The training of AI models, he notes, consumes an “enormous amount” of electricity, with a ChatGPT query using nearly 10 times the electricity compared with a Google search.
As high-density server racks of up to 120 kilowatts (KW) per rack become standard, advanced liquid cooling solutions become a requirement, resulting in AI data centres being built with added infrastructure to supply large quantities of electricity and water.
Koh adds: “The huge increase in demand and inherent economies of scale has led to a preference to build hyperscale data centres.”
Hyperscale data centres have large power capacities of 20 to 50 megawatts (MW) and above. According to Synergy Research, the number of hyperscale data centres increased by 144 to 1,136 in 2024 with the three global cloud providers, namely Amazon Web Services (AWS), Google Cloud and Microsoft Azure, accounting for 59% of all hyperscale data centre capacity.
See also: OCBC Investment Research initiates ‘buy’ on Info-tech with target price of $1
Globally, there are more than 500 hyperscale data centres in the planning and construction phases. Synergy Research expects hyperscale capacity to double every four years with 130 to 140 hyperscale data centres added per year.
These facilities are optimised for high-density computing, enabling companies to support thousands of servers and petabytes of data with minimal latency. Hyperscale data centres also benefit from economies of scale in power usage and cooling, making them more sustainable and financially viable for large-scale cloud providers.
In the US, Meta chief executive officer (CEO) Mark Zuckerberg has recruited a team of experts to form Superintelligence Labs to unify Meta Platforms’ AI efforts. Zuckerberg plans to build a few clusters of gigawatt (GW) data centres in New Albany, Ohio and Richland Parish, Louisiana.
See also: DBS and Maybank stick with ‘buy’ on Singtel, backed by resilient business units
A huge scale is required to achieve superintelligence — AI that surpasses human intelligence across various domains. Stargate, a joint venture (JV) established by OpenAI, Softbank, Oracle and UAE-backed investor MGX, has committed to invest US$100 billion ($128.4 billion) to build data centres in the US.
“The first data centre with a power capacity of 2GW would be built on its flagship site in Abilene, Texas. These gigawatt data centres need dedicated power plants,” notes Koh.
Meanwhile, although many data centres were designed and built during the pre-AI era, the analyst notes that they can be retrofitted with high-density racks and liquid cooling systems to extend their useful life.
He adds: “Unfortunately, tenants have signed long-term leases, often with options for extensions, which limits the landlords’ ability to comprehensively upgrade their data centres. It is also difficult and time-consuming to obtain larger power capacity from the utility providers.”
Overall, with the emergence of generative AI, Koh sees an accelerated growth in the data centre market.
The data centre capacity required to support growth in computational workload grows as AI models scale up in size and capability. According to consulting firm McKinsey, global demand for data centre capacity is projected to expand by a compound annual growth rate (CAGR) of 22% from 2023 to 2030 to reach an annual demand of 219GW.
Koh writes: “The demand for AI-ready data centres, which provide high computational power and power density required for AI workloads, is expected to increase by a CAGR of 33%. About 70% of demand is for AI-ready data centres by 2030, which could be in a potential supply deficit.”
For more stories about where money flows, click here for Capital Section
He adds that besides the shift towards hyperscale data centres, the lease versus build mix balance is also changing from the more usual 50-50 ratio to 70-30, in favour of leasing.
Already, three global hyperscalers, namely AWS, Google Cloud and Meta Platforms, are switching from owning to leasing.
“Leasing is asset-light, reduces upfront capex and allows hyperscalers to scale up quickly without waiting for new builds. In markets with low vacancy, hyperscalers have to pre-lease capacity years in advance to secure the data centre. On the other hand, enterprise on-premises data centres are expected to fall to 30% of total capacity,” explains Koh.
He continues that data centres in Singapore are “more suited” to mission-critical and low latency applications, such as financial services and AI inference applications.
Singapore is power-constrained and has an “extremely tight” vacancy rate of 2%, notes Koh, adding that the city-state is a connectivity hub linking Southeast Asia to the global network through its 26 international subsea cables and three cable landing sites.
With this, the country is targeting investments of at least $10 billion to double its capacity for international subsea cables and landing sites over the next decade to support the pervasive usage of new AI applications.
“The domestic infrastructure will be upgraded to provide broadband speed of 10 gigabytes per second over the next five years,” writes Koh. He adds: “The government is working with the private sector and research institutions to scale the usage of autonomous systems using new technology, such as low-earth orbit satellites.”
Back in the US, Koh notes that the average vacancy rate for primary markets fell to an all-time low of 1.6%.
Hyperscalers and AI companies are both “racing” to secure power and lock in capacity years ahead of delivery to support future growth, with demand outpacing supply in all major markets.
