Selective allocation and cooling efficiency
On September 29 and 30, local time, the Marina Bay Sands Expo and Convention Centre in Singapore was crowded with visitors from around the world attending Data Centre World Asia (DCWA), Asia's largest data center exhibition. More than 38,000 industry participants, including global technology companies, telecommunications operators, equipment manufacturers and investors, visited over the two days. The event drew attention from across Asia as companies sought a leading position in the AI infrastructure market.
How did Singapore, with its limited land and chronic electricity and water shortages, become an Asia-Pacific AI data center hub attracting global companies despite such unfavorable geographic conditions for construction? The explanation lies in its decision to move beyond simply expanding installed capacity toward a higher-value ecosystem that maximizes both infrastructure energy efficiency and computing efficiency within servers.
According to Singapore's Ministry of Digital Development and Information, Singapore had more than 70 data centers with combined capacity exceeding 1.4 GW as of February this year. More than 70 facilities are concentrated in a country slightly larger than Seoul. Singapore developed into a data center hub early, benefiting from submarine cable networks connecting Asia with major global markets and a low risk of natural disasters.
But rapidly increasing data center demand put greater pressure on limited land, electricity and water supplies. In 2019, the Singapore government introduced a moratorium temporarily halting new data center expansion. When it resumed selecting new developers in 2022, three years later, its solution was selective allocation.
A leading example is the Data Centre Call for Application (DC-CFA), a competitive application process for data center development. The government assesses energy efficiency, sustainability and economic contributions together, allocating scarce land and power capacity only to developers with verified high efficiency.
Singapore has also pursued technological innovation to address its physical constraints. With temperatures above 30°C throughout the year and humidity reaching 80%, it cannot use free cooling, which brings in cold outside air. Cooling is particularly challenging: cooling servers alone consumes around 40% of a data center's total electricity.
To address this, the Singapore government introduced a tropical data center standard in 2023 that allows operating temperatures to rise to 26°C or higher, reducing cooling energy consumption. It is also working with the National University of Singapore (NUS) and Nanyang Technological University (NTU) to support research into next-generation cooling technologies suited to tropical conditions, including liquid cooling, with a focus on improving data center energy efficiency.
As Singapore concentrated on upgrading its industry around high-efficiency data centers, the industrial effects quickly spread across the border to Johor, Malaysia. The region is becoming a base not only for large data centers but also for manufacturing and supplying related equipment. In July this year, global data center power and cooling infrastructure company Vertiv opened a manufacturing plant in Johor, while Delta Electronics opened an office in Johor Bahru the same month. A cross-border data center cluster linking Singapore and Johor is taking shape.
AI inference and computing efficiency
Global consultancy McKinsey & Company identifies Singapore's approach as a leading model for the future of the global data center industry. In an interview in Singapore on October 1, McKinsey Senior Knowledge Expert Lorraine Salazar said: "Singapore has turned its geographic and resource constraints into strengths, becoming a key AI hub in the Asia-Pacific region, which attracts the greatest global demand after the United States."
Salazar's assessment is that improving computing efficiency, alongside energy efficiency, will become central to competitiveness in the data center industry. "The basis of competition in the data center industry has now shifted completely from how many megawatts of installed capacity a facility has to how much AI computation it can perform with limited electricity, or computing efficiency," Salazar said.
According to McKinsey, global data center demand is expected to increase more than threefold in five years, from around 90 GW in 2024 to 278 GW in 2030. Related investment alone amounts to $7 trillion, or approximately 9,400 trillion won. AI inference is driving this growth. Its share of total data center demand is expected to rise from 29% this year to 42% in 2030. "While total demand triples, AI inference demand will increase 19-fold," Salazar said.
The rapid growth of AI inference is changing where data centers are located and how they operate. AI training consumes large amounts of electricity over relatively short periods as it processes vast datasets and optimizes models. Inference, by contrast, performs computations continuously, 24 hours a day, to respond in real time to requests from users worldwide. Cumulative electricity consumption during inference is estimated at four to nine times that of training.
Growing inference demand is increasing the importance of urban locations and major telecommunications hubs close to users. Yet these locations already have high electricity demand, and securing large sites and expanding power grids is difficult. This explains the rapid shift in the data center ecosystem from expanding installed capacity to maximizing computation per unit of electricity.
McKinsey's analysis is that, with transmission bottlenecks and other constraints making it difficult to increase electricity supply substantially in the short term, the only viable response is to maximize efficiency across both infrastructure and IT equipment. Salazar emphasized that competition over data center efficiency has expanded beyond infrastructure energy efficiency, measured by power usage effectiveness (PUE), to computing efficiency that maximizes processing output per unit of electricity.
"Efficiency improvements must work together across every area, not only by reducing physical electricity losses in transmission and distribution networks and cooling systems, but also in semiconductors and memory inside racks and the algorithms used by AI models," Salazar said. "Because improving energy and computing efficiency can directly reduce enormous electricity costs, greater efficiency is more than a matter of complying with government regulation. It is a fundamental economic incentive directly tied to developers' survival."
This reporting project received support from the Journalists Association of Korea and Climate Media Hub.
Original Korean article: Money Today
English brief by Korea Wind Intelligence, koreawindintel.com
