AI can raise a data centre’s energy and materials footprint quickly. The best results come from a few practical choices: pick a cooling strategy that matches real rack densities and site conditions; tidy the power path so less energy becomes waste heat; and treat end-of-life equipment as part of a circular programme, not an afterthought. Do these well—and report them clearly—and AI growth can align with efficiency and sustainability targets.

Set the Rules of the Game: Energy and Water Baseline

Before any new hardware ships, build a simple model: expected rack densities over time, hours of use, heat-rejection method, and water implications. Define how to calculate PUE (power usage effectiveness) and WUE (water usage effectiveness) so year-on-year comparisons are fair and consistent.

Industry surveys show average PUE has been largely flat, with a weighted average around 1.54 in 2025—suggesting most improvements now come from design choices rather than easy tuning. Pair that baseline with a view of demand: independent research expects data centre power demand to rise sharply through this decade, with scenarios up to 165 percent higher by 2030 versus 2023—reinforcing the need to use each kilowatt well.

Not Just a Slogan: Choose Cooling That Delivers Real Gains

Cooling should follow a decision tree tied to density and site constraints, not a slogan. In practice, many operators move along a path from contained air systems to rear-door heat exchangers (RDHX), then direct-to-chip (DTC) cooling, and finally immersion at the highest densities. Typical RDHX envelopes today range from ~40–60 kW per rack, DTC ~60–120 kW, and immersion ≥100 kW; dual-phase reports higher.3 Use these as planning ranges, then confirm with site tests.

On environmental performance, recent life-cycle assessment work indicates liquid-cooled designs can reduce operational and embodied carbon in modelled scenarios—sometimes by roughly up to half versus conventional air systems—while also lowering water use, though results depend on local grid mix, climate, and specific design choices. If any dual-phase fluid under consideration contains PFAS, a health and environmental review needs to be part of the selection process.

Tighten the Power Path—and Right-size Resilience

Energy lost in conversion becomes heat that has to be removed. Some operators are evaluating 48-volt server power trains to cut conversion losses; published tests show loss reductions of at least 25 percent in certain trials. Another lever is resilience strategy that reflects workload type: training jobs are typically less time-critical than transactional or customer-facing systems and can run with lower power redundancy; inference often keeps stricter requirements. The aim is not to take risks—it is to match protection levels to business need and avoid waste.

Track the effect of these choices by watching PUE alongside electrical measurements at key nodes. Fewer conversions and better alignment between workloads and redundancy targets should reduce waste heat and improve stability.

Put Workloads Where Energy Is Cleanest—and Capacity Is Real

Location matters. Training usually tolerates higher latency and can be sited in power-abundant regions; inference often benefits from proximity to users and resilient network paths. Plan the interconnect early, including any transmission constraints, so the promised benefits are achievable in practice. The macro picture adds urgency: demand growth continues to pressure available capacity in many markets.

When comparing sites, include grid carbon intensity and water availability in the decision, not just space and price. A slightly different location can yield lower emissions for the same compute.

Retire Equipment Securely and Support Circularity

AI refresh cycles retire equipment quickly. That is both a security obligation and a chance to recover value. Use IT asset disposition (ITAD) services that sanitise media to recognised standards—NIST SP 800-88 and IEEE 2883—and issue a certificate that records who, what, where, when, and how for each asset. Keep an asset-exit ledger so audits are fast and complete.

For environmental assurance, many buyers look for recognised programmes such as R2 (administered by SERI) and e-Stewards (Basel Action Network) to show responsible reuse, recycling, and worker-safety practices across the chain.

Report What Matters—and Make It Comparable

Keep reporting simple and consistent so trends are clear quarter to quarter. Use the same boundaries, time window, and definitions each cycle.

  • Energy & water: Track PUE and WUE with clearly stated boundaries. Add a short note when adverse weather conditions or operating hours shift the numbers.
  • Design changes & embodied carbon: When changing cooling or the power path, include a line highlighting “what changed and why,” plus a brief nod to embodied carbon so recipients interpret the trend correctly.
  • Asset exit & circularity: Maintain an asset-exit ledger listing retired items, the share with NIST SP 800-88/IEEE 2883 data-sanitisation certificates, and the share reused/resold vs recycled. This closes audit gaps and shows progress on circularity.

Consistency matters more than precision; the aim is a record that helps steer decisions next cycle, not a marketing figure.

What This Delivers

An energy-first plan does not slow AI adoption; it prevents rework. The right cooling choice for actual densities improves stability and lowers emissions; a tidier power path reduces waste heat; and a circular exit plan protects data, recovers value, and cuts material impacts. Given the growth outlook for data centre power demand, these choices are how AI scale can fit within energy, water, and compliance limits.

Feature written by Ian Shearer; Managing Director, APAC & EMEA, Park Place Technologies

For more news, click here.