AI-ERA DATA CENTERS IN SOUTHEAST ASIA – PART 1 OF 6
The Speculative Build Dilemma Under Pressure from AI Data Center Clients
James Soh
Somewhere in Southeast Asia right now, a data center developer is about to pour concrete on a new facility—and they are making a commitment that did not exist three years ago.
Build speculatively, the way the region has always built, and risk handing over a shell that cannot host the AI workloads, battery layouts, and safety regimes your most valuable prospective tenants will demand. Or slow down to co-design around specific AI clients, battery chemistries, and authority requirements—and risk missing the market window while a competitor brings good-enough capacity to market first.
That dilemma is now the central strategic question for data center developers, operators, and investment committees across the region. It touches power architecture, structural design, fire engineering, regulatory strategy, and—critically—the operations teams who will be asked to run whatever gets built.
This article is the first in a six-part series examining the AI build-out in Southeast Asia through three lenses: Design for Safety, Design for Operations, and technology relevance. What has to be decided up front—at planning and concept design—so that AI-era facilities can be built, permitted, and operated safely over time, instead of becoming expensive shells that never quite match the clients or staff they attract.
Later in this article, we map how the AI data center build-out is landing differently across Singapore, Malaysia, Thailand, Indonesia, Vietnam, and the Philippines—and what that means for developers and operators in each market.
To understand how we arrived at this dilemma, it helps to look at the playbook it is replacing.
———
The Old Speculative Playbook: “Any Tenant, Any Time”
Southeast Asian data center developers perfected a particular rhythm over the past fifteen years. Acquire the land. Lock in the power allocation. Build a generic white-space shell with room for standard rack densities, conventional chilled-water or DX cooling, and centralized VRLA battery rooms tucked away in dedicated spaces. Fit-out could be tailored later, once a tenant signed.
Battery chemistry and placement sat deep inside engineering and procurement documents. They rarely surfaced in board slides or investment committee memos. Nobody needed them to. The assumption was that a well-designed centralized battery room, sized for N+1 redundancy with industry-standard VRLA strings, would serve any reasonable tenant mix.
And for the most part, it did. When your tenant base is primarily rack-centric IT—financial services, government, SaaS platforms, content delivery—densities are moderate, cooling loads are predictable, and fire-safety expectations are familiar territory for both engineers and regulators. The speculative playbook was not careless. It was proportionate to the market it served.
The problem is not that the old model was wrong. The problem is that the market it was designed for is no longer the only market walking through the door.
———
AI Clients Change the Physics—and the Timeframe
When an AI-focused client approaches a data center operator in Southeast Asia today, the conversation sounds different from what most leasing teams have rehearsed. The discussion moves quickly past available white space and into specifics: what is the per-rack power envelope, what is the floor loading capacity, what cooling topology is in place or can be retrofitted, and—increasingly—what is the battery strategy.
AI clients increasingly standardise the top three to four layers of the stack, while operators control the facility and utility layers. The speculative build dilemma lives in the gap between these layers.
Current-generation GPU clusters (think NVIDIA B200/B300-class systems and the upcoming Rubin-class platforms) push rack densities well beyond what most speculative builds in the region were designed for. A single AI training rack may draw 40–70 kW or more, compared with the 6–12 kW per rack that traditional enterprise tenants require. That difference cascades through every layer of the infrastructure stack: power distribution, UPS sizing, cooling capacity, structural floor loading, and the fire-safety regime that wraps around all of it.
Battery strategy becomes a live design question under these conditions. A centralized VRLA or LFP battery room sized for conventional loads may be undersized, improperly located, or architecturally incompatible with the power-path topology an AI tenant needs. Some AI clients arrive with a preference—or a requirement—for distributed battery architectures: in-hall UPS with lithium battery cabinets positioned closer to the load, or even rack-level battery systems integrated into their own equipment specifications. These are not exotic edge cases. They are emerging as standard expectations from the most commercially valuable tenants in the pipeline.
