Compare GPU reservation cost with actual useful work
On this page
Use the finest common scope and period to compare Effective Cost, the commitment program's own utilization measure, GPU activity, and a locally defined accepted-output measure. Do not treat GPU activity as proof of commitment use or useful work.
Why this is worth a look
FOCUS defines Effective Cost for resources, services, or contract commitments recognized in a charge period and distinguishes it from Billed Cost. Azure exposes reservation utilization, history, and resource details. AWS defines GPU utilization as the percentage of a sample period when one or more GPU kernels ran, while GKE describes utilization as an indication of how busy a monitored GPU is. Define workload output explicitly from local records and do not substitute GPU activity for accepted output. Use a shared period and the finest scope common to the available records. If billing is only available at reservation or resource scope, keep node or job activity as a separate rollup instead of forcing a join.
Run this check
CHECKLISTA read-only worksheet for one commitment program, scope, and period. Record provider-reported utilization where available, add GPU telemetry only for supported services, and calculate ratios only when their units and denominators are present.
READ-ONLY WORKSHEET
Execution surface: manual review of billing, commitment, monitoring, and local workload records. No resource changes.
Required access: read access to the applicable cost or commitment view and monitoring service or workspace. Workload records are local inputs.
Scope: one provider, account or project, reservation or commitment, compatible resource scope, and reporting period.
Units: currency in the billing currency; time in hours; utilization and activity in percent; output in one fixed local unit.
IDENTIFIERS
Period start: ____________________ Period end: ____________________
Provider: AWS / Azure / GCP / other: __________
Account, subscription, or project: ____________________
Commitment or reservation program and ID: ____________________
Resource, cluster, or pool scope: ____________________
GPU type and region, if exposed: ____________________
COST
[ ] Billed Cost for the period: __________ [currency: ______]
[ ] Effective Cost for the period: ________ [currency: ______]
[ ] Effective Cost source or field: ____________________
COMMITMENT UTILIZATION
[ ] Provider-reported utilization value: __________ [% or program unit]
[ ] Provider-reported utilization period and scope: ____________________
[ ] Eligible usage or capacity: __________ [unit: ______]
[ ] Committed capacity or eligible spend: __________ [unit: ______]
[ ] Use the next ratio only if both units are compatible:
eligible usage or capacity / committed capacity or eligible spend = __________
Label: commitment-use ratio
[ ] If the program does not expose compatible units, write N/A and retain its reported utilization.
CAPACITY UNITS, ONLY WHEN THE PROGRAM REPORTS THEM
[ ] GPU count: __________ Availability schedule: ____________________
[ ] Committed GPU-hours: __________ [GPU-hours]
[ ] Available GPU-hours: __________ [GPU-hours]
[ ] Coverage ratio, only with compatible capacity units:
covered eligible GPU-hours / total eligible GPU-hours = __________
[ ] Allocation ratio, only when local allocation records define both terms:
allocated GPU-hours / available GPU-hours = __________
GPU ACTIVITY, ONLY FOR SUPPORTED TELEMETRY
[ ] Metric name: ____________________
[ ] Sample period and scope: ____________________
[ ] Average GPU activity: __________ [%]
[ ] Memory activity or used memory, if available: __________ [%, MB, or other metric unit]
[ ] For an AWS Linux server where the CloudWatch agent is configured to collect NVIDIA GPU metrics and an NVIDIA driver is installed, record nvidia_smi_utilization_gpu: __________ [%]
[ ] For AKS configured with NVIDIA DCGM Exporter and Azure Monitor managed service for Prometheus, record available utilization and memory metrics: ____________________
[ ] For GKE with DCGM metrics, record available utilization and memory metrics: ____________________
LOCAL WORKLOAD OUTPUT, IF RECORDED
[ ] Output unit definition: ____________________
[ ] Successful output units: __________ [fixed local unit]
[ ] Accepted output units: __________ [fixed local unit]
[ ] Accepted output units / allocated GPU-hour: __________
[ ] Effective Cost / accepted output unit: __________ [currency / unit]
[ ] If accepted output is zero, record N/A, not a numeric unit cost.
REVIEW
[ ] Every comparison uses the same period and the finest common scope.
[ ] Billed Cost and Effective Cost remain separate.
[ ] Unsupported or incompatible fields are marked N/A.
[ ] A commitment change is not recommended from GPU activity alone.How to confirm it
- 01
Set the common scope
Choose one provider, commitment or reservation program, account or project, resource scope, and reporting period. Use the finest scope shared by the records. If billing is only at reservation or resource scope, report node or job telemetry separately rather than joining it to billing at that grain.
- 02
Record both cost views
Capture Billed Cost for reconciliation and Effective Cost for period analysis, both in the billing currency. FOCUS says Effective Cost includes recognized portions of related purchase charges for usage charges and can differ from Billed Cost when covering and covered charges span periods or billing accounts.
- 03
Use the program's own utilization measure
Record the commitment program's reported utilization, eligible usage, or eligible spend only where that program exposes it. Azure's reservation view provides utilization percentage, history, and details about resources that used the reservation. Do not convert a spend-based measure into GPU-hours without a compatible provider-reported unit.
- 04
Add supported activity and local output
For an AWS Linux server where the CloudWatch agent is configured to collect NVIDIA GPU metrics and an NVIDIA driver is installed, record nvidia_smi_utilization_gpu and its sample period. For AKS configured with NVIDIA DCGM Exporter and Azure Monitor managed service for Prometheus, record the available utilization and memory metrics. GKE DCGM metrics provide documented GPU activity and memory measures. If local workload records contain an approved output unit, define it once, keep it stable, and mark unit cost N/A when accepted output is zero.
Before making changes
This comparison assumes the billing, commitment, telemetry, and local workload records cover the same period and a documented common scope. Provider-reported utilization and eligible units vary by commitment program, so mark unavailable or incompatible fields N/A. GPU activity is sampled kernel or device activity, not proof that output succeeded. Local allocation and accepted-output measures are assumptions that require records and definitions from your environment.