Idle capacity nobody releases
Instances provisioned for a training run that finished weeks ago, still allocated, still billing, with no process that notices.
AI Cost Intelligence
GPUs are your most expensive compute.
And your least accountable.
The GPU Optimization Assessment inventories every GPU instance across training and inference workloads, measures real utilization, links each one to its team and project with full IaC lineage, and delivers a right-sizing roadmap with estimated annual savings. Reclaim decisions become defensible because the evidence travels with them.
Read-only, approval-gated scan
Findings in days, not months
No credentials stored centrally
No tagging required to start
The Problem
GPU capacity gets provisioned for peak and forgotten at idle.
Training runs end, experiments conclude, and teams move on. The instances stay. At GPU prices, the gap between what was provisioned and what is actually used is often the single largest line of recoverable spend in the cloud budget.
Over-provisioned for peak
Inference fleets sized for launch-day traffic that never returned, running at a fraction of capacity around the clock.
No utilization visibility
Billing shows the instance count. It does not show that average utilization is low, or which specific instances are doing real work.
Unclear ownership
The engineer who requested the capacity has changed teams. Without lineage from instance back to code and requester, nobody will sign off on reclaiming it.
Reclaim decisions stall
Finance flags the spend, engineering fears breaking a workload nobody understands, and the safe answer is always to leave it running.
AI growth compounds the waste
Every new model, experiment, and inference endpoint adds capacity. Without a baseline, waste grows at exactly the rate your AI program does.
The Visibility Gap
Five questions this assessment answers with evidence.
If these cannot be answered today, right-sizing conversations stay stuck at opinion versus opinion.
1
How many GPU instances are running right now, across every account and both training and inference?
2
What is each one's real utilization over the last 90 days?
3
Which team, project, and environment does each instance belong to?
4
Which instances are safe to reclaim, resize, or schedule, and what breaks if we do?
5
What would right-sizing be worth per year?
How It Works
Two scoped phases. Deterministic results. Human approval at every gate.
The assessment runs as two scoped, deterministic, approval-gated phases: inventory discovery through approved connectors, then deterministic policy checks against that inventory. Oscar never performs open-ended network scanning, and no action executes without explicit human sign-off.
- 1
Connect with read-only credentials
Oscar connects through approved connectors using read-only access and the permissions your engineers already hold. Setup takes under 30 minutes. No credentials are stored centrally and no changes are made to your environment.
- 2
Oscar inventories the GPU fleet
Every GPU instance across accounts and providers is mapped in the Cloud Intelligence Graph with its utilization history, workload type, owner, project, and environment, and its lineage back to the IaC that created it.
- 3
Deterministic checks classify the fleet
Idle, underutilized, right-sized, and over-provisioned classifications are applied deterministically from 90-day utilization data. Same inputs, same findings, every time. Each classification carries its evidence.
- 4
You receive the right-sizing roadmap
A prioritized roadmap with estimated annual savings, where every reclaim or resize action is an Oscar-ready task carrying its lineage, owner, and rollback classification, so approval conversations take minutes instead of quarters.
What You Get
Four deliverables that make reclaim decisions defensible.
The difference between a savings estimate and a savings outcome is evidence. Every recommendation carries the lineage that lets an engineer approve it with confidence.
Full GPU Fleet Inventory
Every GPU instance across training and inference: utilization, owner, project, environment, and IaC lineage. The complete picture in one place for the first time.
Cost Attribution & Idle Classification
Spend attributed by team and project with a 90-day utilization trend per instance and a deterministic idle, underutilized, or over-provisioned classification.
Right-Sizing Roadmap
Prioritized reclaim, resize, and scheduling actions with estimated annual savings, rollback safety classification, and the owner who needs to approve each one.
Ongoing Fleet Monitoring
New GPU capacity detected, alert fires. Utilization drops below threshold, alert fires. The baseline stays current as your AI program grows.
Why OpsCanvas
Lineage is what makes savings real.
Utilization data alone does not get instances reclaimed. Defensible ownership does.
Most cost tools can list underutilized instances. The reason that list rarely converts to savings is that nobody can say who owns each instance, what created it, or what happens downstream if it goes away. So it stays running.
The Cloud Intelligence Graph links every GPU instance to its IaC source, its requester, its project, and its dependencies. When the recommendation says reclaim, the evidence for why it is safe travels with it, and the person who must approve it is named.
✓
Full IaC lineage from instance to source code and requester✓
90-day utilization evidence behind every classification✓
Rollback safety classification on every recommended action✓
Training and inference workloads covered in one engagement✓
Every recommendation is an Oscar-ready remediation taskCapability
OpsCanvas
Discovery method
Live APIs via approved connectors
Tagging required
No
Utilization window
90-day trend per instance
Attribution granularity
Team, project, environment, requester
Savings estimate
Annualized, per action
Continuous monitoring
Available as add-on
Common Questions
Frequently asked.
Does this cover both training and inference workloads?
Yes. Training and inference have different waste patterns: training waste is usually capacity that outlived its run, while inference waste is usually fleets sized for peak that never arrived. The assessment classifies and reports them separately because the remediation differs.
Does Oscar scan our network to find GPU workloads?
No. Oscar performs no open-ended or autonomous network scanning. The assessment runs as two scoped, deterministic, approval-gated phases: inventory discovery through approved connectors using existing permissions, then deterministic policy checks against that inventory.
How are the savings estimates calculated?
Per action, from observed utilization against current pricing for the instance type and commitment model. Estimates are conservative by design and each carries its evidence basis, so finance can trace any number in the roadmap back to the instances behind it.
Will the assessment make changes to our environment?
No. The assessment is read-only. Remediation is a separate, optional step where Oscar works through the roadmap in co-pilot mode: Oscar proposes each action, an engineer approves it, and every change is logged with a full audit trail.
Can this be combined with the AI Token Optimization Assessment?
Yes. GPU capacity and token spend are the two halves of AI infrastructure cost, and both assessments run on the same Context Graph. Bundling them gives a complete AI cost picture with one graph build.
Get Started
Find out what your GPU fleet is actually doing.
Read-only connection through approved connectors. Full fleet inventory with utilization, ownership, and lineage. A right-sizing roadmap with estimated annual savings, in days.
Read-only access | No changes to your environment | Setup under 30 minutes