The analyst notes: “About 74% of new capacity under construction in primary markets is pre-leased, mainly to cloud providers and AI companies. The average monthly asking rent increased 15% y-o-y to US$195-235 per KW for Northern Virginia in the 1H2025.”
Northern Virginia remains the largest market, with capacity increasing 33% to 3,480 MW, while capacity under construction increased 80%. Atlanta, Georgia, meanwhile more than tripled its capacity to 1,279MW, driven by hyperscalers and AI start-ups, with the average monthly asking rent rising 13% y-o-y to US$160-180 per KW in the 1H2025.
In general, Koh notes that data centre REITs have long weighted average lease expiry (WALE), which provides stable cash flows.
Keppel Data Centre REIT (KDCREIT) has a long portfolio WALE weighted by a lettable area of 6.9 years, while Mapletree Industrial Trust's (MINT) WALE for its North America portfolio is 6.1 years. Digital Core REIT’s (DCREIT) portfolio WALE is marginally shorter at 4.5 years.
Of the three REITs, Koh notes that KDCREIT and DCREIT have the largest exposures to hyperscale tenants, at 67% and 75% of rental income respectively. He adds: “Similarly, hyperscale data centres accounted for a bigger share of capacity at 26% for KDCREIT and 42% for DCREIT.”
Koh’s take
KDCREIT
With its pivot towards hyperscale data centres with a power capacity of 20 to 50MW and its “watershed” 2024 year of acquiring three hyperscale data centres, Koh likes KDCREIT, with his “buy” call and target price of $2.69 for the stock.
The REIT has also recently divested three sub-scale data centres — Intellicentre Campus in Sydney, Australia, Kelsterbach Data Centre in Frankfurt, Germany and Basis Bay Data Centre in Cyberjaya, Malaysia.
Presently, five of KDCREIT’s top 10 customers are hyperscalers. In aggregate, the five hyperscale tenants accounted for 66% of total rental income as of June. By industry sector, internet enterprises accounted for a similar 67.4% of total rental income for KDCREIT.
Additionally, the REIT also continues to sustain positive rental reversion going into 2026.
Koh writes: “KDCREIT achieved a positive rental reversion of 51% in 1H2025, driven by the sizable amount of colocation leases renewed in 2Q2025, largely from SGP4. Management guided positive rental reversion at high single- to low double-digit levels in 2026.”
He adds that with the Trump administration’s lifting of restrictions on the export of Nvidia’s and Advanced Micro Devices’ AI chips to China, this could spell potential tenants for KDCREIT’s Guangdong data centres, as they would require more data centre capacity upon procuring enough graphics processing units to train their AI models.
DCREIT
On DCREIT, Koh also has a “buy” call, with a target price of 88 US cents.
Six of DCREIT’s top 10 customers, he notes, are hyperscalers, comprising five cloud service providers and one social media platform. In aggregate, the six hyperscale tenants accounted for 75.2% of total annualised rent as of June. By customer profile, hyperscale cloud service providers accounted for a similar 64% of total annualised rent.
The REIT has also acquired data centres in Frankfurt and Osaka, both of which are hyperscale data centres.
“DCREIT owns stakes in three hyperscale data centres, namely Wilhelm-Fay-Straße 15 and 24 in Frankfurt, Digital Osaka 2 and Digital Osaka 3. In aggregate, the three hyperscale data centres accounted for 41.8% of total annualised rent and 34.2% of portfolio valuation,” writes Koh.
MINT
Lastly, Koh has downgraded his call on MINT to “hold” from “buy” previously, at a target price of $2.30.
The REIT presently owns some 62 data centres located in Singapore, Japan, Canada and the US.
Although it did not disclose the power capacity of its data centres, Koh notes that based on the latest FY2025 annual report, fitted hyperscale data centres accounted for 19.2% of MINT’s rental income from its data centre portfolio as of March.
He adds: “MINT’s US data centres are also geographically dispersed and fragmented across various states. Exposures to data centre hubs Northern Virginia and Northern California are quite limited at 20.0% and 4.9% respectively of total portfolio valuation for data centres as of March.”
He notes that data centres serving enterprise customers, often smaller in size, are “susceptible” to non-renewals, as enterprise customers often prefer to outsource and migrate to the public cloud hosted by cloud service providers, such as AWS, Microsoft Azure and Google Cloud.
As of June, enterprise customers accounted for 26% of MINT’s rental income from its data centres. Comparatively, cloud and hyperscale tenants accounted for a smaller 22% of total rental income.
In general, sector catalysts noted by the UOBKH analyst include rising demand generated by increased usage of generative AI and reasoning AI models and supply constraints in Singapore and the US.
Conversely, the main risk would be regulations on ethical and safe usage of AI, which could inhibit growth in the usage of AI.
As at 4.23pm, units in KDCREIT, DCREIT and MINT are trading at $2.37, 51 US cents and $2.07 respectively.
