AI Data Centres: Hidden Costs Singapore Must Plan For

AI isn’t just “another workload” you run in the cloud anymore. It’s starting to dictate where capacity gets built, how fast it comes online, and what it costs to keep it running. When the infrastructure behind AI tightens—power, cooling, water, approvals—every business that depends on AI tools feels it, even if you never plan to own a server rack.

For Singapore companies adopting AI business tools (for marketing, operations, and customer engagement), this matters because the real constraint isn’t ideas. It’s compute availability at a predictable price. If your 2026 plan assumes you can scale AI usage endlessly with stable costs, you’re budgeting for a world that’s disappearing.

The good news: you can plan for this. The companies that get ahead won’t be the ones chasing the newest model every quarter. They’ll be the ones building an AI adoption strategy that treats infrastructure as a risk factor—the same way you treat compliance, cybersecurity, or supply chain resilience.

AI is redesigning the data centre business model (and your cloud bill)

AI is forcing a shift from “build once, operate for decades” to continuous refresh and reconfiguration. That’s not a small operational tweak; it changes financing, project timelines, and how operators think about risk.

Meiske Sompie (Asia partner at TBH, a construction planning and management consultancy) put it plainly: the pace of AI demand growth is now outstripping the ability of energy systems, water planning, and regulatory frameworks to respond. The mismatch has moved from being a technical headache to becoming a hard constraint on expansion.

Here’s the practical consequence for Singapore businesses using AI tools:

• When data centres can’t add capacity quickly, compute becomes scarce, and pricing power shifts to providers.
• When power and sustainability requirements tighten, operators spend more on compliance and engineering—costs that eventually show up in contracts and consumption pricing.
• When timelines matter (and they do), delays translate into higher risk premiums—and higher prices.

A line I keep coming back to: AI turns infrastructure limits into business limits. If you’re building AI into customer support, lead qualification, fraud detection, or content production, you’re implicitly tying your growth plans to data centre realities.

The new ROI math: from static assets to lifecycle planning

The core shift is simple: AI hardware ages faster than traditional data centre gear.

Sompie notes that AI workloads require higher rack densities and significant power, which accelerates depreciation and shortens the lifecycle of key assets—especially chips and networking equipment. In other words, the old “capex up front, stable ops for years” model breaks down.

Why lifecycle thinking is now the only sensible approach

For operators, lifecycle planning means designing for frequent hardware refreshes and phasing capital expenditure around expected upgrade cycles. For Singapore businesses that don’t own data centres, it still matters because your vendors and cloud providers are making the same calculation—and passing it through in product structure.

What I’m seeing more of in the AI tools market:

• Usage-based pricing that can swing more sharply as compute costs change.
• More “AI add-ons” and tiered packages that separate basic automation from compute-heavy features.
• Greater emphasis on regional capacity (where the workload runs starts affecting performance and cost).

Actionable move for Singapore SMEs and mid-market teams

Stop approving AI tools purely on demo performance. Add two procurement questions:

1. What drives our cost if usage doubles? (tokens, seats, API calls, model tier, peak-time pricing)
2. Where does compute run and how is capacity reserved? (Singapore region vs elsewhere, dedicated capacity options, SLA clarity)

If a vendor can’t answer clearly, you’re taking on price and availability risk blind.

Power and water constraints aren’t “data centre problems” anymore

The uncomfortable truth: AI adoption has a physical footprint. Power constraints affect communities, developers face rising costs to build and operate greener facilities, and utilities need long-term planning. This isn’t theoretical; it’s already shaping project viability across Asia.

Sompie highlights the most concerning data point: the widening gap between AI growth and energy infrastructure deployment. That gap forces a rethink of how facilities are financed, built, and operated.

Why water is now part of the AI conversation

Water scarcity has become a standard planning requirement for new data centre projects, particularly in Asia. Sompie points to water stress in places like Johor, parts of Peru, and India. In Johor Bahru, operators have reportedly been told they may need to wait until mid-2027 for sufficient water supply access.

The takeaway for Singapore businesses is not “don’t use AI.” It’s this:

If your AI strategy depends on cheap, infinite compute, you’re depending on cheap, infinite power and water. That’s not the direction the region is heading.

As a result, data centre clients are increasingly discussing renewable or recycled water solutions, including on-site water recycling plants and developing data centres alongside dedicated water treatment facilities.

Practical implications for your AI rollout

If you’re planning to scale AI across departments in 2026:

• Build scenarios where AI unit costs rise 15–30% over 12–24 months (not a prediction—just prudent planning).
• Treat latency and regional hosting as a product requirement, not an afterthought.
• Prioritise AI tools that offer efficient modes (smaller models, caching, retrieval-augmented generation, batch processing).

