GPU FinOps Audit: scope, deliverables, and the spreadsheet

This is the audit template we use for GPU FinOps engagements. It runs two weeks, has fixed scope and deliverables.

Week 1 — Discovery and instrumentation

Read-only access required

• Cloud cost data (AWS Cost Explorer, Azure Cost Management, or Harness CCM / Kubecost export)
• Kubernetes API read access (cluster-wide view is enough)
• Prometheus / Grafana read access
• Read access to GitOps repo (Argo / Flux) for deployment manifests

What we instrument

If not already in place:

• DCGM Exporter → Prometheus for per-pod GPU utilization
• kube-state-metrics for label correlation
• A shared Grafana dashboard for the audit period

We bring this stack up via Helm in under a day. Removed cleanly at end-of-engagement if not adopted.

Week 2 — Analysis and remediation plan

Spend categorization

Every dollar of spend lands in one of these buckets:

Category	Definition
Productive	Workload running, utilization above threshold, attributable to a product / cost center
Allocated idle	Workload running, utilization below threshold (often headroom or over-provisioning)
Unallocated idle	Resources running, no workload (stale node groups, churn cycles)
Unattributed	Resources missing labels — unknown owner
Tax	Cluster overhead — control plane, observability, system pods

What we look for

• Stale node groups — utilization < 1% for the audit window
• GPU node churn cycles — count of GPU node creations / deletions per day
• CPU pools at < 5% utilization — over-provisioning
• Pods missing cost-center labels — unattributable
• Inference pools sized for peak, not p95 — over-provisioned by definition
• GPU workloads that don’t saturate a full GPU — candidates for time slicing
• Long-running pods with low utilization — candidates for scale-to-zero

Deliverables

At end of week 2, you receive:

1. A line-item findings spreadsheet — every issue with cost impact, severity, and remediation effort
2. A prioritized remediation plan — quick wins (≤1 week), medium (≤1 month), longer projects
3. A label and policy proposal — what to enforce via OPA / Kyverno to keep the savings
4. A Grafana dashboard pack for ongoing monitoring
5. One executive-summary slide for the cost story

Optional follow-on

• 4–6 week remediation engagement to implement the prioritized plan
• Monthly cost review with the platform team

Common findings (across engagements)

The Pareto we see most often:

• 70% of unallocated spend is two or three large issues, not a hundred small ones
• The biggest hidden cost is GPU node churn, not over-provisioning
• The biggest preventable cost is missing labels — without attribution, no team owns the waste
• Quick wins recover 50–60% of waste in the first month; the rest requires structural change (workload migration, app re-architecture)

What this is not

• A cost-cutting hatchet job. We won’t recommend changes that risk SLOs.
• A replacement for ongoing FinOps practice. The audit is the diagnosis. The team owns the cure.

---

Interested? Email us with a short summary of your environment and we’ll come back with a fixed quote.