�鶹

Last weekend, social media was seemingly spirited away into the world of Studio Ghibli.

From everyday folk to governments, online users flooded the internet with images in the distinct hand-drawn style of films produced by the award-winning Japanese animation studio, including My Neighbour Totoro and Kiki’s Delivery Services.

However, rather than being hand-drawn, the images inspired by the animation studio were generated using ChatGPT, a generative artificial intelligence (AI) tool created by OpenAI.

In a post on X (formerly Twitter) on Mar 30, OpenAI’s chief executive officer Sam Altman joked: “Can y’all please chill on generating images. This is insane, our team needs sleep.”

Days later, he said on the same channel that a million users had signed up for ChatGPT in an hour – a feat that eclipsed ChatGPT’s record of drawing a million users in five days following its public launch two years ago.

Data from market research firm Similarweb showed that the viral trend drove a spike in ChatGPT use and that the average weekly active users breached the 150 million mark for the first time this year. 

Studio Ghibli itself has not commented on the craze, but the trend has sparked fresh debate about .

Some critics have argued that it counts as an infringement when AI tools trained on copyrighted original works are used to generate images closely resembling the unique style of a particular artist.

In the midst of the debate, past comments made by Studio Ghibli's co-founder, animator and filmmaker Hayao Miyazaki that criticised AI have resurfaced and made their rounds. 

For example, in a 2016 documentary called Never-Ending Man: Hayao Miyazaki, a group of developers showed him an AI-generated animation demo of a zombie that could be used for a video game. 

After watching it, the filmmaker said: “Whoever creates this stuff has no idea what pain is whatsoever. I am utterly disgusted … I strongly feel that this is an insult to life itself.”

While the debate rages, there is another troubling effect of generative AI that has been relatively overlooked: Its environmental toll.

The International Energy Agency said that a single ChatGPT query submitted by a user requires 10 times the amount of electricity as a Google search – 2.9 watt-hours (Wh) compared with 0.3 Wh.

A study by researchers at Carnegie Mellon University in the United States and the machine learning platform Hugging Face found in 2024 that generating 1,000 images using several generative AI tools requires about 2.907 kilowatt-hours (kWh) on average. 

This means that producing a single photo with generative AI can consume as much energy as fully charging an average smartphone, though this may vary depending on the capabilities and computing power required by the generative AI tool.

The energy required to generate text is typically less.

A study by the University of California, Riverside in the US found that a 100-word email generated by an AI chatbot using ChatGPT’s GPT-4 model requires 0.14 kWh. That is still enough electricity to power an LED light bulb for 14 hours.

The same email generated with GPT-4 uses 519 millilitres of water – slightly more than a bottle of water.

Mr Sharad Somani, partner at professional services firm KPMG in Singapore, said that generative AI systems require a large amount of computing power to train and operate. 

“This power demand translates to significant electricity use and unless it is drawn from renewable sources, it contributes to high carbon emissions,” he added.

Mr Somani is also head of environmental, social and governance consulting at KPMG.

Guzzling power aside, experts noted that generative AI’s operation requires a large amount of water and rare earth materials, the extraction of which carries significant environmental costs.

WHY GENERATIVE AI IS BAD FOR THE ENVIRONMENT

Traditional search engines and other traditional AI technologies require computing power as well, but the immense amounts needed to train and operate generative AI models have an environmental impact far outweighing that of the former, experts told �鶹 TODAY.

So why does generative AI need all this energy?

Mr Laurence Liew, director of AI innovation at AI Singapore – a national programme focused on enhancing the country’s capabilities in this area – said that every component in the AI supply chain has its own environmental cost.

“Training a single large language model can consume more electricity than hundreds of households use annually," he added.

"The training phase alone for models like GPT-4 or Claude requires thousands of specialised GPUs (graphics processing units) running at full capacity for weeks or months.”

Enter data centres: Mega facilities providing the computing infrastructure that IT systems require such as servers, data storage drives and equipment that includes GPUs.

However, in providing the computing power, these systems generate a significant amount of heat. As a result, data centres consume large amounts of energy and water to power extensive cooling systems, Mr Liew explained.

In places with warm tropical climates such as Singapore, data centres naturally need more energy and water.

Mr Jon Curry, vice-president of operations for the Asia Pacific at Digital Realty, which operates three data centres in Singapore, said: "Cooling systems in data centres typically account for a large percentage of total energy consumption, as many operators maintain their equipment at temperatures of 22°C and below."

As of May 2024, the more than 70 data centres in Singapore provided 1.4 gigawatts of computing capacity. They contributed 82 per cent of greenhouse gas emissions produced by Singapore’s information and communication technology (ICT) industry last year, though the exact amount of emissions is not publicly available.

