Translated by ChatGPT
For Subscribers Only
https://www.zaobao.com.sg/news/world/story20260719-9361201?utm_source=android-share&utm_medium=app
19 July 2026
Lianhe Zaobao
Compiled by International News Reporter Tan Jie Ming (陈婕洺)
=====
When the server rooms, servers, and cooling systems of an artificial intelligence (AI) data center are all ready, the costly facility may still remain an unusable "empty shell." Even when everything appears to be in place, what invisible hurdle could prevent an entire data center from being activated, forcing it to wait years before it can begin operating?
=====
When you open the ChatGPT chatbot late at night and type a prompt asking it to plan a long-awaited overseas trip for you, it takes only a few seconds before a detailed itinerary and hotel recommendations appear on your screen. A typical text query like this consumes about 0.3 watt-hours of electricity—roughly equivalent to the amount of electricity a household oven uses in half a second.
Viewed individually, the electricity consumed by a single query seems insignificant. However, when hundreds of millions of prompts pour in continuously, the servers and cooling systems behind the screen bear a tremendous energy burden.
According to data from the International Energy Agency (IEA), global electricity consumption by data centers grew by 17 percent in 2025, far exceeding the approximately 3 percent increase in overall global electricity demand. The IEA forecasts that global electricity consumption by data centers will rise from approximately 415 terawatt-hours in 2024 to around 945 terawatt-hours by 2030—an increase of more than double.
Based on Singapore's annual electricity consumption of about 58 terawatt-hours in 2024, the projected annual electricity consumption of global data centers by 2030 would be equivalent to approximately 16 years of Singapore's nationwide electricity demand.
Over the past 30 years, electrification has continued to expand globally, and by 2023 nearly 92 percent of the world's population had access to electricity. Reliable electricity has become an everyday part of modern life: flip a switch, and power is simply expected to be there.
Now, the artificial intelligence (AI) boom has pushed electricity back to the forefront of geopolitical competition and technological rivalry. Since the explosion of generative AI at the end of 2022, market attention initially focused on chips. However, over the past two years, as data centers have expanded rapidly, the capacity of substations and transmission lines—and whether projects can actually be connected to the power grid—has become the new focal point in the competition over AI infrastructure.
Professor Wen Yonggang, President's Chair Professor at the College of Computing and Data Science at Nanyang Technological University, told Lianhe Zaobao that electricity has now become the first hurdle determining whether a data center project can proceed.
Wen said that 10 years ago, companies selecting sites for data centers typically looked first at land availability, fiber-optic networks, transmission latency, and tax incentives. Electricity supply was generally taken for granted. Today, the situation has reversed. Developers must first determine whether they can obtain hundreds of megawatts of stable, affordable electricity within a reasonable timeframe before considering other factors.
"Chips determine how fast you can compute; electricity determines whether you can compute at all... If a data center cannot begin operations because it cannot be connected to the power grid, then no amount of tax incentives will help."
Competition for electricity among data centers will intensify further as existing facilities are upgraded.
AI's Electricity Needs: Stable, Low-Carbon and Affordable
Dr. David Broadstock, Partner at Asia-Pacific energy consultancy The Lantau Group, said in an interview that data centers continuously upgrade their computing equipment throughout their operational lives. Even if the building itself remains the same size, electricity consumption may increase significantly, causing power demand to grow faster and become more difficult to predict.
"More accurately, the electricity that can be supplied immediately today is insufficient to meet the future electricity demand of data centers. Power infrastructure can be expanded gradually, but the key question is whether expansion can keep pace with the growth in electricity demand from data centers."
However, Wen Yonggang pointed out that this does not mean the world has run out of electricity. Rather, there is a mismatch between electricity supply and demand in terms of timing, location, and supply conditions.
"In terms of timing, data centers and computing infrastructure are expanding rapidly, while power infrastructure cannot keep up at the same pace. In terms of location, data centers generally want to be close to users and fiber-optic networks, but abundant, inexpensive clean energy is often located farther away. Moreover, AI does not require just any form of electricity—it requires electricity that is stable, affordable, and low-carbon."
Investment in Power Generation and Supply Continues to Rise, But Grid Construction Has Fallen Behind
A major reason for the mismatch between electricity supply and demand is that although global investment in power generation continues to grow, grid construction has not kept pace.
