Governing AI Data-Centre Resources Before Consumption: The ZERO Framework
This paper introduces the ZERO framework, an operational governance model that shifts AI data-center sustainability decision-making upstream by enforcing multi-resource admissibility constraints before workload execution, thereby prioritizing holistic resource availability over post-hoc efficiency metrics.
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
Artificial intelligence is often talked about as a digital phenomenon, a race of algorithms and data moving through invisible clouds. But the reality is far more physical. Every time a model learns or answers a question, it requires a massive amount of electricity to power processors, water to cool the machines, and raw materials to build the infrastructure that holds it all together. As these systems grow, they are becoming a significant strain on the planet's resources. The current way we manage this strain usually happens after the fact. We build a data center, we run the computers, and then we measure how much energy or water was used. We might try to make the process more efficient, but the decision to start the work in the first place is rarely questioned. The question remains: just because we can run a computation, does it make sense to do so right now, in this specific place, given the local water levels and power grid conditions?
A new perspective from independent researcher Maria Kollia proposes a different way to handle this, called the ZERO framework. The name is not an acronym, nor does it promise that machines will use absolutely nothing. Instead, it represents a shift in timing. The framework argues that we must make a hard decision about whether a task is allowed to run before any resources are consumed. It acts as a gatekeeper that checks the local environment and the necessity of the work before the computer is even turned on. If a data center is in a region where the water supply is critically low, or if the local power grid is unstable, the framework suggests that the work should be stopped or moved, even if that center is otherwise very efficient. This approach moves the focus from simply measuring how well a machine runs to deciding whether it should run at all.
The core idea is built on a simple hierarchy of rules. First, the system asks if the computation is truly necessary. If it is, it then checks if the work can be done with less power or in a smarter way. Only after these questions are answered does it look at where the work should happen. The framework introduces a strict set of "hard rules" that a facility must pass before it is even considered as an option. These rules might include local laws, safety requirements, or environmental limits, such as a minimum amount of water available for cooling. If a facility fails even one of these hard rules, it is immediately excluded from consideration. No amount of efficiency or low carbon emissions can make up for a failure to meet these basic safety or environmental standards. This is a crucial distinction from current methods, which often try to balance trade-offs, such as accepting high water use in exchange for low carbon emissions. ZERO says that if the water is critical, the trade-off is not allowed.
To test how this logic would work in practice, the author created a detailed simulation involving three different data centers. One was a local facility, another had very clean electricity but was located in an area with a severe water shortage, and the third had moderate electricity but plenty of water and a way to reuse the heat it generated. In a standard scenario, a computer program looking only for the lowest carbon emissions would choose the second facility, the one with the clean power. However, under the ZERO framework, that facility is immediately blocked because of its critical water state. The system then looks at the remaining options and chooses the third facility. While this choice uses slightly more electricity to move the work there, it avoids the environmental damage of draining a water-scarce region. The simulation showed that by following these rules, the selected option used significantly less water and still produced a much lower carbon footprint than simply running the work at the local default facility.
The framework also accounts for the fact that conditions change. A data center that is safe to use this morning might not be safe in an hour if the water levels drop or the power grid fluctuates. The system is designed to keep checking these conditions while the work is running. If a facility crosses a safety threshold, the work is paused or moved to a different location. This continuous monitoring ensures that a decision made at the start remains valid throughout the process. It also requires that the information used to make these decisions is fresh and reliable. If the data about a facility's water or power status is too old, the system treats that facility as unavailable until new, verified information arrives. This prevents the system from making choices based on outdated maps of the world.
The author is careful to note that this is a proposal for how to govern resources, not a proven solution that has been tested in the real world yet. The results presented are based on a synthetic example designed to show how the logic works, not to prove that the method will save specific amounts of energy or water in every situation. The paper argues that the current way of managing AI infrastructure is incomplete because it waits until resources are already being used to start measuring and managing them. By moving the decision point earlier, the ZERO framework aims to prevent waste before it happens. It suggests that the most effective way to be sustainable is not just to build better machines, but to have the authority to say "no" or "not here" when the local conditions cannot support the demand.
The path forward for this idea involves a series of real-world tests. The author suggests starting with simulations using past data to see how the rules would have performed in different weather or grid conditions. The next steps would involve testing the system in a "shadow mode," where it runs alongside real data centers to make recommendations without actually controlling the machines. Eventually, controlled pilots would test the system in live environments, measuring not just the environmental benefits but also the cost and complexity of running such a governance system. Until these tests are done, the framework remains a structured way of thinking about the problem, offering a clear set of rules to ensure that the growth of artificial intelligence does not come at the expense of the local environment. It is a call to treat the physical limits of our planet as a primary constraint on digital progress, rather than a secondary detail to be managed later.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.