OpenWorkload / Kairos

Sample decision memo ยท September 18, 2026

The cheapest cluster is not always the right cluster.

Synthetic example. All inputs are invented. This is not a customer case study, vendor quote, measured benchmark, or delivered savings claim.

The decision

A batch workload needs 72,000 accepted classifications per day, within 18 hours. Should the team retain its eight-worker cluster, tune and shrink it, or move to elastic capacity?

Provisional recommendation: test A, the tuned four-worker cluster. It lowers modeled recurring cost while retaining enough capacity after one worker fails. The cheapest option, B, misses the delivery target under that same failure.

Modeled results, before implementation or benchmark validation
MetricBaselineA: TunedB: Elastic
GPU workers844
Compute + storage + network / month$18,080$9,440$7,448
Daily capacity after one worker fails126,00081,00058,500
Meets 72,000 daily demand after that failure?YesYesNo
First three months, including transition$54,240$34,320$34,344

What could change the answer?

Inspect the full work

The full memo includes billing assumptions, exclusions, transition allowances, security tradeoffs, sensitivity tests, and acceptance gates. Keep the existing cluster until the performance, security, and budget checks pass.

Discuss an infrastructure decision