The International Energy Agency notes that since 2015, around US$1 trillion (approximately S$1.3 trillion) has been invested annually in power generation facilities worldwide, while investment in electricity grids has risen only to about US$400 billion, with a growth rate of less than half that of power generation investment.
As a result, although power generation capacity has increased, electricity cannot necessarily be delivered promptly to areas with the strongest demand. According to the IEA's Electricity 2026 report, more than 2,500 gigawatts of renewable energy, energy storage, and large electricity-consuming projects—including data centers—are currently waiting to be connected to power grids worldwide.
Laura Cozzi, Director of Sustainability, Technology and Outlooks at the International Energy Agency, told Lianhe Zaobao that although data centers currently account for only about 1.5 percent of global electricity consumption, the latest generation of projects is so large that it has already begun placing pressure on local power grids. In many electricity markets, the grids are already congested. Connecting new projects takes time, while constructing new transmission lines may require many years.
Cozzi said, "Data centers can typically be completed within 18 to 24 months, but grid construction may take as long as 15 years."
Xie Weikeng, Head of Media and Publishing at the Asia Artificial Intelligence Association, said that power supply gaps are particularly evident in countries such as the United States and Canada, where electricity grids were built earlier. Much of the power infrastructure has been in use for many years, urban land is limited, and expansion requires lengthy planning, approvals, and coordination, making it difficult to increase electricity supply quickly in the short term.
Xie also noted that in some regions where development started later, power grids could be planned from the outset to accommodate large electricity-consuming projects such as data centers, potentially avoiding many of the difficulties associated with upgrading older grids. However, actual conditions still depend on whether sufficient land, funding, and supporting infrastructure are available, so no broad generalization can be made.
The International Energy Agency warns that if grid bottlenecks cannot be alleviated, by 2030 approximately one-fifth of the world's planned data center capacity could be delayed because of the inability to connect to power grids in time.
Can AI Use Less Electricity While Doing More?
Since expanding power generation facilities and electricity grids takes many years, and technologies such as space-based computing power and large-scale long-duration energy storage remain immature, one of the more practical solutions at present is to tackle the issue from the demand side—enabling AI to accomplish the same tasks with less electricity during both training and operation.
Wen Yonggang believes that pressure on electricity supply and demand will not stop AI development. Instead, it will force the entire system to become more efficient. Companies will improve not only models and chips, but also adopt more efficient cooling technologies and data center designs. They will also schedule certain training tasks at times and in locations where clean electricity is more readily available.
When selecting models, not every application needs to invoke the largest language models.
Xie Weikeng pointed out that for clearly defined and less complex tasks, companies can use smaller, more specialized models. Large models can also activate only the components most relevant to a given task. This is like assigning work to the most suitable specialist team instead of mobilizing the entire organization, thereby reducing unnecessary computation and electricity consumption.
Besides choosing the right model, it is also important to avoid having AI repeatedly perform the same task.
Zhang Fan, co-founder of Singapore-based Advantage Research Consulting, said that some AI agents repeatedly call models, consuming large numbers of unnecessary tokens and computing power. Therefore, reducing AI's electricity consumption requires not only selecting suitable models, but also optimizing workflows to reduce unnecessary calls. Chinese large language models such as Doubao, Tongyi Qianwen, and Tencent Hunyuan have already begun making such optimizations.
However, Broadstock cautioned that although more extensive training or higher electricity consumption during the initial stage may consume more power, it may also enable the model to operate faster during later use. Therefore, evaluating AI energy efficiency should not focus solely on the training stage but should also consider subsequent use. At the same time, AI performance should not be fundamentally reduced or innovation hindered simply because of electricity constraints.
On the other hand, besides reducing its own electricity consumption, AI can also help the energy system save electricity and improve efficiency.
The International Energy Agency notes that energy companies have begun using AI to forecast and integrate solar and wind power generation, allowing these energy sources to be connected to power grids more effectively. AI can also assist with electricity dispatching, identify equipment faults and maintenance needs in advance, and enable existing energy infrastructure to operate more reliably and efficiently.
For example, AI can identify and locate grid faults more quickly, reducing power outage durations by about 30 to 50 percent. Combined with remote sensors and intelligent management systems, AI could also release up to 175 gigawatts of transmission capacity from existing power grids without constructing new transmission lines.
The International Energy Agency estimates that if AI is widely used in power plant operations and maintenance, it could save up to US$110 billion annually by 2035.
Who Will Ultimately Pay for the Surge in Data Center Electricity Consumption?