Efficiency isn’t just engineering purity. It’s cost control.

Stranded assets: the warning signs businesses should watch (even if you’re renting)

Sompie argues that data centres that aren’t modular or AI-optimised are likely to become redundant within the next decade. That raises stranded-asset risk—facilities that can’t adapt to higher-density workloads, new cooling methods, or upgrade cycles.

Even if you never build a data centre, stranded assets show up downstream as:

• providers pushing customers off older offerings,
• sudden contract changes,
• migrations that weren’t on your roadmap,
• performance ceilings you can’t pay your way out of.

Early warning signs (translated for AI tool buyers)

Sompie’s planning signals map surprisingly well to what buyers should look for:

1. Escalating operating costs → vendor price increases or shrinking “included usage” allowances.
2. Outdated cooling / infrastructure → unreliable performance at peak demand, reduced availability of high-end GPUs.
3. Limited flexibility to scale → long lead times for capacity upgrades, queueing for dedicated instances.
4. Hard-to-upgrade systems → tool stagnates, can’t adopt new models or safety features without painful replatforming.

Cooling is a big piece of this. Sompie lists three commonly used cooling technologies: closed-loop systems, immersion cooling, and direct-to-chip liquid cooling (currently widely adopted in Asia due to a more mature supply chain).

If your AI vendor is vague about how they maintain performance and cost as models scale, assume they’re exposed to these constraints.

Who pays to future-proof AI infrastructure—and why you should care

Sompie’s answer is basically: everyone has a role—operators, governments, chipmakers, utilities, and especially banks/investors because financing shapes incentives. The tension is real: lenders often optimise for short-term profitability, while sustainability investments pay off over longer horizons.

Singapore is a useful example of coordinated standards. The Building and Construction Authority (BCA) and IMDA have introduced green building standards that data centre projects must comply with. In a constrained market, compliance becomes a functional incentive—meet the standard or you don’t build.

What this means for Singapore businesses adopting AI business tools

Expect more AI procurement to include sustainability and resilience questions:

• Where is the workload hosted, and what’s the provider’s energy strategy?
• Do they offer reporting aligned to your ESG requirements?
• Can they support data residency and performance needs without cost blowouts?

This isn’t “nice to have” paperwork. Carbon and sustainability targets are increasingly tied to business resilience—infrastructure reliability, insurance costs, capital availability, and supply chain stability.

A practical playbook: scale AI in Singapore without getting surprised

You can’t control the regional power grid. You can control your AI operating model. Here’s what works in practice when you’re rolling out AI for marketing, ops, and customer engagement.

1) Classify your use cases by compute intensity

Not every AI workflow needs the biggest model.

• Low intensity: summarisation, email drafting, internal FAQ search
• Medium: customer service copilots with retrieval, sales call analysis
• High: video generation, advanced agent workflows, large-scale forecasting

Then match tooling accordingly. For many teams, using smaller models for 70% of tasks cuts cost dramatically with minimal quality loss.

2) Build a “cost per outcome” metric

If marketing is using AI, don’t track “tokens used.” Track:

• cost per qualified lead,
• cost per campaign variant produced,
• reduction in turnaround time (hours saved × loaded salary rate).

This keeps the discussion grounded when compute prices wobble.

3) Design for caching, batching, and retrieval

Three tactics that reduce ongoing compute:

• Caching: reuse answers for repeated questions (support, policy, product info)
• Batching: run non-urgent tasks off-peak
• Retrieval-augmented generation (RAG): keep prompts smaller and reduce hallucinations

You’ll feel this immediately in both cost and latency.

4) Don’t ignore vendor concentration risk

If one provider throttles capacity or raises prices, do you have a Plan B?

• Keep prompts and system instructions portable.
• Avoid proprietary lock-in where you can (especially for embeddings and vector databases).
• Maintain a fallback model tier for “good enough” operation.

5) Treat sustainability as procurement hygiene, not PR

Community opposition is a real risk in some markets (Sompie mentions South Korea as a cautionary example). In Singapore, regulation is already pushing greener builds. Your suppliers will be asked to prove sustainability; you might as well pick vendors who are ready.

AI makes infrastructure a board topic—so bring it to the budget early

AI demand is pushing renewable energy, sustainability, and long-term infrastructure planning higher on executive agendas than previous technology waves. That’s the real shift. AI isn’t only changing how work gets done; it’s changing the cost structure of the systems underneath.

If you’re leading AI adoption in Singapore, plan like compute constraints are normal, not exceptional. Pick AI business tools that can flex across model sizes, measure ROI in outcomes, and build workflows that are efficient by design.

The question to bring to your next planning meeting is simple: If AI usage doubles next year, do we have a cost-and-capacity plan—or just hope?