Artificial intelligence’s surge from lab curiosity to everyday utility has created a parallel surge in physical infrastructure: enormous, power‑dense data centres that swallow electricity, water and silicon to deliver the responses users now take for granted. The Oman Observer piece on “The hidden costs of AI data centres” outlines how those resource drains stack up — from per‑prompt energy use and the water needed for cooling to the capital and lifecycle costs of GPUs — and sounds an urgent warning: the economic promise of AI must be matched with environmental and fiscal accountability.
Background
AI’s compute requirements scale nonlinearly. Training state‑of‑the‑art models and serving them at global scale require vast pools of GPUs or accelerators operating continuously in facilities designed around very high power density. That shift has consequences across three interlinked domains: energy demand (electricity and its carbon intensity), water (cooling and the water embedded in electricity), and hardware lifecycle costs (purchase, refresh, disposal). These are not hypothetical: international agencies and industry reporting show data‑centre electricity demand has already moved from a niche operational concern to a major electricity sector driver. The International Energy Agency (IEA) estimates global electricity consumption for data centres, cryptocurrencies and AI at about 460 TWh in 2022 and projects a range that could exceed 800 TWh — and in high‑growth scenarios approach or surpass 1,000 TWh by 2026.Those numbers matter because they transform data centres from enterprise infrastructure into grid‑scale consumers. In the United States, for example, energy officials now explicitly warn that data‑centre growth tied to AI could push regional electricity demand to new records and complicate decarbonisation plans.
How AI workloads amplify resource intensity
Density and duty cycle: why AI data centres are different
Modern AI models are trained on massive datasets using clusters of accelerators and then operated as always‑on inference services. Two characteristics of these workloads make them particularly demanding:- Power density: AI racks commonly draw tens of kilowatts each; high‑density GPU racks (50–80 kW or more) are now standard in AI‑first facilities. That is roughly an order of magnitude above typical enterprise racks and forces different designs for power distribution, cooling and site resiliency.
- Duty cycle: Training runs can run for days to weeks and inference fleets often serve continuous, unpredictable loads at global scale. That eliminates much of the seasonal or partial‑load slack that once allowed air economization to do most of the cooling.
Energy: from single queries to gigawatt campuses
Estimating per‑query energy use is fraught, but the best available analyses show that modern inference is far more efficient than early alarmist figures suggested — and still material at scale. Recent industry and independent analyses converge on a ballpark that a single interactive large‑model inference can consume on the order of a few tenths of a watt‑hour (roughly 0.3–0.4 Wh) for efficient production deployments; older studies that compared early ChatGPT instances to a Google web search overstated the difference because both the models and serving stacks have improved. At global scale, even a fraction of a watt‑hour per prompt becomes terawatt‑hour‑level electricity demand.Training presents a different scale of cost. Public disclosures about the compute used to train frontier models are sparse and sometimes contradictory; independent reconstructions are often based on model FLOP estimates and known hardware performances, and therefore contain large uncertainties. Open reporting rarely publishes a single, auditable kilowatt‑hour tally for training flagship models, and the widely circulated figure that “GPT‑4 training used ~1.5 GWh” appears in secondary reporting rather than as a verified operator disclosure. Multiple industry analysts report that training modern foundation models frequently requires clusters composed of tens of thousands of accelerators over extended periods, which translates to megawatt‑scale power draws during those runs — and that aggregate training and inference demand is reshaping hyperscaler capex and utility planning. These claims are credible, but they vary widely by model, optimization, precision mode, and data‑centre efficiency; treat single‑number training estimates as illustrative, not definitive.
