OnpremBench.aiby Understand.tech

Help improve this record. Suggest a correction

Field notebook

Economics / All platforms

What does an accepted result actually cost?

Put quality, retries, utilization and operating effort beside token throughput.

Evaluation checklistUnderstand TechUnderstand Tech LabReviewed 9 September 2026

CONTEXT

Why this comes up

UT’s founder reported substantial savings after moving a cloud workload to two Sparks. Another organization needs its own workload and cost assumptions.

THE USEFUL LESSON

Compare only options that meet the workload’s requirements. A low token cost is useful when the resulting work is accepted and the system can deliver the required service.

ON YOUR MACHINE

What to record and test

  1. 01

    Fix an acceptance set and the required quality, latency and data boundary.

  2. 02

    Count completed and accepted tasks, including retries and human correction.

  3. 03

    Include hardware, installation, software, power, support, administration and any required redundancy.

  4. 04

    Use measured demand across the operating period, including idle time.

  5. 05

    Compare with an eligible cloud option; treat confidentiality as a requirement when it rules that option out.

A costing method, not a savings forecast. The calculator contains editable examples rather than supplier or UT quotations.