What Sustainable AI Has to Prove

An AI response arrives in seconds. Behind it sits a building full of chips, power equipment, and cooling systems. Victor Kiani investigated whether that infrastructure can become environmentally sustainable, gathering research on its growing demands and the environmental benefits AI could deliver. He developed and presented an argument about the conditions that make responsible growth possible.

Academic research project completed and presented for a sustainability class in 2026.

Count the whole system

Kiani began by defining the assessment boundary. A server's energy efficiency captures only part of the resources required to deliver an AI service. He organized the research around three connected scales:

1. 01The facilityElectricity, cooling water, backup fuel, and noise: the operating demands of keeping computing equipment running.
2. 02The locationGrid capacity, the watershed, and nearby residents: the setting determines who shares the resources and experiences the impacts.
3. 03The life cycleMining, chip production, and electronic waste: environmental pressures begin before hardware arrives and continue after replacement.

Efficiency and growing demand

The central tension emerged in the electricity research. Energy use per AI task was falling, while wider adoption and more demanding applications drove total consumption upward. The IEA's 2026 assessment estimated that all global data centers used 485 TWh in 2025 and projected 950 TWh in 2030. Electricity consumption at AI-focused sites alone grew 50% in 2025.

Kiani used this contrast to argue for evaluating total environmental impact alongside efficiency. Lower consumption per task matters, but expanding use changes the scale of the problem. Capacity planning and environmental commitments have to account for both.

The opportunity on the other side

The presentation also examined AI's potential to improve building controls, traffic routing, weather forecasting, and grid operation. The IEA's 2025 widespread-adoption scenario estimated that existing applications could avoid 1.4 gigatonnes of CO₂ in 2035. Sensors and AI-assisted grid management could unlock up to 175 GW of transmission capacity on existing lines.

Kiani connected those possibilities to their enabling conditions: investment, access to data, and adoption. His argument preserved the opportunity while making verified environmental benefit part of the assessment. A promising application earns its place through what it delivers.

Water, neighbors, and hardware

Water made the relationship between computing and place concrete. Berkeley Lab estimated that all U.S. data centers consumed 66 billion liters directly in 2023, with nearly 800 billion liters consumed indirectly through their electricity supply. Kiani considered dry cooling and the importance of local drought and groundwater conditions when evaluating a facility.

He also examined noise as a social impact. Virginia's legislative research commission found that the constant sound from some facilities affected nearby residents even when it rarely exceeded local noise limits. The example connected environmental performance to community well-being and the adequacy of existing rules.

For hardware, he used UNITAR's estimate of 62 million tonnes of global electronic waste in 2022, of which 22.3% was documented as properly collected and recycled. These figures describe all electronics. The presentation placed AI hardware within that wider material system, using a truck comparison and an illustration to make its scale understandable.

Power that matches the load

Kiani tested the meaning of a renewable-energy claim against a practical question: what powers the facility at 2 a.m.? The IEA's 2025 outlook projected that renewables would meet nearly half of added data-center electricity demand through 2030, with gas and coal supplying more than 40%.

He argued for new low-carbon generation, storage, and transmission, matched to demand hour by hour. Nuclear offered a further source of steady low-carbon power, with cost, construction time, waste, and cooling-water considerations. The analysis connected energy procurement to the infrastructure required to support it.

Four questions for accountability

The conclusion brought the research into four questions:

1. 01What is the total impact? Account for electricity, carbon, water, and hardware.
2. 02Who experiences it? Identify the communities and resources affected.
3. 03What limits and reporting rules apply? Establish how performance becomes accountable.
4. 04What benefit is actually verified? Connect environmental claims to evidence.

Using the EU's annual reporting requirement for data centers with at least 500 kW of installed IT power as a starting point, Kiani proposed public energy, water, emissions, and hourly power-source data, together with local water and noise limits.

For businesses expanding AI capacity, these questions connect sustainability to operating costs, resource availability, and community trust. Site selection, power contracts, and equipment choices determine whether environmental commitments remain credible as demand grows. Kiani's conclusion was conditional: AI data centers can become more sustainable when total impacts fall and environmental benefits hold up to verification. He translated that argument into an eight-slide presentation with speaker notes, making a complex infrastructure question accessible through research and concrete examples.