As Singapore’s Minister-in-charge of Energy and Science and Technology Tan See Leng warned at the Singapore International Energy Week 2025, Singapore’s electricity demand is poised to climb sharply, driven by energy-intensive loads such as data centres and electric vehicles.
And we can see why.
AI is changing not just the way we compute, but the way we consume power. The exponential rise of AI workloads has transformed data centres from relatively predictable facilities into dynamic, high-density AI factories, industrial ecosystems of servers and silicon drawing electricity at unprecedented speed and scale.
Across the globe, power demands for data centres are rising faster than most grids were designed to handle. Rack densities for data centres have more than doubled, from roughly 8 kW to 17 kW per rack in just two years, and are projected to reach 30 kW or more by 2027. Singapore, one of Southeast Asia’s most concentrated data-centre hubs, may see its data centres consume 12% of national electricity by 2030, up from roughly 7% today.
Yet the challenge ahead lies not only in scaling supply, but in managing the erratic, high-intensity load patterns that AI computing introduces.
Millisecond surges
Unlike conventional computing, AI workloads operate in bursts. When large GPU clusters spin up simultaneously, demand can spike within milliseconds, far faster than legacy systems were designed to accommodate. The result is shockwaves through the grid: transformers heating up, protection systems tripping, and voltage fluctuations cascading faster than operators can respond.
This is why grid resilience can no longer be defined by average demand curves. It must now account for the extremes: the surges, the synchronisation of workloads, and the unpredictable interplay between digital acceleration and physical infrastructure.
Data centres, too, must evolve. Through grid-interactive battery storage, advanced uninterruptible power supplies (UPS), and intelligent energy-management software, data centre campuses can absorb load volatility, store renewable power when it is abundant, and discharge it during grid peak-load periods. Instead of adding stress, they can act as balancing nodes in the wider power ecosystem.
Beyond power: cooling, water and space
Power demand is only one part of the equation. Every watt consumed by AI eventually turns into heat, and managing that heat has become an escalating challenge. At the same time, the increasing density of modern workloads is also pushing traditional air-cooling systems to the edge of what they were designed to handle.
To address this challenge, the industry is steadily moving toward liquid-cooling systems, whether direct to chip or immersion, to handle rack densities exceeding 40 to 50 kW. These systems are more efficient, but they require new plumbing, heat-transfer infrastructure, and water-quality management. In water-scarce regions, that adds another layer of complexity: how to cool sustainably without undermining environmental goals.
Space is another challenge in land-scarce markets like Singapore, Tokyo, and Hong Kong. While liquid cooling enables higher computing density per square metre, which is critical in urban environments, it also increases the need for grid stability and heat-recovery systems to prevent localised thermal stress.
Transforming the transformer
Meeting this new paradigm requires infrastructure that can respond as fast as AI operates. One option is advanced algorithms implemented in modern UPS, and we can also observe developments in solid-state power-electronics technologies, digital successors to traditional copper-and-iron power infrastructure systems.
Using semiconductor power electronics, these technologies actively manage voltage and current in real time, smoothing millisecond load swings and enabling bidirectional power flows that integrate with on-site renewables and storage. This allows facilities to manage volatility at the source rather than passing it upstream to the grid.
For fast-growing economies that are balancing AI investment with decarbonisation goals, such technologies offer a path to smarter, more responsive grids.
Built to adapt in real time
As AI continues to reshape demand patterns, the focus must shift from expansion to agility. The way forward is to design our grids for resilience, anticipating stress, routing energy intelligently, and balancing every node from battery storage to GPU clusters in real time, which also requires protecting the grid from GPU volatility.
But technology alone will not be enough. Collaboration between utilities, policymakers, and operators will determine how fast we adapt. APAC, with its rapid digitalisation and industrial growth, is both the test bed and the proving ground for this shift.
In the end, how we plan, coordinate, and invest today will determine whether AI accelerates progress or strains the systems that sustain it. The test of AI will not be in algorithms alone. It will be in the substations, transformers, power electronics, and cooling loops that keep them operational. AI has made the power grid its next frontier. Whether it drives sustainable progress or deeper strain will depend on how quickly digital ambition can align with physical reality, and whether power systems can be built to sustain the future being created.
