Introduction
Artificial Intelligence has become the defining technology of the decade. From powering medical breakthroughs to reshaping education and business, it is everywhere. But behind every AI chatbot reply or image generation sits a sprawling network of energy-hungry data centers. The world is beginning to realize that the cost of scaling these systems is not just financial. It is environmental, and it is rising faster than policymakers and engineers can adapt.
- Scale and comparative context
Global trajectories shaping the debate
Artificial intelligence is now central to global infrastructure. The International Energy Agency projects that data-centre electricity demand could double to around 945 terawatt-hours per year by 2030 in its base case scenario, driven largely by AI workloads. That level is roughly 2 to 3 percent of global electricity and places AI among the most consequential energy debates of the decade.
How AI compares with streaming, gaming and crypto
For context, streaming video is estimated to account for roughly 1 to 4 percent of global emissions in various studies, while Bitcoin mining has recently consumed on the order of 100 to 150 TWh annually. AI differs because its compute is concentrated in provider infrastructure rather than on billions of end devices. That concentration shifts the policy question from device efficiency to grid capacity, transmission planning and facility siting. Public debate must therefore treat AI as a distinct infrastructure category.
- Per-interaction accounting and measurement risks
Real per-query costs
Viral conversions like “15 bottles of water per prompt” are misleading because they collapse different units and lifecycle boundaries. Measured production instrumentation from recent provider reporting (instrumentation disclosed in 2025) shows median text prompts on a well-optimized serving stack at about 0.24 watt-hours and 0.26 milliliters of water per prompt. Independent academic benchmarks record a wider range: roughly 0.4 Wh for short optimized queries up to multiple watt-hours for longer or less efficient model configurations. This is why per-query statements must always include context on model, prompt length, hardware and cooling approach.
Measurement gaming and the greenwashing risk
Partial disclosures create a false sense of progress. Some providers report time-averaged renewables or offsets while excluding embodied emissions and water use for cooling. To avoid gaming, disclosure should require stack-level, time-resolved metrics: watt-hours per service, grams CO₂e per service by hour and location, and milliliters of water per service, published with independent audits. Only then can procurement, regulation and civil society compare like with like.
- Electricity mechanics and grid politics
Clustering and who pays for upgrades
AI racks are high density. Where hyperscale facilities cluster, utilities must upgrade transmission and distribution. In several data-centre corridors this has triggered urgent investment decisions and questions about who bears the cost. Passing those network upgrade costs to local ratepayers raises equity issues: communities can end up financing infrastructure that primarily benefits global cloud customers.
Grid mismatch and fossil lock-in
AI load profiles, consisting of continuous serving plus intermittent heavy training, do not neatly follow solar or wind production. Utilities therefore procure fast-ramping capacity, often gas peakers, to ensure reliability. Those investments risk locking in fossil generation and increasing local emissions unless new loads are paired with firmed low-carbon supply.
- Water, cooling architecture and the Global South
Cooling systems and their water toll
Cooling is an often invisible source of water demand. Engineering studies and industry analyses show that a 1 MW facility using direct evaporative or adiabatic cooling can consume on the order of 25 million liters of water per year, though actual numbers vary by design and climate. At scale that converts per-query milliliters into billions of liters regionally. In water-stressed basins this is not abstract; it is a direct resource trade-off.
Burden shifted to the Global South
The controversy is shifting beyond deserts in North America. Cities and regions in India, parts of Africa and the Middle East are courting data campuses even as they face water stress. For example, urban water shortages in Bengaluru have been connected in reporting to rapid infrastructure growth. In the Gulf, some operators pair data centers with desalination, but desalination is energy intensive and typically adds to overall emissions. The siting choices for digital infrastructure therefore have geopolitical and justice implications.
- Carbon, supply chains and waste
Training, retraining and the serving tail
Training frontier models can emit on the order of hundreds of tons of CO₂e for a single large run. Repeated retraining and wide deployment scale that footprint. Serving billions of prompts daily adds a continuous carbon tail whose intensity depends on the local grid mix and the time of day the energy is drawn. Because of this variation, identical models can have very different carbon footprints across jurisdictions.
Embodied carbon, materials and e-waste
Chip fabrication and server manufacture are resource intensive. Mining for semiconductors and rare minerals concentrates environmental harms and social costs. Global e-waste reached 62 million tonnes in 2022, and that figure is set to rise as hardware refresh cycles accelerate. These supply-chain and end-of-life burdens are often excluded from per-query metrics but can dominate lifecycle impacts.
- Justice, labor and unequal burdens
Who bears the costs
Local communities frequently bear the externalities: rate increases for network upgrades, competition for water, and local environmental impacts. Low income and rural populations often have the weakest bargaining power while large cloud providers capture global value. Recognizing this distributional reality is crucial to designing equitable policy responses.
Community benefits and possible models
Where permitting and procurement have required community benefit agreements, local outcomes improve. Conditioned approvals that secure local renewable supply, workforce development and community investments provide a governance template for fairer integration of AI infrastructure.
- Rebound effects and alternative futures
Efficiency gains do not guarantee lower demand
Efficiency improvements can be offset by increased use. Modeling of rebound dynamics indicates that material efficiency gains can lead to proportionally larger increases in demand in some contexts. Policy must therefore combine efficiency with demand management, procurement rules and usage controls to capture net emissions reductions.
Plausible 2030 trajectories
Policy choices determine whether global data-centre demand stays near the lower bound of scenarios, plateaus under regulatory and market pressure, or grows into the higher bound that stresses grids, water systems and climate targets. The IEA baseline and alternative scenarios provide quantitative anchors for each path.
- Governance and global coordination
Standards and global reporting baselines
National measures are emerging in the EU and in some U.S. states, but international baselines are needed to prevent jurisdictional arbitrage. Standardized, audited disclosures for energy, water and carbon are a first order governance requirement.
Trade and finance levers
Development finance institutions, export credit agencies and multinational procurement can enforce conditionality. Verified low-carbon and low-water metrics as requirements for financing and procurement significantly reduce incentives to export impacts to weaker jurisdictions.
- Climate feedback quantified and the policy imperative
Climate and engineering literature indicate that warming substantially increases cooling demand in many regions. Multiple studies show cooling demand grows measurably with temperature, and engineering guidance commonly places the sensitivity in the low single digits to low double digits percent per 1 degree Celsius in many climates. The practical implication is clear: each degree of warming increases cooling load, which increases electricity demand and often emissions if marginal supply is fossil. That creates a compounding feedback loop that policymakers can measure and act upon using time-resolved metrics and siting rules.
Action checklist
- Audit and disclose per-service metrics in real time: watt-hours, gCO₂e and milliliters of water, reported by location and time.
- Mandate eco-efficiency in procurement: favor smaller, task-specific models where they meet business needs.
- Ban high-water evaporative cooling in water-stressed basins and require alternatives in new permits.
- Require new hyperscale sites to secure firm renewable power plus storage before connection to constrained grids.
- Enforce environmental justice provisions: community benefit agreements and transparent permit conditions.
- Use international procurement and finance rules to prevent regulatory arbitrage.
Conclusion
The collision between AI infrastructure and planetary limits is visible today in grid planning debates, local water tensions and emerging e-waste challenges. The choice ahead is pragmatic: align disclosure, procurement and siting rules so AI can contribute to prosperity without undermining climate and resource stability. Clear, audited metrics and enforceable standards are the first line of defense. With them, AI can be shaped into a tool that supports resilience rather than accelerates systemic risk.
