MITIGATING THE AI WATER CRISIS THROUGH ECO-ADAPTIVE SPATIO-TEMPORAL WORKLOAD SCHEDULING

The quick rise of Generative AI is sparking a rapid rise in hyperscale data centers. The extreme carbon impact associated with these facilities is being studied in great detail, however, very little attention has been given to the extreme amount of freshwater used in evaporative cooling.

Current hardware-focused solutions like mechanical liquid cooling can be very expensive and shift a large portion of the carbon footprint toward also generating a large amount of electricity.

The Eco-Adaptive Water Brain Framework being discussed in this paper is an innovative software based architecture that will drastically reduce water use by designing solutions at the OS level and at the routing level. This framework will utilize spatio-temporal workload scheduling to continuously evaluate the local server's water stressed condition. The framework will route active urgent queries to energy efficient models Low Water Mode while heavy non urgent training workloads will be queued to run later at night when evaporative cooling requirements are naturally reduced with cooler ambient temperatures.The multi layered algorithmic process used in this new framework provides a hardware agnostic and sustainable method that will greatly reduce the water footprint of generative AI