Green AI Workshop

Title

Deaths Per Token: A Health-Impact Unit for Generative AI Infrastructure

Abstract

The AI sector is locking in multi-trillion-dollar, multi-decade infrastructure with no quantitative answer to a basic question: how many statistical deaths does a token of generative output cost? PUE, WUE, gCO2/API call, and Wh/token are useful proxies, but none speak the unit that users, providers, and regulators actually transact in, nor do they connect engineering choices to population-health outcomes the way grams-per-kilometer does for cars.

This talk argues that the right unit is **deaths per token (DPT)**, the micromort of machine generation, and sketches what changes when we adopt it. We introduce a pathway-additive framework covering datacenter electricity, water, construction, climate, and client devices; show how it lets us "mark to market" the gap between a 100% renewable claim and its actual hourly, local displacement; and apply it to today's frontier model fleet.

What comes out is a wide spread across models and a few uncomfortable surprises, most of them not where the field has been looking, pointing to a handful of levers that matter far more than the rest.

Bio

Laurent Bindschaedler leads the Data Systems Group at the Max Planck Institute for Software Systems, where his work sits at the intersection of operating systems, databases, and machine learning. His group builds next-generation systems for massive-scale data management and AI, with current threads spanning agentic AI infrastructure, LLM-enhanced databases, learned index structures, code intelligence, and benchmarking methodologies for adaptive systems. He holds a PhD from EPFL, was a postdoctoral fellow at MIT CSAIL, lectures on operating and distributed systems at the University of Saarland, and has co-founded technology ventures along the way.