Generative AI workloads are breaking every aspect of the data center. As the capabilities of AI increase, so does its demand. The conventional path of performance improvement in legacy processors has stagnated, providing diminishing returns. This raises concerns on whether we may ever fully realize the potential of AI. In this talk we will share how combining our software-first methodology and novel computer arithmetic leads to breakthrough performance gains in general purpose accelerated computing, pushing developers to the limit of physics.