US$20 Billion Into Infrastructure, US$8 Billion Into Deals: Who Will Fill Southeast Asia's AI Capacity?

Southeast Asia committed US$20 billion to digital infrastructure in the last 12 months, led by the US$11 billion acquisition of ST Telemedia Global Data Centres. Digital economy deals in the same period totalled US$8 billion, in line with the three year run rate. The gap between those totals is the question hanging over the region's AI buildout.

The e-Conomy SEA 2026 report, by Google, Temasek and Bain and Company, puts live data centre capacity at about 4 gigawatts and says meeting projected demand requires roughly a threefold expansion over five years. Capacity additions already exceeded 52 per cent a year, with low vacancy in third party operated facilities. Announced plans underline the scale of intent: Microsoft plans US$5.5 billion in Singapore from 2025 to 2029, and AWS has announced planned investment totalling more than US$33 billion by 2039. Those plans stretch over several years, so they should not be read as money spent in the last 12 months.

Private capital, by contrast, is concentrating fast. AI focused companies captured 80 per cent of private deployment in the last six months, up from 30 per cent in the second half of 2024. That concentration suggests investors believe applications will arrive. It does not yet prove who will pay for the capacity at a price that covers power, cooling and financing costs.

Demand, measured by use instead of by users, is still thin. Generative AI reached 243 million users in the six largest Southeast Asian economies, equal to 58 per cent of the working population. Use per person stands at about 10 per cent of United States levels. Only 1 to 2 per cent of the population pays for generative AI. Consumers account for about 60 per cent of regional AI token use. Enterprise workloads form the bulk of use only in Singapore and Malaysia.

The test for the infrastructure wave is therefore specific. Will enterprise buyers in Indonesia, the Philippines and Vietnam move from consumer led use, at about 80 per cent of token use, to paid workflows that run at scale? Will inference costs fall fast enough for free consumer traffic to become sustainable, instead of staying subsidised? Will vacancy stay low as capacity triples, or will the region build ahead of the paying market?

Data centres and power are long lived assets. AI products change in months. Southeast Asia is funding the long lived side first. The next reporting cycle should show whether the short lived side, paid applications and enterprise contracts, can fill it.