Water: the overlooked constraint
Cooling is where electricity converts into a water problem. Evaporative cooling and wet recirculation remain among the most energy‑efficient ways to reject heat at very high densities, but they consume water. Technical analyses and reporting indicate that training a large model or operating millions of inferences can carry a non‑trivial water footprint once on‑site cooling plus the water embedded in electricity generation are both included.One oft‑cited calculation attributes roughly 500 millilitres of water to every 20–50 conversational prompts when counting both on‑site evaporative losses and the water used in electricity production; the figure originated in academic and technical breakdowns that emphasize sensitivity to location, cooling architecture and grid mix. Separately, a university researcher’s estimate that a GPT‑3‑scale training run could consume on the order of hundreds of thousands of litres of cooling water has appeared in multiple reports; these estimates are highly dependent on whether the accounting includes construction and ramp phases, closed‑loop makeup water, and whether non‑potable or reclaimed sources are used. These numbers show why communities near proposed AI campuses are concerned about water allocations and why municipalities are increasingly asking for volumetric metering and enforceable limits.
True costs beyond the utility bill
Capital and replacement: GPUs, racks, and refresh cycles
Leading AI accelerators do not come cheap. High‑end accelerators used in hyperscale training clusters — for example, modern Blackwell/H100‑class cards — have been listed at tens of thousands of dollars per unit; industry pricing guides and resale markets place new H100 boards commonly in a range roughly between $25,000 and $40,000 depending on variant and supply conditions. Multiply that by hundreds of thousands of GPUs and the capital cost for a megasite quickly reaches into the billions. The hardware lifecycle — typically a 2–3‑year refresh cadence for competitive performance — turns replacement into an ongoing multibillion‑dollar annual item for the largest operators.Retirement and disposal add real costs, too. Secure decommissioning, data eradication, and responsible recycling of high‑value boards and server chassis are nontrivial; vendors and specialists charge for secure wipes and handling, and e‑waste infrastructure is uneven across jurisdictions. The environmental and regulatory burden of disposing of high‑density compute gear should be treated as an operational cost — not as an afterthought.
Operational engineering complexity and grid interaction
High‑density AI facilities have non‑linear cost behavior. When rack density increases, incremental megawatts require stronger electrical substations, upgraded distribution systems, and advanced cooling plants. Efficiency gains from economies of scale show diminishing returns as thermal management, transformer losses, and distribution inefficiencies kick in. Operators sometimes lease on‑site generation, procure firm power contracts, or commit to long‑term PPAs — each choice influences grid stability, local electricity prices, and emissions profiles. Recent reporting shows utilities and regulators are having to rework interconnection planning because data‑centre growth is now a system‑level driver of demand.Reputational and community costs
When a new campus requests municipal water allocations or conditional potable draws, downstream political and social costs arise. Residents, farmers and local businesses see data centres as new industrial neighbours that may strain shared resources. Where early permit estimates have excluded construction or peak‑day scenarios, measured consumption has sometimes exceeded public expectations — generating legal disputes, permit renegotiations, and reputational losses. These social externalities can be costly and materially delay operations.Where operators are making (and can make) progress
Cooling and water strategies
- Air‑first designs: prioritize free‑air economization where climate permits so water use remains minimal for most of the year.
- Closed‑loop liquid cooling and immersion: these approaches reduce evaporative make‑up but still require heat rejection strategies and occasionally makeup water; closed loops also enable dense packing and lower energy per operation.
- Non‑potable sourcing and reuse: treated wastewater, captured stormwater, and on‑site reclamation can replace potable supply for cooling, subject to regulatory approval and treatment costs.
- Heat recovery for district heating: when feasible, recovered heat can be redirected to local district heating systems — a notable example is Google’s Hamina site in Finland, which uses seawater cooling and will route waste heat into the municipal district heating network, offsetting local fossil‑fuel heating and improving system‑level efficiency. Those projects require careful planning and local partnership but illustrate practical industrial symbiosis.
Energy strategies
- Model compression and distillation: reducing model size where possible and deploying smaller specialized models for routine tasks can cut inference energy dramatically.
- On‑site renewable generation and PPAs: siting near hydro, wind, solar, or low‑carbon firm power reduces lifecycle emissions but must be paired with grid integration strategy to prevent fossil fuel firming during peaks.
- Demand‑side flexibility and scheduling: shifting non‑urgent training to low‑carbon periods, using batteries to smooth demand spikes, and optimizing batch sizes help reduce marginal carbon intensity and costs.
