Behind every humming data center is a quiet supply chain drama: water. It’s not just used to cool a bank of chips; water moves through the entire computing stack—from the splashy towers of cooling towers to the ultra-pure baths that manufacture silicon. As artificial intelligence workloads surge, the thirst of the digital world becomes a strategic constraint. People tend to compare carbon footprints, but water footprints behave differently: a lot of water used in a place where it rains all year feels very different from the same amount used in a drought-prone basin. This is the tension the new framework tries to quantify.
Traditionally, researchers have measured how much water computing drinks, but they’ve paid little attention to where stress is highest or how it might shift over years or decades. The Purdue University team—Yanran Wu, Inez Hua, and Yi Ding—built SCARF, a framework that weighs water consumption by local stress and by time. In plain language, SCARF asks: if two data centers pull the same amount of water from different basins, which one leaves a bigger mark on its watershed? The answer isn’t always the obvious one, and the framework provides a way to plan with that difference in mind.
SCARF is more than a clever calculator. It’s a way to embed water risk into decisions that previously hung entirely on dollars and watts. The researchers show three bite-sized stories across the ladder of computing—language-model serving, data centers, and semiconductor fabrication—where location and timing shift the environmental bill by orders of magnitude. The work comes from Purdue University, led by Wu, Hua, and Ding, and it offers a language and a toolkit for engineers who want to design tomorrow’s systems without draining the local rivers dry.
A framework that weighs water stress
SCARF works by turning water use into a two-channel map: on-site consumption and off-site consumption. On-site water use includes things like cooling towers and humidification—the direct, visible taps in a building’s life-support system. Off-site consumption captures the water used to generate the electricity that powers the equipment—water that disappears from the local scene but reappears downstream as steam, energy, or heat. By combining these two streams, SCARF builds a raw water footprint that can be compared across sites and tasks. The unit is simple: liters of water, tied to how much power the equipment burns and how long it runs. It’s the starting line for a more nuanced story about scarcity.
Next, the framework grabs the actual weather of water: where the site sits in a watershed. It maps the facility to a hydrological basin, then draws from Aqueduct 4.0’s stress indicators—the ratio of water demand to water supply in that basin. The trick is to move from national or state-level assumptions to basin-level realities. Two facilities in the same country can live in entirely different water futures if they sit in different basins. That spatial twist matters because water stress is not a uniform climate; it’s a pool of scarcity that shifts with river flows, drought cycles, and policy decisions.
Finally, SCARF layers in time. It creates a Water Stress Factor by aggregating stress across years, but with a twist: it lets you discount the future. A higher discount rate gives more weight to today’s stress and downplays what might come tomorrow; a lower rate values future drought risk more heavily. The short-term version matters for tasks that come and go in months, like serving a burst of queries from a language model. The long-term version is for facilities that live for a decade or more, where climate projections and population growth can redraw the water map. Multiply the raw water footprint by this Water Stress Factor, and you get the Adjusted Water Impact.
Where your servers drink their water
That AWI number is where the drama unfolds. In the real world experiments, two data centers can drink the same amount of electricity but produce wildly different AWIs because their basins face different stresses. A high-desert site with parched rivers and complex supply chains can yield an AWI thousands of times higher per request than a cooler, water-rich region that happens to be nearby. The key point is not just how hungry the data center is, but how hungry the watershed is on the day you pull water. The same workload, deployed in different corners of the country, can generate very different environmental receipts.
Seasonality compounds the effect. Water stress is not a constant weather pattern; it lurks in the background and spikes in dry months. In practice, this means a deployment that looks healthy in January might be unnecessarily expensive in April if the basin is draining faster than the reservoir can fill. The paper’s side-by-side look at two locations over a year shows seasonal swings can tilt the balance, turning a seemingly safe site into a riskier bet during droughts. In other words, timing matters as much as geography when you’re building a digital empire that runs all year long.
These ideas might feel abstract at first, but they translate into real decisions. If a cloud provider could shift a workload to a lower-stress basin just for a handful of minutes during a spike, the AWI would drop; if it could schedule maintenance or hardware refreshes to avoid peak stress periods, water risk could be reduced in ways that electricity and carbon accounting alone never captured. SCARF gives managers a way to monetize that intuition into a concrete plan.
Three case studies across the stack
In the first slice of the study, language-model serving—think of a service that handles natural-language questions and conversations—became a canvas for AWI to move. The researchers showed that the same question in a single hour could produce very different water footprints depending on which data center answered it. The takeaway isn’t that one place is magically cleaner than another; it’s that the timing and location of service delivery can compound water stress in private ways, especially when a system has to scale up to meet surging demand.
Datacenters tell a subtler story. When the team looked at Google’s U.S. footprint, they found that long-term planning and short-term calculations pull in opposite directions. Some sites sit in medium-stress basins but drink a lot of power, which can push their AWI higher in aggregate. Once you switch from a strictly annual accounting to an AWI that folds in time, the ranking of sites can flip. A place that looks less sustainable if you focus on this year’s water buffet may look more favorable when future stress is given more weight, and vice versa. It’s a reminder that sustainability isn’t a snapshot; it’s a moving picture, with frames that can shift depending on how far ahead you’re looking.
And for the hardware layer—the factories that turn silicon into chips—the message is equally stark. Semiconductor fabs are water scoops, pulling vast volumes of ultra-pure water for wafer cleaning and processing. When placed in basins with high current and projected stress, their AWI climbs even if their on-site consumption isn’t enormous. The result is a blunt lesson in siting: the same plant in a water-stressed region can carry a heavier environmental burden than a larger plant in a more forgiving landscape. It’s a case study in how the location of water itself shapes the cost of making the stuff that runs the future.
SCARF isn’t just a new calculator. It’s a way to reframe how we think about the economics of computing. Water is not a background externality; it’s a resource that fluctuates with climate, policy, and local demand. If the industry wants to build systems that scale without draining local basins, it needs to couple performance, cost, and water stress in the same decision framework. The Purdue team suggests practical steps: schedule workloads by watershed stress, site facilities where water risk is lower, and design operations that can tolerate future shifts in water supply.
Beyond the practical knobs, SCARF asks a deeper question about responsibility across generations. If future water availability is uncertain, should today’s data centers sacrifice a portion of near-term efficiency to protect tomorrow’s rivers and wells? The framework’s inclusion of discounting lets leaders explore that trade-off explicitly, offering a bridge between engineering optimization and long-horizon stewardship. In the end, the authors—Wu, Hua, and Ding at Purdue—offer more than a metric. They provide a language for a more humane conversation about how the digital economy taps into the natural world, and they invite the wider industry to build with water in mind rather than as an afterthought.
