In January, TM Nxera announced an agreement securing 280 MW of electricity for its planned data center campus in Johor. By April, it had marked the building’s topping-out. For enterprise buyers, these milestones raise a practical question: when will that investment become capacity their applications can actually use?
Power agreements and construction progress are steps towards delivery; neither establishes that the full capacity is operational.
That distinction should shape AI procurement across Asia-Pacific. A successful pilot can justify a business case without establishing whether the production service can obtain the capacity it needs, in the required location, when demand rises. CIOs need to connect rollout promises to infrastructure commitments they can verify.
To understand how infrastructure constraints are becoming more critical as enterprises race to scale AI, iTNews Asia spoke with Seonji Lee, Consulting Associate, ICT & Security Advisory at Frost & Sullivan. “The biggest constraint is likely to be reliable power at the right location and on the right timeline,” she says.
Make availability part of the business case The International Energy Agency projects that global electricity use by data centres will rise from 485 TWh in 2025 to approximately 950 TWh in 2030. Its 2026 assessment also identifies tightening energy supply chains and stretched planning systems. These are global projections and constraints; the consequences for an APAC deployment depend on the specific site and service.
Lee draws an important distinction between national electricity supply and a project’s ability to connect to it. She sees constraints across APAC from grid connections and permitting through to transmission infrastructure. A country’s power surplus therefore cannot substitute for evidence that a particular facility can serve an enterprise’s deployment timetable. McKinsey’s Technology Trends Outlook 2026 similarly places energy and physical infrastructure at the centre of technology deployment. For enterprise buyers, however, investment momentum offers limited assurance about an individual delivery date.
Demand is changing too. The IEA finds that reasoning and agentic tasks can consume much more energy than simple text generation, even as efficiency improves. Capacity planning needs to reflect the work enterprises intend to automate, alongside user numbers.
Procurement should therefore distinguish planned capacity, contracted power, and commissioned capacity available to the customer. A colocation buyer needs evidence on energisation, supported rack density, cooling readiness, and expansion milestones. A cloud or managed-AI customer needs commitments on regional service availability, quotas, and the conditions under which capacity can expand.
Choose locations around the work
The infrastructure response is already taking different forms across APAC. Lee points to Korea as an example of coordinated capacity planning. “The government has designated AI data centres as a strategic industry and is coordinating electricity, land, permitting, and power and cooling infrastructure around large-scale development.”
For an enterprise, this scenario raises a due-diligence question: can a provider align power, facilities, and approvals with the promised deployment date?
Japan illustrates a second shift. Lee notes that, as of 2023, around 90 percent of its data center space is concentrated in Tokyo and Osaka. The country is now pursuing greater regional diversification through its “watt-bit collaboration” approach, coordinating data centre locations with electricity and telecommunications infrastructure and seeking regions with available land, industrial water, and grid capacity.
That geography creates architectural choices. An overnight forecasting job may accommodate a more distant facility, while an interactive customer assistant needs response times tested against its users and the systems supplying its data.
“The physical characteristics of a location will increasingly determine which AI workloads make economic sense to run there,” Lee says.
China takes the principle to the national scale. Its “East Data, West Computing” strategy is building computing capacity in western regions with greater renewable-energy availability while much of the demand remains in the east, Lee explains, “The location of compute does not always need to match the location of AI demand.”
The same distinction applies across APAC. Training and batch processing generally offer more flexibility in where and when they run. Inference and production workloads may be more sensitive to latency, connectivity, data sovereignty, and proximity to enterprise systems. Before choosing a region, teams should document those requirements alongside recovery needs and data-movement costs.
Lee argues for a portfolio of locations matched to workload characteristics. “Over time, such an approach could make workload portability and geographic flexibility as important to enterprise AI architecture as the underlying compute itself.” That flexibility only has value when the application, data, and controls can move safely.
Price the complete service
Moving compute to a cheaper location can introduce costs elsewhere. Data transfer, replicated storage, network connections, and the engineering needed to operate across environments all belong in the comparison. Buyers should ask how cooling requirements and energy charges affect the commercial offer, particularly as rack density increases.
TM Nxera’s inclusion of liquid cooling in its Johor campus plans illustrates why “AI-ready” needs a technical specification behind it. Enterprises should establish what density is supported at launch, what requires a later upgrade, and how are water availability and heat rejection addressed? A proposed design should not be treated as measured operating performance.
The business metric should be the cost of completing an acceptable task at the required service level. A cheaper model or region may lose its advantage if it generates more retries, demands more human review, or misses response-time targets. Testing under representative demand makes those trade-offs visible before a wider rollout.
Prove the alternative before depending on it
Geographic flexibility also incurs an operational cost. A second provider is useful only if the application, its data, and its controls can work there. Recovery exercises should establish whether permissions, model behaviour, and data synchronisation remain adequate when a workload moves.
Lee cautions that “additional compute capacity does not necessarily translate into additional business value." Her point also extends to ownership: someone must be accountable for the production service, including its cost, performance, and business outcome.
Data needs to be accessible and governed, security needs to be built into production workflows, and accountability for model performance and business outcomes needs to be clear.
- Seonji Lee, Consulting Associate, ICT & Security Advisory at Frost & Sullivan.
These responsibilities should travel with the workload: moving an application must not leave its monitoring, access controls, or business owner behind.
For a critical AI workflow, the investment decision by CIOs and technology leaders should therefore include evidence of usable capacity, a tested cost per completed task, and a workable response to disruption. Business, infrastructure, and security owners need to agree on what happens if any of those assumptions fails.
APAC’s changing infrastructure map gives enterprises more potential locations. To turn those options into dependable AI services, buyers must establish what they can actually deliver and ensure their rollout commitments match that evidence.
Nivash Jeevanandam is a contributing writer to iTNews Asia.
