Illustrative method · not a client case study

An illustrative GPU decision memo

This memo uses a fictional workload to demonstrate an evidence register, cost boundary and decision gates. It does not define or promise a paid service or deliverable. Paid services are under definition, including scope, deliverables, price and capacity. No paid offer or intake is active.

Read before using this sample

Demonstration only: no customer, quote or outcome is implied

Provider names and rates below came from the cited official public sources in the frozen public-price snapshot. The workload, shortlist and disposition are invented solely to demonstrate the review format. Pricing and availability may now be historical; verify the current catalog before any decision.

Illustrative decision memo

H100 inference launch · pre-shortlist challenge

Disposition: hold for evidence

1. Executive conclusion

Do not approve a provider from these public price rows alone. The selected candidates produce a published compute-only boundary of $3,420–$5,400 per month for1,200 GPU-hours, but the rows do not establish allocatable quantity, workload throughput, full regional fit or the non-compute cost boundary.

The next decision is not “which hourly rate is lowest?” It is “which candidate can prove the required capacity, region, topology and fully loaded cost without invalidating the launch plan?” Keep all three in evidence collection; approve none until the blocking questions below have written answers.

2. Scenario and calculation boundary

Known4 concurrent H100 GPUs · 300 runtime hours
Formula4 GPUs × 300 hours × published USD/GPU-hour
ExcludedStorage, egress, support, tax, engineering time, discounts and idle capacity
Not inferredThroughput, latency, quota, reservation, compliance or framework compatibility

3. Evidence and cost register

The table compares only the selected public rows. It is not a complete market ranking.

CandidatePublished rateCompute boundaryRegion evidenceCapacity evidenceOfficial source
HyperstackH100 · on demand$2.8500 / GPU-hour$3,420 / monthNo named region in the selected public rowNo current capacity proof in the selected public rowOpen source ↗Observed 2026-11-01
LambdaH100 · on demand$3.7900 / GPU-hour$4,548 / monthNo named region in the selected public rowNo current capacity proof in the selected public rowOpen source ↗Observed 2026-11-01
Nebius AI CloudNVLink 80GB · on demand$4.5000 / GPU-hour$5,400 / montheu-north1Deployability observed; allocatable quantity not publishedOpen source ↗Observed 2026-11-01

4. Approval blockers and provider questions

Blocking evidence gaps

  • Written confirmation of 4 simultaneously allocatable GPUs in the required region.
  • Exact GPU variant, node topology, interconnect and tenancy model.
  • Workload-representative throughput and latency under the proposed configuration.
  • Storage, data transfer, support, minimum-commit and termination cost assumptions.
  • Security, data residency and incident-response evidence required by the buyer.

Questions sent to every candidate

  1. Can you allocate the full configuration by the required date, and what evidence makes that commitment enforceable?
  2. Which region, GPU variant, host shape and interconnect does the quoted rate cover?
  3. Which charges are excluded from the hourly price, and under what minimums or commitments?
  4. What happens commercially and operationally if capacity is delayed or withdrawn?
  5. Which workload benchmark can the buyer reproduce before signature?

5. Decision gates and revalidation triggers

Gate 1 · technical

The reproducible workload test meets the buyer's documented threshold on the exact proposed topology.

Gate 2 · capacity

Region, quantity, start date and reservation or quota mechanism are confirmed in writing.

Gate 3 · commercial

The fully loaded cost and downside case fit the approved budget with material exclusions resolved.

Reopen when

Price, capacity, workload, region, topology, contract term or evidence freshness changes materially.

Explore the method

Use the free decision tools with your own assumptions

The sample explains how evidence, costs, uncertainties and revalidation conditions can be recorded together. It is an educational example, not a commitment to provide a review, memo or clarification round.