Illustrative draft · provisional fictional sample
GPU DECISION REVIEW
Illustrative Draft GPU Decision Brief
AWS 8 x H100 Inference Renewal Review. Rendered from a provisional fictional sample Decision Record using captured Cloud Advisor and GPU Audit evidence.Review the current AWS 8 x H100 path before the next inference commitment.
Northstar AI operates a continuously running 8 x NVIDIA H100 inference workload in Northern Virginia. The captured run uses the published AWS p5.48xlarge on-demand rate because no private paid rate was entered. GPU Audit applies 730 billed hours per GPU-month and the visible 12% idle or underutilized-time planning default. The company, owner, timing, and decision narrative are fictional; the captured inputs and product evidence are preserved.
Keep AWS as the published baseline while validating alternatives before the next commitment.
The AWS configuration and published price are traceable for the stated workload and operating scope. No lower exact-market, technically comparable, orderable alternative has been confirmed. The present evidence supports a verification cycle, not a provider switch.
Provisional · owner confirmation pendingThe workload and commercial basis this draft covers.
- Workload
- Inference, 24/7
- Quantity and GPU
- 8 x NVIDIA H100 on one instance
- Minimum shape
- 176 vCPU and 960 GiB memory
- Geography
- Northern Virginia operating scope
- Purchase basis
- On-demand published catalog price
- Planning inputs
- 730 h/GPU-month; 12% modeled idle time
What is retained, investigated, and unsupported.
Current path
RETAIN AS BASELINEAWS p5.48xlarge in US East (N. Virginia), using the provider-published on-demand catalog rate. This is not the buyer's paid rate.
Broader-market signal
HELD FOR VERIFICATIONA $3.90 per GPU-hour Crusoe/global signal was captured outside the accepted exact-market comparison. Geography, terms, capacity, fees, and technical comparability remain unconfirmed.
Immediate provider switch
NOT SUPPORTABLEThe captured evidence does not establish a comparable offer, orderability, capacity, quote, workload performance, or migration case.
Accepted facts, explicit models, and held signals stay separate.
The unresolved external-market signal is excluded from comparable alternative economics.
AWS published baseline
ACCEPTED PUBLISHED FACT$6.88 / GPU-hour$55.04 per instance-hour and $482,150.40 annualized at 8 GPUs x 730 hours x 12 months. This is a catalog basis, not a paid rate.
Modeled idle exposure
EXPLICIT MODEL$57,858.05 / year12% of the AWS published annual basis. This is planning math, not measured telemetry and not verified savings.
External-market signal
UNQUALIFIED · HELD$3.90 / GPU-hourCaptured as a broader/global signal only. It is not annualized, compared as an alternative, or used in the decision economics while exact-market review remains open.
The draft separates what is known from what cannot yet be claimed.
Confirmed facts
- The captured query specifies an 8 x H100, 24/7 inference workload in the Northern Virginia operating scope.
- AWS p5.48xlarge supplies 8 x H100, 192 vCPU, and 2,048 GiB in the captured catalog evidence.
- The provider-published AWS on-demand basis is $55.04 per instance-hour, or $6.88 per GPU-hour.
- The evidence package is dated through August 24, 2026; the captured run warns that some pricing rows were excluded.
Open confirmations
- The buyer's actual paid rate, current term, renewal proposal, and reservation structure.
- Exact-market mapping and like-for-like eligibility for the $3.90 external-market signal.
- Provider orderability, capacity window, quote, deposit, minimum, and service commitments.
- Measured utilization, workload performance, software stack, networking, storage, egress, support, and tax.
Excluded implications
- No verified savings, comparable alternative annual cost, or actionable price gap is established.
- No catalog entry proves availability, orderability, capacity, a live quote, or final paid rate.
- No workload benchmark or telemetry establishes performance, utilization, migration feasibility, or fit.
- No provider selection, customer confirmation, Decision Record seal, or monitoring activation has occurred.
Close the evidence gaps before commitment.
Commercial basis risk
The AWS benchmark may differ from the buyer's paid or proposed rate.
Comparability risk
The lower signal is outside the accepted exact-market comparison until geography and offer terms are proven.
Workload-fit risk
No benchmark or confirmed architecture-sensitive requirement proves that another configuration can run the workload.
- Confirm the buyer's current AWS paid rate, proposed term, minimums, and capacity commitment.
- Ask each provider whether the exact 8-GPU configuration is orderable for the required dates and geography.
- Collect networking, storage, egress, support, tax, deposit, and reservation terms outside compute.
- Confirm drivers, frameworks, interconnect, service commitments, and workload benchmark requirements.
Proposed checkpoints, not active monitoring.
- A verified, technically comparable offer becomes available in the required market.
- AWS pricing, term, capacity, or service commitments materially change.
- Measured idle or underutilized GPU time differs materially from the 12% planning assumption.
- Workload size, duty cycle, GPU requirement, region, or performance constraints change.
Next review: October 27, 2026
The recheck conditions are proposed for owner confirmation. They are not active monitoring triggers.
No monitoring is included in this illustrative draft. Any future monitoring would require separate confirmation, consent, and activation.
The Brief renders dated evidence; it does not rerun the tools.
Cloud Advisor captured evidence
Workload-shape and AWS published-baseline evidence
Captured August 28, 2026 · Evidence through August 24, 2026GPU Audit captured evidence
AWS annualization and explicit 12% planning model
Captured August 28, 2026 · Evidence through August 24, 2026Crusoe/global catalog signal
Unqualified external-market signal held for verification
Captured August 28, 2026 · Evidence through August 24, 2026Evidence supports the record; it does not author the decision.
- Cloud Advisor and GPU Audit provide dated evidence packages. They do not author, approve, or seal the decision.
- This Brief performs no calculation or offer selection. It renders accepted facts and explicit unknowns from the provisional fictional sample Record.
- This artifact is an illustrative sample, not a delivered Brief or a recommendation.
- This fictional public sample contains no customer data. Customer decision data must remain protected inside the authenticated AIForge Works boundary and may not be externally processed or delivered without explicit authority.