Data centers require expanded power generation facilities, substations, and transmission lines. One of the public's biggest concerns is who will ultimately bear these additional costs—technology companies, power companies, or ordinary consumers.
Some parts of the United States have already experienced situations where data center expansion has driven up electricity costs for traditional businesses.
PJM, which serves 13 U.S. states, is the country's largest regional grid operator. To ensure sufficient electricity supply during peak demand periods, PJM pays power generators in advance to reserve generation capacity that can be activated when needed. This payment for ensuring electricity is available when required is known as the capacity price.
In recent years, the rapid increase in data centers within the PJM region has caused electricity demand to grow faster than generating capacity. To ensure sufficient future electricity supply, PJM has had to purchase reserve generating capacity from power producers at much higher prices. As a result, the capacity price has risen from US$28.92 per megawatt-day in 2024 to US$329.17—an increase of more than tenfold.
The problem is that these additional costs are not charged solely to data centers. Under PJM's pricing mechanism, users within the same electricity market share the costs, which are reflected in the capacity charges they each pay. As a result, manufacturers and businesses already operating in the region also have to bear the higher costs.
Belden Brick, a long-established brick manufacturer in Ohio, is one such example. Although it has nothing to do with the AI industry, its monthly capacity charge has risen from US$1,600 to US$12,000, while its total electricity bill has increased by about 90 percent within a year.
This case demonstrates that the higher regional electricity supply costs driven by data center expansion may be passed on to other businesses through electricity market pricing mechanisms. If businesses cannot absorb these costs over the long term, they may ultimately pass them on to consumers through higher prices for goods and services.
Will Data Center Expansion Push Up Singapore's Electricity Prices?
Singapore's electricity market differs from that of the United States, but similar cost issues are equally relevant.
Singapore has stable infrastructure, international connectivity, a sound legal system, a mature cloud services ecosystem, and a large base of financial institutions and corporate users, making it a regional data center hub. However, limited land, high energy costs, and the tropical climate's cooling requirements also raise the barriers to further expansion.
According to data from construction consultancy Turner & Townsend, the construction cost of data centers in Singapore increased from US$11.40 per watt in 2023 to US$14.50 per watt in 2025—an accumulated increase of about 27 percent over two years, second only to Tokyo among the world's major markets.
In addition, although Singapore's current electricity system is generally reliable, approximately 95 percent of its electricity is generated from imported natural gas. Fluctuations in international fuel prices and supply disruptions could affect local electricity prices and energy security.
Zhang Fan pointed out that whether the additional electricity costs arising from increased data center demand are ultimately passed on to residential electricity bills and everyday prices depends largely on how each jurisdiction differentiates and prices industrial and residential electricity.
"In some U.S. markets, electricity prices are relatively market-driven. If industrial and residential electricity are not clearly separated, or if their pricing is linked, higher industrial electricity costs may be passed on to household electricity bills relatively quickly. In most Asian markets, however, industrial and residential electricity are priced separately. Therefore, the additional electricity costs brought by data centers are usually borne by businesses first and do not necessarily show up directly in household electricity bills."
AI expert Xie Weikeng believes that some of the additional costs will most likely eventually be passed on to consumers, as this is the normal way costs are transmitted in many industries. However, data centers also stimulate economic activity and enhance economic potential. Therefore, when evaluating the impact of rising costs, it is also necessary to assess whether they bring corresponding investment, productivity gains, and economic growth.
IEA energy expert Laura Cozzi emphasized that there is no strict linear relationship between growing data center loads and electricity prices. "If policies can integrate data centers into electricity networks in an intelligent, efficient, and flexible manner, there is no reason to believe they will necessarily drive up electricity prices."
Whether electricity prices will rise and who ultimately bears the costs are only one aspect of AI's electricity challenge. More importantly, countries must be able to deliver stable, affordable electricity to ever-expanding computing facilities in a timely manner.
Moreover, competition in AI has never depended on a single factor. While electricity is the entry ticket determining whether projects can proceed, advanced chips, computing architectures, and model efficiency still determine how fast and how far AI can advance.
Different countries also face different resource conditions and development bottlenecks. Some lack electricity; others are constrained by advanced chips; still others are held back by shortages of talent, capital, or infrastructure. In the future, the countries and companies that will be most competitive are those that can coordinate chips, computing power, algorithms, and energy, while transforming technology into practical applications at lower cost and higher efficiency.
Compiled by Tan Jie Ming
Posts
