About This Talk
Modern AI Workloads are exposing gaps in the infrastructure built to run them.
Cloud infrastructure was designed to treat compute as an interchangeable resource. That model works well for many applications, but it becomes less reliable when the workload depends on the placement of thousands of accelerators, the shape of the network connecting them, and fast access to enormous volumes of data.
At Arcee AI, Gerred Dillon works across the infrastructure used to train and serve open-weight models. In this closing keynote, he’ll share what that experience has revealed about operating AI compute at scale. He’ll look at how different workloads place different demands on the same fleet, why hardware specifications reveal only part of a cluster’s capacity, and where conventional measures of efficiency begin to lose meaning.
AI infrastructure is becoming more specialized because the work itself is becoming harder to abstract. The next generation of systems will need a clear-eyed view of this imminent future.