Then there is the calendar. AI clients frequently arrive with pre-set deployment timelines driven by GPU supply windows and model-training schedules. They do not want to wait 18–24 months for a redesign, new permits, or a fresh round of authority approvals. If your facility cannot accommodate their requirements within their window, they move on. The competitive pressure is real: other developers across Jakarta, Johor, Batam, and further afield are working to bring AI-ready capacity to market as fast as they can.
———
Centralized Battery Rooms vs In-Hall and In-Rack: Where the Real Design Commitment Sits
To understand the core trade-off, it helps to strip the battery decision down to its physical and regulatory fundamentals.
Centralized battery rooms—whether VRLA or newer LFP cabinets—concentrate the energy storage in dedicated spaces. Structurally and from a fire-engineering standpoint, this is familiar ground: clear separation from white space, well-defined fire ratings, straightforward egress, and established inspection routines. It is also easier, in many jurisdictions, to explain to authorities and insurers.
Distributed batteries—UPS systems and lithium cabinets in the hall, or sidecar and in-rack batteries integrated with GPU clusters—change the picture. The “fuel load” is now spread across the data hall. Floor loading in hot aisles and containment zones matters more. Detection and suppression strategies have to be reconsidered, especially for chemistries with different off-gas and thermal-runaway behaviour. Approval discussions move from a single room to the entire layout.
Most facilities in Southeast Asia will continue to rely primarily on centralized battery rooms for the foreseeable future. But halls designated for high-density AI workloads may increasingly push toward some form of distributed battery architecture—especially where the AI client’s own equipment specifications assume or require it. That is where the real design commitment sits today.
When you pour foundations and define your structural grids, you are no longer just committing to “demand will come.” You are committing, implicitly, to:
• Whether your anchor tenants will be AI-heavy or mostly conventional IT.
• Whether your design can practically and legally evolve from purely centralized battery rooms to selected in-hall or in-rack deployments without structural surgery.
• Whether your chosen chemistries and layouts will still be acceptable to authorities and insurers three to five years into the AI cycle.
———
Questions Boards, Investment Committees, and Regional Leaders Should Be Asking
Before the next foundation is poured or the next investment memo is signed, decision-makers across the region should be testing their assumptions with a sharper set of questions:
• Which tenant archetype is this facility genuinely designed for—AI-heavy, mixed, or legacy IT?
• What battery chemistries and placements does the base design assume, and which does it effectively rule out without major retrofit?
• If a key AI tenant insists on in-hall or in-rack batteries for specific halls, what does that mean for floor loading, fire zoning, capex, and approval timelines?
• How early have we engaged local authorities and insurers on our likely battery and fire-protection strategies, especially for any distributed lithium deployments?
• Do we have an internal view on how talent and operations will need to evolve if we shift from centralized VRLA/LFP rooms to distributed systems around AI clusters?
These are not questions to push down into technical appendices. They go to the heart of which projects will be truly “AI-capable” in practice and which will be restricted to a narrower tenant pool—and that is a strategic decision, not an engineering one.
———
Where These Commitments Play Out: SEA’s Emerging AI DC Lattice
The answers to those questions will look different depending on where in Southeast Asia the facility sits. AI-era capacity is not spreading evenly across the region. It is sorting into a lattice of hubs, spillover markets, and specialist campuses—each with its own mix of power, policy, talent, and capital.
