Data Centre Operators Confront Power Crunch as AI Loads Surge in Singapore

Singapore's data centre operators are turning to liquid cooling and on-site generation as AI workloads push power density beyond traditional limits.

SINGAPORE — Data centre operators in Singapore are racing to re-engineer their facilities as the power density of artificial intelligence workloads outpaces the capacity of conventional cooling and electrical systems. The shift is forcing providers to adopt liquid cooling, explore on-site generation, and rethink the way they allocate power across their campuses, according to industry executives and equipment suppliers.

The pressure is most acute in Singapore, where a moratorium on new data centre construction that lasted from 2019 to 2022 created a backlog of demand. When the government lifted the pause, it attached conditions requiring operators to demonstrate higher efficiency and to support Singapore’s economic priorities. Since then, AI has become the dominant driver of new capacity requests, with training clusters and inference servers demanding racks that draw far more power than the web-hosting equipment that defined the previous build-out.

Cooling becomes the bottleneck

The most immediate constraint is heat removal. Traditional air-cooled data halls were designed for racks drawing 5 to 10 kilowatts. AI servers equipped with high-end GPUs can draw 40 to 100 kilowatts per rack, and the next generation of accelerators is expected to push higher. Air cooling cannot keep pace at those densities without excessive energy use, prompting operators to evaluate direct-to-chip liquid cooling and immersion systems.

Several operators in Singapore have begun retrofitting existing halls with rear-door heat exchangers and direct-to-chip loops, while new builds are being designed with liquid cooling as the default rather than an option. Equipment vendors say the shift is changing procurement patterns: facilities teams now ask for coolant distribution units and leak-detection systems alongside conventional uninterruptible power supplies and switchgear.

The challenge is not only technical but financial. Retrofitting a live data centre with liquid cooling requires careful planning to avoid disrupting existing tenants, and the capital expenditure can be significant. Operators are weighing whether to pass costs through to customers or absorb them to retain anchor tenants.

Power procurement under scrutiny

Singapore’s limited land area and its reliance on imported natural gas make power procurement a sensitive topic. Data centres already account for a meaningful share of the country’s electricity demand, and the government has signalled that new facilities must demonstrate efficient use of power. Operators are responding by signing renewable energy certificates, exploring regional grid imports, and in some cases evaluating on-site generation using hydrogen-ready turbines or fuel cells.

However, on-site generation remains expensive and faces regulatory hurdles, particularly around emissions and fuel storage. Most operators expect the grid to remain the primary source of power, with energy attribute certificates and efficiency measures serving as the near-term compliance path.

AI workloads reshape design

The nature of AI workloads is also changing data centre design. Training jobs run for hours or days and produce relatively steady power draw, while inference workloads can be bursty. Operators say this variability complicates capacity planning and makes it harder to oversubscribe power the way they once did with less demanding enterprise applications.

Some providers are introducing power-smoothing software that shifts non-urgent workloads to periods of lower demand or caps accelerator power draw during peak hours. Others are designing halls with higher per-rack power envelopes and modular electrical distribution that can be scaled as tenants add GPUs.

Demand for AI capacity in Singapore shows no sign of abating. Cloud providers, AI startups, and enterprises building private inference systems are all competing for limited space and power. That competition is pushing operators to differentiate on efficiency, connectivity, and their ability to support high-density racks — factors that are becoming as important as location and price.

For now, the industry consensus is that the power crunch will not be solved by a single technology. Liquid cooling, better power management software, and more efficient accelerators will each contribute, but operators say the next few years will require sustained investment and closer coordination with regulators and energy suppliers. The data centre boom in Singapore is entering a phase where engineering discipline matters as much as land acquisition.