Pathways to Sustainability: Carbon-Aware Routing for Global AI Data Transfers

AI-driven applications require massive amounts of training data that are not necessarily located close to the computational infrastructure that processes them. The energy consumption of transmitting data, both for training and for routine applications of deployed AI methods, can be substantial, but is often overlooked. However, carbon emissions from data transfers could be reduced through carbon-aware routing that selects lower-emission paths through the network. These paths are selected based on the carbon intensity and the time of day in the countries whose network infrastructure is used along the path, reflecting the country-specific share of green energy. Here, we present a carbon-aware routing based on a weighted graph representation of the global internet infrastructure where time-dependent edge weights capture both the energy consumption and carbon emissions associated with data transmission across submarine cables and terrestrial links. We performed an empirical evaluation of the savings that could be achieved if such a routing was implemented in practice, showing that the carbon emissions of data transfer could be reduced by on average 40.07% if the "greenest" path was chosen over the baseline. Our work raises awareness for the fact that cloud computing causes substantial carbon emissions through the data transfer alone, and that intelligent routing could serve to reduce the carbon footprint of AI in the future.