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–$4,620 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
3. Evidence and cost register
The table compares only the selected public rows. It is not a complete market ranking.
| Candidate | Published rate | Compute boundary | Region evidence | Capacity evidence | Official source |
|---|---|---|---|---|---|
| HyperstackH100 · on demand | $2.8500 / GPU-hour | $3,420 / month | No named region in the selected public row | No current capacity proof in the selected public row | Open source ↗Observed 2026-09-09 |
| LambdaH100 · on demand | $3.7900 / GPU-hour | $4,548 / month | No named region in the selected public row | No current capacity proof in the selected public row | Open source ↗Observed 2026-09-09 |
| Nebius AI CloudNVLink 80GB · on demand | $3.8500 / GPU-hour | $4,620 / month | eu-north1 | Deployability observed; allocatable quantity not published | Open source ↗Observed 2026-09-09 |
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
- Can you allocate the full configuration by the required date, and what evidence makes that commitment enforceable?
- Which region, GPU variant, host shape and interconnect does the quoted rate cover?
- Which charges are excluded from the hourly price, and under what minimums or commitments?
- What happens commercially and operationally if capacity is delayed or withdrawn?
- 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.