Based on the most recent data from the Ministry of Trade and Industry, data centres were responsible for around 7 per cent of Singapore’s total electricity consumption in 2020.

The hardware needed to support generative AI also has an environmental impact.

Mr Vivek Kumar, chief executive officer of the World Wide Fund for Nature, Singapore (WWF-Singapore), said that these include the lithium-ion batteries used in GPUs.

“The extraction, transportation and supply chains for these materials contribute to carbon emissions, while the reliance on rare earth minerals sourced through mining operations leads to ecosystem destruction, resource depletion and pollution.”

To add to the hefty price tag, building data centres also requires land and other resources, which might result in the loss of natural habitat when land is cleared for the infrastructure.

And as generative AI capabilities and outcomes become more accurate, the computing power required often increases. This is due to the rise in parameters, or internal variables that enhance a generative AI’s accuracy.

Dr Kirti Jain from technology consultancy Capgemini, said: “To do anything with GPT-3 requires a substantial amount of energy and resources, and that has about 1.75 billion parameters.

"The newer GPT-4, while more accurate, has parameters estimated to be in the trillions.

“So you need more energy to power the data centres providing the computing power and more water to cool the data systems.”

He is the consultancy's vice-president and head of insights and data global business line for the Asia Pacific and Middle East.

WHY THIS IS TROUBLING

With generative AI driving energy usage in data centres, the International Energy Agency projected that electricity consumption from these facilities could double by 2026.

This could result in a jump from an estimated 460 terawatt-hours in 2022 to more than 1,000 terawatt-hours in 2026, or roughly the electricity consumption of all of Japan in one year.

Despite this dark side, it is impossible to stop the growth and adoption of these advanced models as they become increasingly integrated into corporations both big and small.

The efficiencies that this AI tool can bring to industries from manufacturing to healthcare present significant opportunities, experts said.

Adjunct Professor Ngiam Kee Yuan, head of the Artificial Intelligence Office at the National University Health System (NUHS), said that its secure clinical large language model tool has helped enhance healthcare services.

"A study of NUHS clinicians showed that users of the platform's summarisation tool achieved 40 per cent greater efficiency," he added.

"Now adopted by more than 3,600 NUHS staff members as the default generative AI tool, RUSSELL-GPT continues to grow in both capability and adoption."

Analysts at real estate firm Knight Frank Singapore have also seen their output increase by roughly 1.5 times with the use of generative AI tools, its head of strategic consulting and workplace Samarth Kasturia said.

For example, it takes menial work such as data preparation out of the picture, allowing employees to focus on research and conduct data analytics more quickly.

And in the longer term, he believes that generative AI may be good for the environment.

"For example, it could be used for facility management and reduce energy wastage in an office, like maintaining the air-conditioning system," he said.

In the meantime, though, Mr Kumar of WWF-Singapore said that AI development must prioritise sustainability.

“Investing in energy-efficient models, renewable power sources and responsible resource management is essential in ensuring that technological progress benefits both people and the planet without deepening environmental harm."

Mr Matthew Oostveen, chief technology officer and vice-president for the Asia Pacific and Japan at data storage company Pure Storage, said that with Singapore setting its sights on becoming an AI hub, it is important to address these resource concerns.

“Data centres in countries such as Singapore draw their energy supply from power stations, which can potentially strain energy resources with demand from other sectors such as electric vehicles also growing," he added.

Mr Oostveen also said that as the significant amount of water required for cooling data centres puts added pressure on Singapore’s already strained water resources, it is important for the country to balance technological growth with sustainable resource management.

In response to queries, the Ministry of Digital Development and Information (MDDI) said that although Singapore is a small country, it is committed to sustainability so that it may grow the tech ecosystem while meeting its climate change commitments.

For example, the ministry has launched a Green Data Centre Roadmap that sets out how Singapore should grow its data centre sector sustainably. 

With plans to increase data centre capacity by at least 300 megawatts, it will prioritise "efficient and green energy deployments", MDDI added.

Beyond that, the Building and Construction Authority and Infocomm Media Development Authority (IMDA) refreshed the Green Mark for Data Centres 2024 to raise sustainability standards for these facilities, among other things.

"The government is also exploring ways to uplift data centre sustainability through regulations. We are studying other jurisdictions  and are  in early engagement with the industry  to develop a framework for Singapore’s context," the ministry said.

It also said that IMDA, a statutory board under the ministry, rolled out two programmes to lower carbon emissions of software last year.

One of these is the Green Software Trials to assess and quantify the impact of carbon reduction techniques. These techniques include resource redistribution, application modernisation, AI optimisation and computational offload.

