top of page
Image by Akif Waseem

APAC TMT Notes

A collection of articles sharing operator-side perspective on telecommunications, digital infrastructure, governance, and technology across APAC covering key topics in investments, market dynamics, emerging technologies, and organisational development.

The Real AI Infrastructure Race: Lighting Up Capacity

Writer: Kaye Hau
Kaye Hau
Nov 2
5 min read

Updated: Nov 4


data centre farm
Credit: Canva

The race to build AI infrastructure is afoot.


Hyperscalers have committed billions to data centres, chips, networks and energy infrastructure. Developers are securing land and power, while equipment makers are expanding production. And investors are creating new financing structures to fund the build-outs.


J.P. Morgan estimates that hyperscaler capex could reach US$697B in 2026, making AI infrastructure one of the largest capital deployment themes in the market.

The demand for AI compute is real. The more difficult question is whether the market is making multiple investment decisions against the same future demand. We're nowhere near to concluding that the market is heading into oversupply. The more immediate underwriting question is whether planned, reserved, contracted, powered, commissioned and utilised capacity are being treated as though they are the same thing.


The real race then, is to secure critical inputs and resources, complete construction, install equipment, and turn the capacity on quickly enough to generate revenue.

From Planned Capacity to Live Compute

AI infrastructure requires significant upfront investment. Land must be secured, power connections arranged, buildings constructed, cooling systems installed, networks deployed and accelerators purchased; much of this investments take place before customer workload is fully operational.


Therefore, customer offtake agreements matter. A developer with long-term customer commitment can use the contracts to support construction financing and provide comfort to equity investors. The customer, in turn, secures future capacity in a market where suitable land, power, and data centre space are becoming increasingly scarce.


The Bank for International Settlements (BIS) has described the growing use of long-term leases and capacity offtake agreements in AI infrastructure financing, including structures where hyperscalers work with private capital and special-purpose vehicles.


While it's all rational behaviour, a project only becomes an operating asset when it is capable of serving and hosting customers.


Different Definitions of Capacity

A data centre project can have different categories of capacity and they're not interchangeable. They span from announced, planned, contracted, reserved, powered, installed, available, and utilised.


A customer may commit to 100MW initially but only uses 40MW. The remaining capacity may be reserved for future deployments, subject to ramp-up schedules, options for expansion, or minimum payment obligations.


For the developer, that commitment can be used to support financing. For the market, however, it can create the impression that more demand has been secured than is currently operational.

The distinction matters as multiple projects may be built against similar assumptions. Each decision may be commercially rational in isolation but collectively, could potentially produce more capacity than the market ultimately absorbs; this is a risk to monitor and not yet a proven market outcome.


Circularity of AI Financing

A further complication is the growing circularity of capital within the AI infrastructure ecosystem.


Chipmakers and hyperscalers are increasingly involved not only as suppliers and customers, but also as investors and financing partners. The BIS has described arrangements in which chipmakers and hyperscalers take equity stakes in AI labs or neocloud providers, which then commit to purchasing chips or compute capacity.


This is not necessarily improper. Strategic investments and long-term purchase commitments can help secure supply and accelerate deployment; but they can make the demand picture more difficult to assess.


Capital flows into an AI company, which then uses the capital, directly or indirectly, on chips purchase, cloud services, or compute. These commitments goes into revenue forecasts, and financing of other companies in the same ecosystem.


The underlying demand may be real, but it may not represent fully independent end-user demand.

Public filings illustrate how closely these relationships are becoming inter-connected. NVIDIA disclosed cloud services commitments, data centre leases, and arrangements to support customers and partners in securing land, power, and data centre capacity. It has also disclosed arrangements under which it might purchase capacity if AI cloud partners are unable to sell that capacity to third-party customers.


The concern is whether circular financing amplifies the apparent scale of demand and makes it harder to identify who ultimately bears the downside if workloads, prices or utilisation fall short.


Capacity can be overbuilt without anyone deliberately double-booking the same physical infrastructure. The market could be building against a series of commitments that are individually valid but collectively dependent on the same underlying assumptions.

Cloud vs Capacity Overbooking

Cloud providers have long used pooled infrastructure and statistical multiplexing. Many customers share a platform because they don't use all their maximum allocation at the same time.


That model is not the same as dedicated data centre capacity. A powered shell, dedicated building or AI cluster has physical limits. The same megawatt, rack or GPU cluster can't generally be sold as firm, simultaneous take-or-pay capacity to multiple customers unless the agreement explicitly allows sharing.


The risk therefore, is how the market defines a commitment.


Whether it's a firm take-or-pay obligation, or an option for future capacity; or it's subject to financing, power delivery, ramp-up schedule or acceptance.


An offtake agreement can reduce demand risk significantly but it doesn't eliminate the risks from construction, power, technology and customer concentration.

Current Demand Remains Strong

The available public data still shows strong current demand and constrained supply for capacity. In North America, JLL reported 25GW of data centre absorption in the first half of 2026, with vacancy at approximately 1%. The same report described a debt wav e of roughly US$700B supporting the build-out.


But current absorption doesn't eliminate future oversupply risk. Projects take time to build, power and commission. The commitments made today might only become operational years from now. By then, customer demand, chip efficiency, model architectures, power availability, and financing conditions could have changed.


The more relevant question is whether today's commitments will still translate into sufficient utilisation and cash flow when the capacity comes online.

Discerning Public Markets

Recent capital market activities provide a snapshot of the sentiments surrounding data centre and AI infrastructure.


Firmus Technologies withdrew its planned ASX listing after muted investor demand, while AirTrunk postponed its planned Singapore REIT listing, amid market volatility, higher interest rates, and renewed scrutiny of AI-linked valuations.


DayOne's preliminary Nasdaq filing shows that the public markets remain an important source of capital for data centre operators, but it doesn't provide a clear read on investor appetite. The offering size, price range and timing remain to be advised.


Taken together, these developments suggest that while there is still support for the long-term AI infrastructure theme, investors are becoming more discerning and selective about valuations, interest-rate sensitivity, customer commitments, and the timing of future cash flows. The market is asking for more evidence before underwriting the next phase of growth.
Unit Economic Value of Power

The conversations are also shifting from the absolute amount of capacity built to the economic value generated from each unit of power.


Power, land, water, grid access and specialised equipment are all becoming constraints. A project that has secured 100MW is not necessarily more valuable than one with 50MW. It depends on the quality of the customer, efficiency of the facility, revenue generated, utilisation profile, and the durability of demand.


And this is where regulation becomes part of the investment equation. Singapore's Digital Infrastructure Bill introduces licensing regimes covering the sustainability of data centre operations, as well as the security and resilience of major data centres and cloud providers. It also creates a basis for strategic, economic and green commitments made in exchange for new data centre capacity to be enforced through licence conditions. Likewise, regulators and policymakers across APAC are beginning to push the industry beyond capacity expansion, asking whether each unit of power, water and land allocated to data centres generates sufficient economic value.


The broader policy push signals that access to land and power is increasingly being treated as a scarce public resource that must generate greater efficiency and strategic value.
The Race to Realised Revenues

The AI infrastructure race is real and public evidence still points to strong demand and tight current supply.


However, the market should be careful not to count every announcement, reservation, option and future phase as though it's realised revenue.


The next winners might not be the ones with the largest capacity pipeline. They would be ones that can secure power, turn on the infrastructure, and convert it into cash-generating compute.

Getting Started

Reach out for a confidential initial conversation on how we can help.

Every engagement begins with a complimentary discovery session to understand your objectives, challenges and priorities. If we're not the right fit, we'll tell you honestly and, where possible, point you in the right direction.

bottom of page