Service-preserving carbon- and water-aware resource allocation for geo-distributed AI inference
This paper proposes a trace-driven, two-stage lexicographic framework for geo-distributed AI inference that prioritizes service feasibility while optimizing carbon emissions under strict physical water constraints, demonstrating significant environmental trade-offs and providing a reproducible benchmark for sustainable resource allocation.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
When artificial intelligence systems answer a question or generate an image, they do not exist in a vacuum. They run on massive networks of computers scattered across different cities and countries, a setup known as geo-distributed computing. While these systems are powerful, they consume vast amounts of electricity and water to cool the servers that keep them running. The electricity comes from power grids that vary in how clean they are; some regions rely heavily on coal, while others use wind or solar. Similarly, the water used to cool these machines is drawn from local rivers and reservoirs, which face different levels of stress depending on the season and the local climate. The challenge for engineers is to decide where to send each request. They must balance the need for speed, the availability of computer power, and the desire to minimize harm to the environment. If a system simply sends work to the cleanest power grid, it might accidentally drain a local river in a drought-stricken area. If it sends work to save water, it might burn more carbon. Finding the right balance without slowing down the service is a complex puzzle that researchers are only beginning to solve.
A team of researchers at North China Electric Power University has developed a new way to solve this puzzle, focusing specifically on the massive AI systems that generate images and text. They created a two-step planning method that ensures the AI service never slows down or fails, even while the system tries to be more environmentally friendly. Their approach treats the service itself as the most important thing to protect. First, the system calculates the absolute maximum amount of work it can complete given the current limits of computer capacity, network speed, and travel time for data. Once that maximum is locked in, the system then looks for the best way to distribute that exact same amount of work to reduce carbon emissions and water usage. This method prevents a common mistake where a system might claim to be "greener" by simply doing less work or skipping difficult tasks. Instead, it finds the greenest path for the full workload.
The researchers tested their method using a real-world record of over 26,000 successful AI requests that took place over 554 hours. They simulated a network with four different regions, each with its own mix of power sources, water availability, and computer limits. In their most detailed test, they created a map of 21 different scenarios, ranging from allowing all water usage to completely banning water use in the most stressed areas. They found that by strictly avoiding water use in regions where water is already scarce, the system could save 2.4250 liters of direct water compared to a standard approach that only cares about carbon. This saving came with a very small cost: the system produced an additional 0.0582 kilograms of carbon dioxide equivalent. However, even with this small increase, the total carbon footprint remained significantly lower than if the system had simply refused to move work between regions to save time.
The study revealed that the cost of being greener is not the same at every level. When the researchers relaxed the water rules slightly, the system could avoid a lot of water with almost no extra carbon cost. But as they pushed the rules to be stricter, trying to avoid the very last drops of water in stressed areas, the carbon cost began to rise sharply. This creates a clear trade-off curve, or frontier, that managers can use to make decisions. They can see exactly how much extra carbon they would have to accept to save a specific amount of water. The researchers also tested thousands of different combinations of power grid conditions, computer speeds, and network delays to ensure their findings held up under pressure. They found that while the exact numbers changed depending on the conditions, the overall direction remained the same: it is possible to protect the service, save water in stressed areas, and still keep carbon emissions low.
One of the most important findings was that the location of the water matters more than the total amount. Because the water used to cool servers is calculated based on how much electricity is used, the total volume of water consumed by the system did not change just because they moved the work. What changed was where that water was taken from. By shifting work away from regions with high water stress to regions with plenty of water, the system avoided putting pressure on vulnerable local ecosystems. The researchers confirmed that this result was not a fluke of their specific data; they ran thousands of simulations with different time patterns and water efficiency rates, and the core conclusion held steady. The method provides a transparent, auditable way to manage these massive AI systems, allowing operators to set clear limits on water use and see the exact environmental price of those limits before they are applied. This gives decision-makers a tool to balance the demands of a growing digital world with the physical limits of the planet, ensuring that the AI of the future can run efficiently without draining the resources it depends on.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.