Observability Summit NA 2026: What the Community Is Thinking About

Two days in Minneapolis with the OpenTelemetry community, talking about where telemetry pipelines are headed and what the AI wave is doing to them. Two topics dominated everything: AI and cost reduction. Not as separate conversations, either. The more the community talked about AI telemetry, the more the cost question followed right behind it.

I joined Diana Todea from VictoriaMetrics and Antonio Jimenez Martinez from Cisco ThousandEyes on the Telemetry That Matters panel. Here's what I took away from the full two days.

AI Showed Up in Five Different Ways, and None of Them Were Hype

AI dominated the agenda, and the conversations were more concrete than the usual hype.

GenAI Semantic Conventions Are Settling, but "Compatible" Means Different Things

Zach Groves from Datadog gave one of the more practically useful sessions for teams shipping LLM telemetry today. The OTel GenAI semantic conventions are approaching stability, but compatibility claims across the ecosystem are uneven. Some implementations emit proprietary attributes while advertising spec compliance. Zach walked through which projects track closely and which don't. If you're building on top of GenAI telemetry, that distinction matters for portability.

Tracing Agent Decisions, Not Just Actions

Dhiraj Kumar Jain and Vikash Agrawal from AWS extended the GenAI conversation in a direction I found more interesting: how do you capture why an agent made a decision, not just what it did? Their answer used standard OTel primitives. Confidence scores, options considered and rejected, chain-of-thought captured as span events. No new instrumentation surface. The framework already supports it.

Agents for Incident Investigation

Divya Mahajan from Amazon and Achin Gupta from Intuit presented the clearest end-to-end version of an agent-led RCA workflow I've seen: topology investigation to scope blast radius, parallel fan-out to collect signals, MCP-fed runbooks and past incidents for context, then hypothesis generation and ranking.

They also raised something worth sitting with: humans investigating incidents start from a hypothesis and look for confirmation. Agents approach the same problem without that anchoring bias. That's not a small difference in high-stakes investigations.

AI-Generated Code Needs Better Runtime Visibility

May Walter from Hud came in from a different angle. AI-generated code, in their experience, is producing more critical and major bugs than human-written code, and standard OTel instrumentation isn't catching them. Hud's answer is a runtime intelligence layer on top of telemetry: detect issues in production, propose fixes, open PRs for review. The observation that our instrumentation models weren't designed for AI-generated code is worth taking seriously.

Intelligence at the Edge: How Much Autonomy Is Right?

Alex Degitz from ElastiFlow talked about pushing ML models onto the collector itself: keep, drop, and route decisions made locally, with dynamic sampling driven by anomaly scores. It works. The open question is how much autonomy to hand these systems. That tension between automation and operator control came up across multiple sessions.

Bindplane is investing in this space with Pipeline Intelligence. Our position is a human-in-the-loop model: the system proposes changes, operators approve them. We think that's the right balance for production environments where a misconfigured drop rule has real consequences.

Cost Reduction Is Still Central

Cost reduction ran through the whole agenda, not just the sessions explicitly about it. As LLM and agent telemetry volumes grow, the question of which data is actually worth keeping gets harder, not easier.

Telemetry That Matters: Our Panel's Take

Our panel hit this from three directions. Diana made the case that roughly half of collected metrics today are never queried. Her team at VictoriaMetrics is working on retroactive sampling they plan to donate to OTel upstream: sampling decisions made centrally on lightweight metadata, with full trace data pulled back only for selected samples.

Antonio walked through concrete pipeline techniques at scale: semantic conventions for signal consistency, log deduplication, and the new cardinality limiter processor. My contribution centered on the pipeline as the control plane. That's where you enforce attribute consistency for correlation, where you optimize without code changes, and where you manage fleets at scale through OpAMP.

The Hidden CPU Tax of Format Conversion

Cijo Thomas and Joshua MacDonald from Microsoft gave the most novel talk of the two days. Their argument: data format conversions inside the pipeline (SDK to OTLP, OTLP to collector-internal, collector to backend) consume significant CPU without producing any new telemetry. They're calling this the invisible tax.

Their proposed answer is OTAP, an Apache Arrow-based columnar protocol with stateful streams. In early testing: roughly 40% smaller payloads and meaningfully lower transform overhead. Apache Arrow language support is uneven right now, and this isn't production-ready. File it under "watch closely," not "deploy next quarter."

Build vs. Buy at Petabyte Scale

Diego Rocha from AWS and Otavio Valadares from Nubank tackled cost from a different layer entirely. Nubank replaced their observability vendor with an in-house log platform built on Fluent Bit, S3, Parquet, and Trino. One petabyte of logs per day. Costs cut in half. It's an impressive build. It's also a team-years engineering investment that puts it out of reach for most organizations. Worth knowing it's possible. Worth being honest about what it takes.

AI Expands the Surface. The Pipeline Is Where It Has to Be Managed.

AI is expanding the telemetry surface. Agents emit more data, from more hops, with less predictable structure. The community is still working out the right instrumentation primitives for that world. At the same time, costs are already high and getting higher. The pipeline is where those two pressures meet.

If you want to dig into any of these talks, the Sched site has the full schedule with slides linked on most session pages.