Procurement and hardware lifecycle management
- Purchase vs. cloud tradeoffs: cloud bursting and use of provider‑managed H100 fleets can avoid upfront capital and refresh logistics, though unit economics change at sustained scale.
- Extended use and secondary markets: refurbishing older accelerators for lower‑duty tasks extends useful life and reduces e‑waste flows.
- Transparency and auditing: require audited lifecycle carbon footprints and WUE (Water Usage Effectiveness) metrics as part of procurement and permitting.
Policy and disclosure: closing governance gaps
The sector suffers from opaque reporting on two central metrics: energy intensity at the facility/mode level and true consumptive water use. Meaningful public policy levers include:- Mandatory facility‑level disclosure of PUE (Power Usage Effectiveness), WUE (Water Usage Effectiveness) and lifecycle carbon intensity for projects above defined computational or megawatt thresholds.
- Permit conditions requiring volumetric potable‑water metering, peak and hourly reporting during the commissioning phase, and clear separation of construction‑phase vs. operational water use.
- Fiscal incentives (tax credits, grants) conditioned on verified non‑potable water sourcing, heat‑reuse commitments, and demonstrable reductions in energy per inference.
- Utility planning coordination that treats large campus interconnections as system assets — requiring grid‑planning transparency and robust contingency rules for emergency demand events.
- Public R&D funding for model efficiency research (compression, distillation, algorithmic optimizations) to reduce required FLOPS per useful output.
Risks, tradeoffs and hard truths
- Efficiency gains are not a panacea: technical progress reduces energy per operation, but expanding total demand can still raise absolute consumption. Efficiency improvements often encourage more use — the rebound effect — that can offset gains.
- Water‑energy tradeoffs: switching from water‑intensive evaporative cooling to electric chillers reduces freshwater consumption but increases electricity demand and possibly emissions unless paired with low‑carbon power.
- Opacity in training emissions: public statements and vendor marketing rarely provide full lifecycle or peak‑load energy disclosures for major training runs. Reconstructing training footprints from model size and hardware assumptions is possible but uncertain; policy should push for auditable operator disclosures rather than reliance on reverse engineering.
- Inequitable geography: siting decisions matter — placing a water‑intensive campus in an arid region transfers risk to local communities. Conversely, coastal, cool climates, or regions with abundant low‑carbon electricity may be optimal for density but still require careful social license and local benefit.
Practical recommendations for industry and regulators
- Make WUE and PUE mandatory for major AI compute projects and publish monthly operational metrics for the first two years after commissioning.
- Tie public incentives and large procurement contracts to audited lifecycle carbon and water disclosures.
- Prioritise siting near low‑carbon firm power or industrial heat users to enable waste‑heat reuse.
- Fund model‑level efficiency R&D and require that large public grants support demonstrably efficient architectures (compression, distillation, sparse or conditional compute).
- Implement binding volumetric caps on municipal potable water withdrawals for cooling, with clearly defined escalation and remediation steps if peaks approach critical thresholds.
Conclusion
AI data centres are the new heavy industry of the digital age: they produce valuable services but also concentrate electricity, water and material impacts into discrete campuses that intersect with local communities, national energy systems and global carbon budgets. The Oman Observer’s framing — that AI’s promise must be matched with accountability — is accurate and urgent.Credible international analysis shows data‑centre demand is already a grid‑level issue, with projections that could push consumption into the high hundreds of terawatt‑hours within a few years unless moderated by efficiency, siting and policy. Practical solutions exist: model efficiency, smarter procurement, closed‑loop and non‑potable water systems, heat recovery projects like Google’s Hamina initiative, and stronger public disclosure rules would together shift the industry toward what could be called responsible intelligence.
That transition will not be automatic. It will require honest, audited disclosures from operators, regulatory teeth from municipalities and utilities, and continued investment in software and hardware approaches that reduce the compute — and therefore the environmental — cost of every useful AI output. Success will belong not to those who build the biggest clusters, but to those who prove they can deliver AI at scale without outsourcing the true costs to communities, grids and future generations.
Source: Oman Observer The hidden costs of AI data centres
