The Sustainability Paradox of Urban AI Infrastructure: When Smart Cities Meet Energy and Water Bottlenecks

In the past few years, cities worldwide have been racing to deploy AI infrastructure, from autonomous vehicle fleets to real-time urban digital twin platforms, with AI becoming the operating system of digital cities. However, a recent review article published in Nature Reviews Clean Technology reveals a stark reality: the AI data centers supporting these smart applications are consuming energy and water at an unprecedented rate, and without intervention, they will turn cities' climate commitments to dust.

The Exponential Growth of Data Centers

According to the study titled Strategies and Design for Increasing AI Sustainability, the number of global data center racks is expected to double or even triple between 2023 and 2028, with the net increase in electricity demand reaching approximately 650 terawatt-hours per year—equivalent to France's annual electricity consumption. For any city with a hyperscale data center, this means heavy pressure on the local power grid. The case of Northern Virginia (the world's largest data center hub) shows that the local electricity consumption structure is being reshaped by AI workloads, with most power still coming from fossil fuels.

Embedded Carbon: The Overlooked Urban Cost

The study points out that more than half of the carbon emissions from large AI data centers come from "embodied emissions"—the carbon footprint generated by manufacturing chips, servers, and building the data centers themselves. This means cities cannot focus solely on operational efficiency when choosing AI hardware. Using recycled components or previous-generation chips can reduce total emissions by 10% to 20%. For local governments pursuing "near-zero carbon" smart city goals, this is a practical strategy: prioritize refurbished or low-energy inference hardware over chasing the latest training chips.

Water-Carbon Trade-off: A Dilemma for City Operators

AI is a "big water guzzler." The cooling systems used during model training and inference consume large amounts of fresh water. The study reveals an unsettling trade-off for city managers: measures to reduce water consumption (such as replacing evaporative cooling with air cooling) can lead to a more than 10% increase in carbon emissions. In water-scarce cities on the Arabian Peninsula or in the southwestern U.S., water conservation often takes priority over carbon reduction; but in Nordic cities, the opposite is true. Therefore, each city must develop differentiated AI infrastructure operation strategies based on its local water resources and grid carbon intensity.

Temporal and Spatial Scheduling: Let AI Dance with Renewable Energy

Unlike traditional data centers that run under constant high loads, AI training tasks are time-flexible.The Scheduling of Time and Space: Letting AI Dance with Renewable Energy

Unlike traditional data centers that always operate under high load, AI training tasks have temporal flexibility. Research shows that scheduling training tasks to periods of renewable energy surplus (e.g., windy nights) can reduce the carbon intensity of model training by about 10%. For cities, this means they can collaborate with power system operators to establish a "carbon-aware scheduling" mechanism: AI workloads dynamically migrate or pause based on real-time grid carbon signals. This model is being piloted in California and Ireland, but large-scale deployment requires support from edge computing nodes and 5G networks.

Fragmentation of Inference: The Potential of Urban-Scale Distribution

While AI training is concentrated in large data centers, inference workloads are moving to the edge. Studies indicate that inference contributes 40% to 60% of the equivalent carbon emissions over a model's lifecycle, yet inference can run on lower-performance hardware. Cities can distribute inference tasks to smart light poles, traffic signal controllers, or building energy management controllers on the streetside—essentially replacing centralized data centers with edge devices. The "AI at the edge" project in Amsterdam, Netherlands, is validating this approach: processing traffic video analysis through thousands of low-power terminals, reducing reliance on centralized cloud computing.

A Greater Challenge: System-Level Trade-offs and Governance

It must be acknowledged that AI's environmental impact is not an isolated technical issue. It involves urban energy systems, water supply networks, land planning, and digital sovereignty. For instance, a city that declares "carbon neutrality by 2030" while simultaneously encouraging the deployment of hyperscale AI data centers must incorporate carbon budgets and water budgets into the planning stage. Singapore's "Green Data Center Roadmap" requires new data centers to have a PUE (Power Usage Effectiveness) below 1.2, and mandates indirect evaporative cooling technology to reduce water consumption. Similar policy tools are becoming the new standard for competition in digital infrastructure among cities.

Conclusion: Cities Need a New "Infrastructure Balance Sheet"

AI is not going away, and smart cities are irreversible. But cities must realize that behind every line of code lies a real physical cost. The successful smart cities of the future will not be those that deploy the most AI sensors, but those that can accurately measure and sustainably manage AI's carbon, water, and grid impacts. As the review emphasizes: technological efficiency is never sufficient to offset scale growth; cities must redesign the relationship between AI workloads and local resources from a systemic level.

This is not only a technical challenge but also a test of governance artistry.