Smarter AI design, paired with green data infrastructure, can help Southeast Asia meet its digital ambitions without compromising energy transition goals. The region doesn’t just need more green data infrastructure – it also needs AI applications that do more with less.
Southeast Asia’s digital economy is booming. With rapid growth in e-commerce, fintech, and AI services, the region is seeing a surge in electricity demand – particularly from data centres. These facilities run 24/7 and require intensive cooling, placing large, round-the-clock loads on national power grids.
Globally, data centres consumed about 415 TWh of electricity in 2024 – more than Indonesia’s entire national consumption. By 2030, their electricity use is expected to surpass that of Japan today. While much of the global data centre expansion is occurring in the United States, China, and Europe, Southeast Asia – anchored by the Singapore-Malaysia data hub – is quickly catching up, with regional demand projected to more than double by 2030.
Country-level estimates highlight the scale of the challenge. In Malaysia, electricity demand from data centres could rise sevenfold by 2030, reaching roughly 30 per cent of national consumption. In Indonesia, demand is expected to nearly quadruple, while in the Philippines, it may soar more than eighteenfold. Surging demand from data centre also risks competing with residential and community needs for electricity and water – especially in areas with constrained grids and limited water supply – raising broader social and equity concerns.
If this growing demand is met predominantly by fossil-heavy grids, it risks slowing – or even derailing – the region’s clean energy transition. As of 2022, fossil fuels, led by coal, still supplied over 70 per cent of Southeast Asia’s electricity, despite the ongoing expansion of renewables.
A key part of the solution lies in improving the hardware, particularly through the development of “Green Data Centres“. These facilities adopt advanced technologies such as high-efficiency cooling systems, waste heat recycling, workload shifting to off-peak hours, and integration of renewables. With these improvements, data centres can become far more energy-efficient and, critically, serve as levers to accelerate clean energy deployment.
Southeast Asian countries are already moving in this direction. Singapore’s 2024 Green Data Centre Roadmap sets top-tier energy efficiency standards and offers incentives for renewable energy use. Malaysia is preparing to launch a sustainable data centre framework by late 2025.
These initiatives mark important progress in improving the hardware layer of digital infrastructure. Yet concerns remain as to whether hardware improvements alone will be sufficient. Meanwhile, other powerful levers remain underused, particularly in the software layer. One opportunity lies in smarter, leaner AI design – building applications that deliver the same results with less computational heavy lifting, reducing demand for both infrastructure and energy.
In practice, this can be achieved by deploying smaller, task-specific AI models instead of sprawling general-purpose ones, using smaller but higher-quality datasets in model training, applying model-compression techniques such as pruning and quantisation to reduce computational load, and adopting more efficient algorithms for both training and inference.
These measures have considerable potential to improve software efficiency and cut energy use. For instance, Google has reported that its Gemini model, which combines more efficient software architectures and algorithms with hardware improvements, consumes significantly less energy than many earlier public estimates suggested.
To fully capture this potential, sustained research support is needed, including greater emphasis on green AI, alongside broader AI initiatives and sector-specific applications. But the more pressing challenge is not technical know-how; it is creating the right enabling environment.
For years, AI developers – from foundation model engineers to application builders – have been rewarded for accuracy, speed, and features, not energy efficiency. That is starting to change as rising computation and token costs force efficiency into the conversation, but most efforts remain ad hoc. Without a clear policy signal to embed efficiency into AI application development, progress could stall, and energy-intensive software could prevail if energy costs dip or priorities shift.
This is where governments and companies can work together. Instead of regulating AI design directly, policymakers can foster an enabling environment by promoting reporting standards for the energy use of AI applications and supporting voluntary efficiency benchmarks. Companies, in turn, can collaborate by sharing data, piloting lightweight applications, and showcasing best practices in algorithmic optimisation. Public agencies should also consider prioritising essential social needs over discretionary use, ensuring grids continue to serve society’s broader interests as AI demand grows.
The path is not straightforward. Many AI applications, such as news summarisation tools, rely on upstream third-party systems whose energy costs are often unknown to application developers. This opacity makes it difficult to assign responsibility, measure impact, and reward efficiency.
To navigate this challenge, the first vital step is recognising that software efficiency is just as critical as hardware upgrades. Such recognition is the necessary starting point for deeper discussion between relevant stakeholders on how best to align digital growth with environmental sustainability.
Dr Muyi Yang is senior policy analyst at Ember. He also holds positions as a non-resident senior policy fellow at Asia Society Australia and adjunct fellow at the Australia-China Relations Institute at the University of Technology Sydney.
Xiwei (Sherry) Xu is a principal research scientist at CSIRO’s Data61.
David Lo is the OUB chair professor of computer science at the School of Computing and Information Systems, Singapore Management University.
This article was first published in Fulcrum, ISEAS – Yusof Ishak Institute’s blogsite.
