: how flexible can an AI facility realistically be?
Storage can help answer part of it. Battery energy storage systems can absorb renewable generation, reduce peaks and support a facility over periods ranging from minutes to hours. UPS systems and supercapacitors operate on much shorter timescales and can respond to power-quality disturbances closer to the computing equipment.
They work best when the systems are coordinated. The energy management system can decide when batteries should charge or discharge, how UPS assets should respond and whether non-essential electricity demand can be reduced. Viewing each technology as a separate layer tends to create duplication and leaves useful flexibility unused.
Compute itself may also offer some room to manoeuvre, but the amount depends heavily on the workload. Certain non-critical tasks could be moved away from periods of network constraint. Battery charging could be delayed. Other processes cannot be shifted so easily.
AI training programmes may be tied to customer commitments or operational deadlines. A gridinteractive data centre therefore needs to be based on a realistic assessment of which loads can move and for how long.
It is essential to implement broad policy expectations carefully. Any requirement for demand flexibility will therefore need to reflect how individual facilities operate and which loads can realistically be shifted.
Water, cooling and location as considerations
Australia’ s proposed standards also place greater attention on water, an issue that becomes more important as AI changes cooling design.
Dense GPU environments produce concentrated heat loads. Liquid cooling is increasingly used to remove that heat efficiently, but the wider cooling system still determines the facility’ s overall energy and water performance.
Since all the methods involve making concessions, evaporative cooling can reduce
38
INTELLIGENT CIO APAC www. intelligentcio. com