"The trials will also generate valuable data and insights for IMDA to create green software guidelines for the industry," MDDI said.

Globally, efforts are underway to address the perils and promises posed by power-hungry generative AI.

The European Union, for instance, introduced the first-ever legal framework on AI last year, which includes provisions for energy consumption reporting, as well as principles for developing and using AI systems in a “sustainable and environmentally friendly manner”.

REDUCING AI'S CLIMATE IMPACT

Several businesses and organisations in Singapore told �鶹 TODAY that they have taken proactive measures to mitigate the environmental impact of generative AI, which they use in their operations.

Singapore’s biggest bank DBS said that it has a four-lever approach to address environmental challenges such as reducing consumption of resources and generating renewable energy.

It also buys green products, energy and renewable energy certificates, as well as carbon offsets. Such certificates act as a mechanism for accounting, tracking and assigning ownership of renewable energy.

Associate Professor Daniel Ting, director of the AI Office at public healthcare cluster SingHealth, said that the organisation is "cognisant of the amount of compute resources used" in training its AI models.

SingHealth has developed several generative AI projects such as Note Buddy, which transcribes and translates doctor-patient consultations into medical notes and is used by 1,700 healthcare professionals across the cluster.

"We endeavour to use small models that require less compute resources where possible, and train AI models for highly specific tasks. This ensures that computing resources are used efficiently and keeps energy usage at an optimal level," Assoc Prof Ting added.

At NUHS, it seeks to use renewable energy sources for computing whenever possible, Adj Prof Ngiam said of the hospital group's generative AI usage.

As for small- and medium-sized enterprises (SMEs), the environmental impact of their usage of generative AI, while not well-studied, is likely minimal, experts said. 

“SMEs typically lack the capital for investing in efficient AI infrastructure and have limited in-house expertise on sustainable deployment,” Mr Liew of AI Singapore said.

"Many struggle just to begin their AI journey, making environmental considerations seem like a luxury rather than a necessity."

Several tech companies behind generative AI programmes also said that they are actively working to minimise environmental costs.

Google told �鶹 TODAY that it has deployed strategies to manage the environmental impact. These include building energy-efficient computing infrastructure and optimising its AI model training.

“Compared to five years ago, (Google’s data centres) now deliver around four times as much computing power per unit of electrical power,” it added. Its data centres are optimised such that they use less power than the industry average.

The company also stated that it has identified, tested and adopted practices that can reduce the energy required to train an AI model by up to 100 times and reduce associated emissions by up to 1,000 times, among other things.

“Across the Asia Pacific, the way Google tackles our carbon, water, waste and ecological impact is tailored to each country … We also know that scaling AI and using it to accelerate climate action is just as crucial as addressing the environmental impact associated with it.” 

The company noted that AI itself could be used for good, referring to a Boston Consulting Group study, which suggests that AI could mitigate 5 per cent to 10 per cent of global greenhouse gas emissions by 2030. 

Mr Arun Biswas, managing partner for strategic sales and sustainability at IBM Consulting APAC, told �鶹 TODAY that the company is “committed to developing sustainable AI”.

For example, the company's new is designed to lower energy consumption during both the training and inference phases of AI, contributing to reduced carbon emissions, he said.

Its data centres are located near renewable energy sources and use a hybrid cloud model, allowing them to reduce energy waste while transiting and computing data.

Mr Biswas added that IBM is collaborating with the National University of Singapore to establish a new AI research and innovation centre here, along with other initiatives.

"The goal is to significantly reduce AI computing power consumption – up to nine times at equal or better performance – while maintaining or improving performance," he said.

"Deployment of these solutions is planned by the second half of 2025.”

Data centre operators here are also trying to green themselves.

At Equinix, Ms Yee May Leong, its managing director for Singapore, said that its fifth data centre here uses a proprietary surface cooling technology that allows it to handle high-density workloads while reducing water and power consumption.

The company uses NEWater for its cooling systems, Singapore’s high-grade reclaimed water.

The company has also retrofitted cooling tower fans with more advanced fans powered by electric motors in its third data centre here. This has resulted in energy savings of 38 per cent to 50 per cent on cooling tower fan operations, Ms Yee said.

She added that the company is developing a sixth data centre in Singapore that will be specifically designed to manage AI workloads efficiently while prioritising energy efficiency and sustainability.

Mr Curry of Digital Realty said that the three data centres it operates run on 100 per cent renewable energy coverage as of last month, with some of its power coming from solar facilities installed on-site in 2023 and 2024.