Market | Emerging Role in AI DC Lattice | Power & Grid Situation | Structural Advantages | Key Constraints / Risks |
Singapore | Command-and-control AI hub; R&D and CoE base driving overflow into region. | Strong, stable grid but tightly rationed new DC allocations; power caps and efficiency demands push bulk AI loads offshore. | Dense AI CoEs, hyperscalers, top connectivity, deep talent pool. | Land, power, and policy limits; must lean on Johor and other hubs for large AI campuses. |
Malaysia | Primary AI capacity leader and Singapore spillover hub (especially Johor). | Large blocks of new power earmarked for DCs in Johor/Klang, but transmission and substation timelines becoming bottlenecks. | Multi-GW pipeline, Johor–Singapore proximity, lower costs, large DC parks. | Risk of grid congestion, approval tightening, and overbuild in certain clusters. |
Thailand | Cost-efficient AI DC and ops talent base. | Ability to add capacity in selected industrial/power corridors; grid strength and timelines vary by province. | Aggressive AI upskilling, competitive labour, growing AI/cloud demand. | Uneven permitting and grid readiness; needs more hyperscale-grade reference sites. |
Indonesia | Sovereign AI and domestic-scale anchor market. | Large generation potential but local grid constraints around Jakarta/Batam; upgrades and fuel mix critical for AI sites. | Huge domestic demand, data sovereignty push, sovereign fund + hyperscaler capital. | Grid build-out, permitting complexity, and policy/FX risk at national and regional levels. |
Vietnam | Low-cost specialist AI campus play. | Attractive tariffs and planned capacity additions but visible grid congestion and curtailment risks in some regions. | Lower build costs, strong AI/DC policy tailwinds, large AI DC pipeline. | Managing AI-era loads against grid constraints; regulatory growing pains at high density. |
Philippines | Late-rising, hyperscaler-anchored digital hub. | High prices and variable reliability; AI DCs concentrate near best-served nodes and new dedicated capacity. | Large digital economy, new hyperscaler DC and cloud projects, improving connectivity. | Power and permitting bottlenecks remain the main brake on very large AI campuses. |
Developers and operators who treat this lattice as part of their concept-stage decision-making—not an afterthought—will own the assets that AI tenants can actually use at scale. That means matching each site’s power realities to realistic AI rack densities, choosing battery and fire strategies that regulators will accept under those conditions, and placing operations teams where talent can be built and retained.
The same logic applies to suppliers and procurement teams: those who align products and buying criteria with where the region’s power and policy can actually support AI loads will quietly determine which projects become true AI platforms and which become stranded shells.
———
The Commitment You Are Already Making
Every data center project in Southeast Asia that is currently in planning or early construction is already making this commitment—whether the board has named it or not. The structural grid, the battery room allocation, the fire-engineering assumptions, the cooling topology: each of these locks in a position on which tenants the facility can realistically serve, which chemistries it can safely house, and which regulatory conversations it will face.
Seen through Design for Safety and Design for Operations, the implications are concrete:
• Battery strategy is now a concept-design decision, not a procurement detail. A generic centralized room will not serve all future tenants. You need an explicit position—before pouring concrete—on which halls may require in-hall or in-rack batteries and what that means for structure and fire engineering.
• AI rack density must be reflected in base-building criteria. If your default design still assumes 6–12 kW racks, you are baking operational strain into the building. Your ops teams will be asked to run AI loads in a facility not sized or zoned for them.
• Authority and insurer engagement has to move upstream. Waiting until late design to surface lithium chemistries and distributed UPS choices increases the risk that approvals will constrain what your operations team can actually implement.
• Operations teams need a seat at the planning table. The people who will live with alarm storms, emergency procedures, and shift-level safety decisions should be evaluating whether a proposed battery and power architecture is realistically operable at AI densities.
This is also how experienced data center staff stay relevant and valuable in AI-era facilities.
The developers who get this right will not necessarily be the fastest or the biggest. They will be the ones who made the commitment with open eyes—who decided, deliberately, what their facility is for, and built the safety regime, the operations model, and the team to match.
The rest will own expensive shells that operations and safety teams spend the next decade trying to work around.
———
This is Part 1 of a six-part series on AI-era data centers in Southeast Asia. Part 2, Tech, Ops and Talent: Building AI-Era Operations Teams, examines what AI-density facilities demand from org structures, skills, and training—and where the human-factor risks are sharpest.
For a deeper treatment of the infrastructure stack, battery architectures, and operations frameworks discussed here, see Data Center Primer (ISBN 9789819439768). I am also available for closed-door briefings and customised training for project teams, operations leaders, and investment committees navigating the AI transition.
Comments