The person who scopes the work is the one who runs the instrument.
Nameplate Analytics is a specialist practice, not a staffing pyramid. There is no junior consultant assigned after the sale and no template report with your logo dropped into the header. You are buying the judgment of the engineer who does the measuring.
Kam Kheri
Seven years building and running cloud infrastructure at federal scale. Currently supporting production cloud infrastructure for a national public research program across AWS, GCP and Kubernetes: Terraform, Ansible, Puppet, and the monitoring stacks this practice depends on. Prometheus, Grafana, the ELK and Splunk side, and the DCGM exporter that produced the measurement on this site.
Alongside that, technical lead on a nine-engineer team building a personnel and security onboarding platform from scratch to replace a commercial Workday deployment. C#/.NET 8, Next.js, TypeScript, multi-schema PostgreSQL. That work is the reason the tooling behind this practice is engineered rather than scripted: gpuaudit has a schema layer, an amortization layer, a cluster allocation module and a regression suite, because that is how you build something whose output a client will act on.
Federal and government clients throughout. Active Top Secret clearance, which is worth mentioning here for one specific reason: you are being asked to send your billing data and cluster readings to someone outside your company, and it is fair to want to know who is on the other end of that.
Generic cloud cost work is finished. This is the part nobody instrumented.
Rightsizing, idle instances, commitment coverage. That work is commoditized, several tools do it for free, and the practitioners surveyed in the 2026 State of FinOps report say the obvious waste is already gone. A firm selling you that today is selling you a report you could generate yourself.
The question almost nobody can answer is what fraction of GPU spend produced work, because answering it requires joining billing data to hardware telemetry that most organizations have never turned on. Granular AI and GPU monitoring is the single most-requested unmet tooling capability in that same survey, and 98 percent of FinOps teams now manage AI spend, up from 63 percent last year and 31 percent two years earlier (2026 State of FinOps, 1,192 respondents).
So the practice is deliberately narrow: GPU cost attribution for teams running machine learning on Kubernetes. We say no to work outside that, including work we could technically do, because a specialist who takes generalist engagements stops being a specialist.
Six commitments, and we will hold to them against our own interest.
Fixed fee, never contingency
A percentage of savings gives the person measuring your infrastructure a stake in the size of the number. We will not take that structure even when a client offers it.
Confidence on every finding
Proven, likely, or unproven. Proven means the evidence covers the period rather than a single reading of it, so that column grows with how long we measure and we say which it was. Unproven dollars stay out of the recoverable total, and the exclusion is enforced in the tooling because under pressure discipline is what bends.
We will tell you when there is nothing
If the snapshot comes back clean we say so on the call and stop. A bad engagement costs more than no engagement, for both sides.
The measurement is reproducible
The measurement record publishes every registered row, and the protocol, scripts and raw logs are on GitHub. Rerun the scripts and check the readings against ours. Every measurement we publish can be checked that way; where we cite someone else's number instead, such as the survey figures above, we name the source so you can check theirs.
We never touch your systems
Reports touch nothing at all. In a Sprint, your engineers apply every change through your own review process; we write the spec, check the result against the bill, and never hold credentials to your systems.
Scope stays where it was written
The SOW names the specific items agreed at the end of the previous tier. Anything outside it gets repriced in writing rather than absorbed and resented.
You are sending a stranger your billing data. Here is exactly what that means.
What we ask for
- Read-only, always. Billing exports and cluster reads. No write permissions at any tier.
- Files, not access. At every tier your engineers run the CUR 2.0 export and the collector and send us the files.
- No credentials at any tier. An IAM role, an access key, a console login or a kubeconfig is declined if offered, so there is nothing to revoke at the end.
- A mutual NDA signed before you send anything for a paid engagement, and for the free snapshot too if you ask.
What we never do
- Install agents on your production nodes.
- Touch production traffic routing, at any tier.
- Subcontract, offshore, or grant a third party access to your environment.
- Keep your files more than 30 days after close-out. The deliverable is yours; the working copies are deleted within 30 days.
Security and data handling questions get a direct answer on the first call, in as much detail as your security team needs. How your files are stored, incident notification and subprocessors are all fair game, and the answer you get is the current state rather than an aspiration.
Start with the free snapshot, or just ask a question.
Forty-five minutes and one month of your own billing export is enough to know whether there is anything here. No credentials to anything, and if there is nothing we will say so on the call.
- Direct line
- 301-802-1073
- Specimen
- Read a specimen report
- Engagements
- Scope, fees and terms
- Practice location
- Miami, Florida
- Response time
- Same business day