Your GPUs are idle. Your bill isn't.

We audit what your cloud GPUs cost and how busy they really are, per team and per job, then hand you a savings and capacity plan you can act on. Two weeks, CAD 15,000 fixed fee, read-only access, AWS or GCP.

Every audit is run by one senior ML platform lead, who runs a $12M+ NVIDIA A100/H100 platform on EKS at a global pharma, with SageMaker and Bedrock in production for 7+ teams.

What the audit finds.

Read-only access to your billing export and cluster metrics. Every finding comes with the change that fixes it and what it's worth a year.

Where every GPU-hour goes, by team and by workload.

Allocation against use
GPU-hours used of GPU-hours allocated, per team, per workload and per hour of the day.
Idle and oversized
Instances that sit idle, and jobs running on GPUs bigger than they need.
Scheduler and quotas
Queues, quotas, MIG and time-slicing that keep the GPUs you have busy.
Commitments, before you renew
The Savings Plan or CUD amount and term your real usage supports, split between Compute and SageMaker plans.
SageMaker and Bedrock
Endpoints, instance types and provisioned throughput, right-sized.
Capacity plan
What you'll need for the next 12 to 24 months, and where to get it: reservations, capacity blocks or another region.
Sample readoutGPU-hours used of allocated, by workload, last 30 days
  1. Notebooks18%
  2. Research31%
  3. Batch scoring44%
  4. Training runs58%
  5. Inference72%
Illustrative, synthetic data. Your readout uses your billing export and cluster metrics.

Why now

40% of enterprises say their actual cloud AI spending exceeds their initial projections. Enterprise AI Infrastructure Survey 2026, commissioned by Cloudian, 203 IT decision-makers, February 2026

Across the tens of thousands of Kubernetes clusters Cast AI analyzed, GPU utilization averaged 5% before automation was turned on. Cast AI, 2026 State of Kubernetes Optimization Report, April 2026

Ten working days, one readout.

Read-only access, nothing that writes. You keep every model and spreadsheet.

  1. Access

    Days 1 and 2

    Read-only roles to the billing export, cluster metrics and job logs, under NDA.

  2. Measure

    Days 3 to 5

    Utilization and allocation per team, per workload and per hour; idle and oversized capacity.

  3. Model

    Days 6 to 9

    Scheduler, quota and commitment changes, each with its saving and its effort.

  4. Readout

    Day 10

    The ranked savings list and the capacity plan, with your platform and finance leads.

Sample readoutGPUs needed against GPUs reserved, next 12 months
0306090Now+6 months+12 months
DemandReserved, in three stepsOn-demand
Illustrative, synthetic data.

Billing reports find idle GPUs. Schedulers keep them busy.

Cost Explorer and Kubecost show what you spent. The savings that last come from deciding who gets which GPUs, when, and on what terms: quotas per team, MIG and time-slicing for small jobs, Karpenter for bursts, and commitments sized to the demand you actually have. Those are platform changes, and the implementation makes them with your team.

Start with the audit. Take the savings after.

Fees in Canadian dollars. The audit is a fixed fee, credited toward the implementation, which is quoted from what the audit finds.

GPU Spend & Capacity Audit

Length2 weeksFeeCAD 15,000

For
Teams running 8 to 200 GPUs on AWS or GCP
You get
Per-team utilization and allocation, a savings list ranked by effort, the scheduler, quota and commitment changes that deliver it, and a 12 to 24 month capacity plan, in a readout with your platform and finance leads.
Works in
EKS, GKE, SageMaker, Bedrock, Vertex AI, Karpenter, NVIDIA MIG, Kubecost, Terraform
Terms
Fixed fee, credited toward the implementation.
Book your GPU audit

Implementation

Length6 to 10 weeksFeefrom CAD 60,000

For
Teams that want the savings taken, not listed
You get
Quotas and queues per team, MIG and time-slicing, Karpenter autoscaling, commitments sized for your finance team to buy, SageMaker and Bedrock right-sized, and chargeback your finance team can read. Built in your accounts with your platform team, in Terraform.
Scope the fix

One senior lead, on every audit.

Sovont is run by a senior ML platform lead who runs a $12M+ NVIDIA A100/H100 platform on EKS at a global pharma: provisioning automated with Terraform for 23% more throughput, and SageMaker and Bedrock models in production for 7+ teams, 26% faster.

Sovont runs alongside that role, so it takes on 2 to 3 engagements at a time and the lead runs every audit personally. Before that: ML platforms and the teams that run them at a national grocery chain and a US pet retailer.

At Sovont, we've helped 5‑person ML teams and 50‑engineer data orgs alike.

Common questions.

Something else? hello@sovont.com

What access do you need?
Read-only: your billing export (AWS cost and usage reports or the GCP billing export), cluster metrics, and the scheduler or job logs. Nothing that writes, all under NDA. If GPU metrics aren't exported yet, turning on NVIDIA's DCGM exporter is the one change we'd ask your team to make.
Our Savings Plan renews soon. Is two weeks enough?
Yes, if access starts at least three weeks before the renewal. The day-10 readout gives the amount, the term and the payment option, and what each saves against on-demand.
Do you replace our cost tool?
No. We start from what it shows and go further: the scheduling, quota and commitment changes that keep the savings.
What if we run our own GPUs, or colocate?
We compare your cloud plan against owning or colocating, on cost and lead time. Facility and electrical design are out of scope.
What size teams do you work with?
The audit is priced for 8 to 200 GPUs. Below 8 the savings rarely cover the fee; above 200 we scope it on a call first.
When can you start?
The lead runs every audit personally, so we take on 2 to 3 engagements at a time. If a Savings Plan or CUD renewal is coming, write early.

Book your GPU audit.

Or ask a question. Tell us what you run, roughly what it costs, and when your next commitment renews. We'll reply within 24 hours with whether the audit fits; next is a 30-minute call, then a fixed scope and start date in writing.

Or email hello@sovont.com.

Topic