Industry analysis identifies AI infrastructure as entering power-constrained era, requiring new grid-planning paradigms.
Headlines about artificial intelligence in Asia Pacific tend to focus on the largest numbers: bigger data centers, greater investment commitments, and ever-higher forecasts for regional spending on AI ambitions. Yet beneath these figures lies a more fundamental question: Can the region generate and deliver enough power to support those ambitions?
As AI scales, power—more than capital or chips—is rapidly becoming a key constraint, reshaping what infrastructure leadership means.
Demand itself is not in question. IDC forecasts data center power capacity across Asia Pacific, excluding Japan, to reach approximately 142,600 megawatts by 2029, while Gartner expects global data center electricity demand to double by 2030. The real bottleneck is supply. With grid connection wait times exceeding four years in major markets, access to reliable power has become a strategic differentiator. For enterprise leaders, this transforms what was once a facilities question into one of corporate strategy.
For the past decade, the response to AI's scaling demands was straightforward: add more accelerators, more racks, more capacity. But when power becomes the constraint, that approach backfires. The critical question shifts: How much value can each unit of power produce, rather than how much computing capacity can be deployed?
Performance per watt is emerging as one of the most important metrics for AI infrastructure. Efficiency stops being a sustainability footnote and becomes a practical factor determining how far an organization can scale with available resources. This requires a tangible mindset shift with real commercial implications. Electricity is already the largest operating cost for many data centers, so the efficiency of every server, rack, and cooling choice belongs in the boardroom, not just the server room. Organizations that extract more useful work from every watt will have greater room to grow. Those treating efficiency as an afterthought may find their options narrowing sooner than expected.
Policymakers are already enforcing this priority. In late 2025, Singapore's Economic Development Board and Infocomm Media Development Authority opened a second Data Centre Call for Application (DC-CFA2), directing at least 200 megawatts of new capacity to operators meeting demanding standards for energy overhead, equipment efficiency, and green power. Malaysia has tied data center tax incentives to energy-efficiency and emissions targets as part of its push for 70 percent renewable energy and net-zero emissions by 2050. Lenovo has also set a net-zero greenhouse gas emissions target for 2050, validated by the Science Based Targets initiative—a measurable commitment rather than a general ambition.
One of the biggest design mistakes is optimizing only for the immediate workload. Public discussion around AI focuses heavily on training large models, which are intensive but occur in bursts. As organizations move from experimentation to deployment, inference becomes the continuous workload: a model answering customers, checking payments, or guiding factory-line processes. The power per task may be smaller, but it is constant and accumulates daily. At a hypothetical cost of one cent per prompt, a service with 10 million users making 10 requests each day would spend approximately $1 million daily on electricity and computing.
JLL projects that inference will overtake training as the dominant data center AI workload around 2027, meaning this operating cost will continue rising as AI deployment matures. The answer is to treat architecture as a strategic decision and design for the entire life of a workload—choosing infrastructure based on performance per watt rather than headline speed, sizing systems for sustained real-world demand rather than peak days, and matching each workload to the environment where it performs best.
Where a workload runs is one of the most direct levers leaders have over cost, performance, and energy use, pointing toward hybrid approaches. Large training runs should reside where dense computing capacity and required cooling are available. Inference, by contrast, often sits closer to where data originates and decisions are made, including at the edge. This reduces latency, bandwidth use, and power consumption, while helping organizations address data-sovereignty and compliance requirements tightening across the region. There is no single correct environment—only the right environment for each workload.
Hybrid AI, spanning on-premises systems, the edge, and public cloud, is becoming the default enterprise architecture. Treating workload placement as a deliberate design choice rather than a default keeps both energy costs and response times in check as AI scales.
No single component can deliver this efficiency alone. Gains emerge when silicon, system design, software, and cooling are engineered to work together, with each element easing the limitations of the others. Progress depends on collaboration across the technology stack among chipmakers, infrastructure vendors, and deploying organizations.
Cooling has moved from the back of the data hall to the center of infrastructure strategy. Enterprises want more of their available power and budget supporting AI growth rather than simply removing heat. As AI racks run hotter, air cooling reaches its limits. Dense cooling with air requires more fans and greater reliance on chillers, adding energy costs beyond the computing equipment itself. Reducing cooling overhead is one of the fastest ways to improve data center efficiency.
Neptune warm-water cooling illustrates what this can look like. By circulating warm water directly to the hottest components, it reduces reliance on power-intensive fans and chillers. Lenovo reports that the technology can reduce data center power consumption by up to 40 percent compared with similar air-cooled systems, with power usage effectiveness as low as 1.1—below typical conventional levels. Such systems are already in use among meteorological agencies, universities, and research institutions across Asia Pacific.
When the entire technology stack is designed for efficiency, savings add up. There is an environmental dividend from lower energy use and lower emissions, helping organizations respond to disclosure expectations tightening across the region. But the more immediate reason efficiency is rising on corporate agendas is simpler: where power is constrained, efficiency is what keeps growth within reach, regardless of energy mix. The commercial and environmental cases point in the same direction.
Asia Pacific has one critical advantage: much of its AI-capable capacity is being built now, allowing efficiency to be designed into infrastructure from the beginning. Leaders that pull ahead will spend well rather than simply spend more. They will treat infrastructure as a series of efficiency and workload-placement decisions rather than as a race to add capacity. In the next phase of AI adoption, competitive advantage will come not from deploying the most infrastructure, but from extracting the greatest value from every watt. Where power, more than ambition, is the scarce input, organizations that make the most of every watt will be best positioned to scale.