Data centre electricity consumption across the region is expected to triple in the coming years, while power availability has overtaken land, capital, and connectivity as the single biggest barrier to new capacity.
New facilities that once averaged tens of megawatts now routinely exceed 100MW. Grids in the region's major hubs were never designed for this kind of load, and queues for new grid connections are lengthening.
At the same time, AI adoption across Asia Pacific is accelerating past the point of experimentation. Organisations are moving pilots into production and embedding AI into core operations, which puts a premium on infrastructure that can support growing AI workloads efficiently and sustainably.
These two forces, surging demand and constrained power supply, are converging on a single question: how do organisations scale AI in a region where every watt is contested?
Why power constraint is a bigger problem than it looks
A few years ago, an AI accelerator might have drawn anywhere from 250 to 300 watts. Today, leading accelerators can draw 1 kilowatt or more. An eight-GPU server now consumes roughly 15 kilowatts on its own, while most enterprise data centres across Asia Pacific might provision 15 to 40 kilowatts per rack in total. A rack that once held a dozen conventional servers may now hold one or two AI systems before hitting the power and thermal ceiling.
This creates a less visible constraint on AI expansion. Organisations may have floor space to spare in their existing data centres, but not enough power in the right locations. Across Asia Pacific, data centre markets are facing growing pressure on power availability as AI drives demand for higher-density computing. Retrofitting existing facilities also mean dealing with floor loading limits, the absence of a water loop, and switchgear specified for a different era of computing.
Alongside compute, power has now joined the list of things that stall AI programs, and neither can be solved by hiring or by better governance. When the supply of watts is effectively fixed for the next few years, the only remaining variable is how much useful work each watt performs.
- Kumar Mitra is Commercial Leader, Greater Asia Pacific, Lenovo.
AI’s next phase: from training to inferencing
Model training demonstrated what AI is capable of. Inferencing is where that capability begins creating everyday business value, by putting trained models to use on customer queries, fraud checks, or agent tasks executed at scale. Industry estimates say inferencing now accounts for 80 to 90 percent of all AI computing power used, and the International Energy Agency projects that electricity consumed by inference workloads will keep growing at 30 percent a year as adoption deepens.
Siting is where the two diverge. Training mega-clusters can be placed wherever power is cheap and plentiful, often far from population centres. Inference must live close to users, data, and devices, because milliseconds matter and because data residency requirements across much of Asia Pacific mean sensitive workloads cannot simply be shipped offshore.
Agentic AI compounds this, chaining together dozens or even hundreds of inference calls across different models, tools, and data sources to deliver a single business outcome.
Inferencing is therefore pulling AI into the region's own data centres, edge sites, and enterprise server rooms, reinforcing the shift towards Hybrid AI. Lenovo's CIO Playbook 2026, conducted with IDC, found that 84 percent of organisations expect to run AI across on-premises or edge environments alongside cloud deployments, placing each workload where it best balances latency, compliance, cost, and energy efficiency.
Measuring the real ROI on AI
That reality is changing how AI success is measured. Power Usage Effectiveness (PUE) remains important, but PUE describes the efficiency of the facility and says nothing about how much useful AI output is produced.
While chip supply will eventually catch up with demand, the same cannot be said about grid capacity, at least not on any timeline that supports deployment plans in the next year or two. The question then is no longer about how many GPUs an organisation can buy; but how much intelligence it can generate per megawatt it can secure.
That turns cooling into a board-level business decision. Every watt lost to inefficient cooling is a watt not spent on inference. In tropical Asia Pacific, where heat and humidity make cooling harder and costlier than anywhere else, the stakes are much higher.
If this sounds theoretical, consider how Singapore has allocated its next tranche of data centre capacity. Just recently, 200 megawatts of new capacity in total were awarded to Digital Realty, Equinix, Keppel Data Centres, and ST Telemedia Global Data Centres, under the second Data Centre Call for Application (DC-CFA2). The bar they had to clear was a PUE of 1.25 or better at full IT load, tighter than the 1.3 required in the earlier pilot; Green Mark Platinum certification; at least half of power drawn from eligible green energy sources; and compliance with a new national standard, SS 715:2025, designed to cut the energy consumption of data centre IT equipment by at least 30 percent.
For the first time, a government is regulating efficiency inside the rack, at the level of the servers themselves, and no longer only at the facility level where PUE lives. Efficiency is now the price of admission, and similar expectations are emerging across Asia Pacific as governments balance AI growth against sustainability commitments and energy security.
Matching the cooling method to the workload
The sensible approach is to match the cooling method to the workload and the facility. Direct liquid cooling comes in where rack densities have moved beyond what air can dissipate, while air or hybrid cooling remain appropriate for less demanding environments, for example. Planning this transition in phases as workloads evolve delivers lower costs, lower emissions, and greater operational flexibility, so that every available watt converts into intelligence.
In the training era, competitive advantage came from access to chips, and the winners were those who could secure allocation. In the inferencing era, advantage comes from infrastructure efficiency, because the organisations that generate the most intelligence per watt will be the ones that can afford to scale AI at all.
Whilst Asia Pacific cannot build new grid capacity overnight, organisations can engineer an infrastructure that makes better use of every available watt. Increasingly, that is what AI’s return on investment will be measured on: how efficiently organisations convert power into business outcomes.
Kumar Mitra is Commercial Leader, Greater Asia Pacific, Lenovo.
