AI infrastructure is increasingly fragmented across NVIDIA and AMD GPUs, private bare metal, public clouds, neoclouds, HPC and sovereign environments. Yet developers expect infrastructure to behave like software: request capacity, deploy a model, scale it and tear it down without caring where it runs.

This talk explores the architecture behind Soveri, a platform to create a globally distributed, sovereign AI compute network fabric using Kubernetes. We are working toward substrate neutrality and heterogeneous accelerators, while using jurisdiction, hardware capability, latency, capacity, availability and cost as inputs for intelligent workload placement.

I will also explore the developer experience on top: isolated AI sandboxes, inference environments and model-serving endpoints that abstract away clouds, clusters and accelerators. The bigger question: can we move from asking for a specific GPU to simply asking for AI capacity, with Kubernetes as the global abstraction?