Use case
Rightsize over-provisioned GPU instances
Find training and inference boxes running on far more GPU than they use, and get the exact smaller instance or accelerator to switch to — safely.
eks · prod-workers
$4,800/mo
High
rds · analytics-staging
$1,180/mo
Medium
ecs · checkout-api
$640/mo
High
ec2 · 14 × gp3
$312/mo
High
The problem
GPU instances are expensive and easy to over-size. A box gets provisioned for the peak of a training run, or a multi-GPU type is chosen “to be safe,” and then it runs for weeks at a fraction of its capacity. Because GPU hours are billed at a premium, a little over-provisioning is a lot of money — and CPU-oriented cost tools don’t even look at GPU utilization.
How Lizrd fixes it
Lizrd reads real GPU and GPU-memory utilization read-only (via the cloud’s GPU metrics), finds accelerators running well under what they’re paying for, and proposes the exact smaller instance type or accelerator — the right accelerator at the right size — with 30-day evidence, a confidence level, and the diff to apply. It works across EC2/SageMaker, Vertex AI, and Azure ML.
The outcome
Right-sized GPUs without a benchmarking project, every change reversible, and the savings proven back against your real bill once it ships.
“A four-GPU training box was sitting at 14% utilization for weeks. Lizrd showed the evidence and the one-line diff to drop it to a single GPU.”
More ways to save
Cover steady usage with commitments
Savings Plans and committed-use discounts are free money on steady baselines — if you size them right. Lizrd finds your safe, always-on floor to commit against.
Cut idle networking waste
Idle NAT gateways, unattached elastic IPs, and load balancers with no traffic bill around the clock. Lizrd finds them and confirms what's safe to remove.
Delete orphaned resources
Unattached volumes, unused IPs, and stale snapshots keep billing with nothing using them. Lizrd finds them and confirms they're safe to remove.
Find this in your own cloud
Connect read-only and Lizrd surfaces the highest-impact fixes — with the exact change to make.