NANOG 96: Re-Engineering the Network for the AI Era

The fog wasn’t the only thing rolling into San Francisco this February. From February 2–4, 2026, the Hyatt Regency Embarcadero became the center of the networking community as NANOG 96 touched down.

While NANOG has always been the place where the “people who keep the internet running” gather to swap horror stories about vendor bugs and maintenance windows, this event felt different. The hallway track was buzzing with a singular, high-voltage energy: Artificial Intelligence. It’s no longer just a buzzword; AI has officially become a reality that is reshaping everything from data center placement and cooling to the way we diagnose outages.

The AI Shift: From Data Centers to AI Centers

One of the most provocative themes of the week was the transition from traditional data center spine and leaf architecture to what Arista’s Tyler Conrad dubbed “AI Centers.” The consensus? A normal DC network simply isn’t built for the workload of a 60,000-GPU cluster. Keynote sessions from AHEAD, NVIDIA, and Meta laid out the brutal reality of modern infrastructure:

• Thermal Demands: We are moving toward rack densities that require extreme-density cooling solutions—far beyond what standard air-cooled facilities can handle.
• Traffic Patterns: Unlike the predictable North-South traffic of the web-scale era, AI workloads create massive East-West bursts and elephant flows. If your congestion control isn’t near perfect, the traffic will crush your performance before you can even see the telemetry spike.

Network Automation isn’t just about configuring network devices!

The most practical takeaway for the operators in the room was the focus on Autonomous Troubleshooting Agents. The theme wasn’t just about building networks for AI, but using AI to run the networks.

I participated in a deep-dive panel from Bill Lapcevic from NetBox Labs and Chris Wade from Itential that was moderated by Ethan Banks from Packet Pushers. We enjoyed the conversation so much we did a longer form video recording in the afternoon. The video from NANOG is here and the longer form video is here.

In these sessions we challenged the old-school view of automation. We argued that automation shouldn’t just be about pushing a config via a script; you have to define the service you are looking to deliver. Once that is defined, you have a bigger picture of the goal of your automation initiative.

The session broke down automation into three critical areas that often get neglected:

• Source of Truth (SoT) as the Foundation: As Bill highlighted, you cannot automate a mess. Tools like NetBox aren’t just “digital spreadsheets”; they are the authoritative engines that tell your automation what the network should look like. If your SoT is out of sync, your automation is just a faster way to break things.
• Observability Loops: I emphasized that automation is a two-way street. For every configuration change pushed, there must be a corresponding verification step. Does the telemetry show the traffic shifted as expected? If not, the automation should be smart enough to trigger an automatic rollback.
• The “Human” Interface: Chris and Ethan touched on the reality that automation must integrate with business tools. Whether it’s ServiceNow tickets or Slack notifications, the network doesn’t live in a silo. Automation should bridge the gap between a business request and a technical execution on the network devices.

By shifting the focus from “writing scripts” to “building systems,” the panel challenged the NANOG community to stop thinking like hobbyist coders and start thinking like Systems Architects. We also argued it should be about Agentic AI—LLM-driven agents that can:

1. Ingest Telemetry: Understand real-time streaming data rather than just setting static thresholds on alerts.
2. Form Hypotheses: During a 3:00 AM incident, an AI agent can correlate a spike in optical interface errors with a BGP flap much better than a tired human being.
3. Execute Guardrailed Actions: Not just alerting a human, but suggesting (or some day performing) a drain of a specific circuit to mitigate traffic loss.

The “Grounded” Truths: BGP and Routing Security

Despite the AI fever, NANOG stayed true to its roots. Geoff Huston delivered his essential “BGP in 2025” report, providing a data-driven reality check. While many feared the global routing table would explode, the growth has actually remained relatively linear.

RPKI adoption was another major win. We saw a significant shift from “talking about” routing security to “enforcing” it, with more operators reporting they are now actively dropping invalid routes. It’s a reminder that while we’re building for the future of AI, we’re still working hard to secure the foundation of the 40-year-old protocols that hold the web together.

Final Thoughts: The Infrastructure Boom

NANOG 96 made one thing clear: we are in a new “land rush” for infrastructure. Between the massive IPv4 shifts (including the unsettling exodus of address space from Ukraine) and the billion-dollar investments in AI clusters, the network is once again the most exciting place to be in tech. If you were not able to make it out to NANOG 96, be sure to check out the recordings in the agenda. Also, be sure to register for NANOG 97 in Bellevue, WA June 1-3!