Kaj Kjellgren – Beyond the GPU Cluster: Building Networks for Distributed AI Inference
Kaj Kjellgren is Senior Peering Consultant at DE-CIX, focusing on global peering and interconnection, particularly in the Nordic market. He has more than 25 years of experience in internet infrastructure, network architecture, and interconnection. Before joining DE-CIX, Kaj spent more than two decades at Netnod, most recently as Technical Ambassador, and previously served as Technical Director for Sweden and Denmark at KPNQwest.
Kaj is a frequent speaker, moderator, and panellist within the international peering community, with a focus on how interconnection, network architecture, and resilience must evolve to support new technologies and changing traffic patterns.
At BalticNOG 2026, Kaj will bring this perspective to the stage with his talk, “Beyond the GPU Cluster: Building Networks for Distributed AI Inference.”
Past the Data Centre Walls
Most AI networking discussions stop at the GPU cluster. Kaj’s talk moves past that boundary and asks what happens when a single inference task starts crossing devices, agents, models, tools, data sources, neoclouds, and multiple network domains, turning one user interaction into a dynamic distributed execution graph rather than a simple request to one model endpoint.
The Network as an Input to Execution
Different parts of that execution graph may run at different distances, with model, provider, and location choices shifting mid-task. Kaj will examine how latency, reachability, compute queues, policy, privacy, and existing state can all factor into where a task actually runs, effectively making the network part of the execution decision rather than passive transport underneath it.
Why Attribution Gets Hard Across Boundaries
Using an illustrative five-second AI task, Kaj will show why attributing delay or failure becomes difficult when every participant only sees one part of the transaction. Existing tools, including OpenTelemetry, W3C Trace Context, inference-server metrics, gNMI/OpenConfig, IPFIX, BMP, STAMP/TWAMP, and IOAM, already provide many of the needed signals. The harder problem is correlating that evidence across organisational and trust boundaries.
Introducing FOX: A Federated Observability Exchange
Kaj will introduce FOX, a deliberately exploratory concept for policy-controlled exchange of cross-domain observability evidence. He’s clear that FOX isn’t a finished architecture, standard, or product, it’s meant to prompt discussion about whether neutral interconnection could help solve cross-domain observability at all.
The talk closes by connecting observability to security and resilience: agentic amplification, inference exhaustion, hidden shared dependencies, stateful failover, and the need for local fallback.
Why You Shouldn’t Miss This Talk
This talk isn’t trying to predict one future for AI networking, it’s giving the networking community a practical vocabulary for the architectures already starting to emerge. Kaj brings decades of interconnection experience to a problem that’s only going to get more relevant as AI workloads spread across more networks and providers.
Whether you operate a network, work in peering, or are just trying to understand where AI traffic is headed next, this talk gives you a framework for thinking about it.
đź“… When: 23-24 September 2026
đź“Ť Where: BalticNOG 2026, Riga, Latvia
đź”— Don’t miss BalticNOG! Register today: https://balticnog.org/tickets/
