Most discussions about AI infrastructure focus on what happens inside the GPU cluster. This talk moves the discussion beyond the data centre and asks what happens when inference becomes distributed across devices, agents, models, tools, data sources, neoclouds and multiple network domains.

The central proposition is that an AI task may increasingly resemble a dynamic distributed execution graph rather than a simple request to a single model endpoint. Different parts of that graph may execute at different distances, while model, provider and location choices change during the task. Latency, reachability, compute queues, policy, privacy and existing state may therefore become part of the placement decision—making the network an input to execution rather than passive transport.

Using an illustrative five-second AI task, the talk then examines why attributing delay or failure becomes difficult when every participant sees only one part of the transaction. Existing technologies—including OpenTelemetry, W3C Trace Context, inference-server metrics, gNMI/OpenConfig, IPFIX, BMP, STAMP/TWAMP and IOAM—already provide many of the required signals. The unresolved problem is correlating selected evidence across organisational and trust boundaries.

The talk introduces FOX, a Federated Observability Exchange, as a deliberately exploratory concept for policy-controlled exchange of such evidence. FOX is not presented as a finished architecture, standard or product, but as a way to stimulate discussion about whether neutral interconnection could help provide cross-domain observability.

Finally, the talk connects observability to security and resilience: agentic amplification, inference exhaustion, hidden shared dependencies, stateful failover and the need for local fallback. The objective is not to predict one future, but to give the networking community a practical vocabulary for discussing the architectures now beginning to emerge.