AI Observability
Track the performance, reliability, and cost of your AI and LLM workloads.
The AI Observability page keeps track of how your AI and LLM workloads are actually behaving: latency, reliability, and cost. That covers both API-based models (hosted LLMs being the obvious example) and GPU inference you run yourself.
What you'll see
- Requests and tokens processed over the selected range
- Cost and cost over time for your model usage
- p95 latency and error rate
- GPU utilization and memory for self-hosted inference
- A by-model breakdown and a recent-calls feed
Getting data in
There are two ways to send AI and LLM events to KubeWatch:
- OpenTelemetry (OTLP): instrument your app and export spans or metrics for model calls.
- AI events API: report each model call (model, tokens, cost, latency, status) directly from your application.
GPU utilization and memory are collected automatically by the agent on hosts with GPUs.
AI Observability is available on the Enterprise plan.