CASE 115 · HAZEL · 2026
Training commitments that match the model release cadence.
A computer vision company shipped a new model version every six weeks, with each release requiring 80+ hours of GPU training. The training was on-demand; the bill was lumpy and expensive. We added SageMaker-aware Savings Plans matched to the release cadence and brought average effective rate to 41% off list.
Computer vision
COST
2026
RESULTS
What changed, by the numbers.
EFFECTIVE TRAINING RATE
−41%
COMMITMENT WASTE
< 3%
TRAINING BILL
−38%
RELEASE PREDICTABILITY
IMPROVED
HOW IT WENT
The release cadence was predictable: training peaks every six weeks, roughly the same duration each time. The team had treated training as a spiky workload and used on-demand for everything, paying full list price. The shape of the demand actually fit a Savings Plan well — they just hadn’t modelled it.
We took eighteen months of historical SageMaker training spend, modelled the floor and the peaks separately, and bought Compute Savings Plans against the floor with a small risk margin. The peaks still ran on-demand and burned through the commitment plus overflow.
Effective rate landed at 41% off list, with commitment waste under 3%. Net training bill dropped 38%. The release-time budget request to finance became unnecessary — training fit inside the existing line item instead of spiking every six weeks.
RELATED · SAME DOMAIN
Other engagements in this space.
READY WHEN YOU ARE
Let's get your AWS bill (and architecture) in order.
The discovery call is free. You walk away with at least one concrete idea — even if we never work together.