It has also taken on other initiatives such as a collaboration with national water agency PUB to pilot a cooling tower that can reduce the amount of water discharged from its cooling systems monthly by 60 per cent, or about 650,000 litres.

Over at Alibaba Cloud Intelligence Singapore, its deputy country manager Hon Keat Choong said that the company has pledged to use 100 per cent clean energy by 2030.

In the fiscal year 2023-2024, the company's self-built data centres improved their power usage efficiency and 56 per cent of electricity consumption came from clean sources, he added.

He also said that the company has developed Energy Expert, a management tool that allows enterprises to measure and analyse their carbon emissions and energy consumption using AI. This has been used by 3,000 organisations globally, including in Singapore.

MAKING GENERATIVE AI GREENER

Even though there are people who believe that the future efficiencies of generative AI will outweigh the current environmental costs, Mr Kumar of WWF-Singapore warned that this assumption is problematic. 

“It shifts our focus away from making sustainable AI a priority. We become so focused on what AI can do and its promised efficiencies that we stop asking how to make the technology itself more sustainable.

“While AI can be beneficial, we still need to make sure that its infrastructure and full life cycle have as little environmental impact as possible,” he said.

However, there are several hurdles in making generative AI models environmentally sustainable, one of which is the design and use of energy-efficient hardware, Mr Somani of KPMG said.

For example, developing cutting-edge AI chips that are more sustainable would require substantial investments in research and development. This also means that it would be costly and harder for small firms to obtain and use newer and greener technology.

“Compounding this problem is the rapid pace of technological innovation, which drives frequent hardware upgrades. This not only increases costs but also contributes to the rise of e-waste, diluting any long-term environmental benefits,” he added.

Getting access to renewable energy can pose a challenge due to limited availability and high costs. This makes it difficult for businesses and data centres to move away from traditional energy sources such as fuel, Mr Somani said.

Sustainability can also get in the way of development.

“While smaller, energy-efficient AI models can reduce environmental impact, they sometimes trade off performance or computational power. For applications requiring high precision, this trade-off makes them less practical,” he explained. 

“Additionally, many companies prioritise rapid market entry and profitability over sustainability, slowing the adoption of energy-efficient practices.”

So, is there a way to unlock the full potential of generative AI while minimising its harmful impact on the environment? 

Mr Oostveen from data storage firm Pure Storage said that companies developing AI should look at improving the efficiency of their models and thereby reducing the computing power needed.

He pointed to DeepSeek, an AI model from China that is on par with advanced models from OpenAI and Meta in the United States, but developed at a fraction of their costs.

“DeepSeek’s approach is rooted in a mixture-of-experts model, where smaller, highly trained models work together in tandem. This sophisticated method selects the most appropriate expert model, optimising for both performance and efficiency.”

Achieving model efficiency ensures that supporting hardware can meet computing power needs and energy consumption, Mr Oostveen said.

“Apart from hardware, prioritising smarter data management can help reduce AI’s environmental impact.”

He pointed out that large data sets and unnecessary data increase the computing power required, which in turn raises electricity and water needs. 

Dr Kirti from tech consultancy firm Capgemini said that users should be mindful of their AI usage and its impact on the environment.

For example, instead of turning to generative AI tools for every question, they may want to consider using other tools. For example, using a map application for directions would be a less resource-intensive option.

Ms Sammie Leung, a partner specialising in sustainability and climate change at PwC Singapore, said that data centres should be strategic with their infrastructure and design.

One way is to have their cooling systems optimised to ensure that they are as efficient as possible.

However, greater efforts are needed to make renewable energy more accessible so that data centres can rely on less carbon-intensive options, Ms Leung added, echoing the views of several experts who mentioned that nuclear energy may be a reliable alternative.

Although there is a strong need to mitigate the environmental impact of generative AI “before it's too late”, it does not mean that generative AI should not be used or improved, she qualified.

“If we do not fix this issue with green energy sources for generative AI, the planet will not forgive us … but there are benefits and efficiencies (that generative AI) provides and we should not limit them either.” 

Mr Liew of AI Singapore said that incorporating energy usage metrics directly into developer tools can help make developers more aware of energy consumption.

“What gets measured gets managed.

"By exposing energy consumption information within AI development environments, tool creators can nudge developers towards more sustainable choices.”

Beyond technical optimisation, fundamentally rethinking how one approaches AI development is another way to reduce the environmental impact, he proposed.

For instance, developers could look into adapting existing AI models rather than training new ones from scratch, avoiding the resource-intensive task.

“The environmental challenges of AI are substantial, but so is our capacity for innovation," Mr Liew said.

"With deliberate focus and collaborative effort across industry, government and academia, we can ensure that AI's benefits don't come at the expense of our planet's future."