Procurement workflow

GPU cloud procurement checklist

Use this checklist before turning a GPU cloud shortlist into committed spend. It connects workload fit, exact hardware, source-backed pricing, capacity evidence, networking, operations, security, contracts, and decision provenance so the cheapest visible rate does not become the entire procurement process.

Start the checklist Compare current prices Model workload cost

54 review checks across nine decision areas
Pricing and availability remain separate evidence categories
Security and legal approval stay with the responsible human owner

How to use it

Separate evidence, assumptions, and approvals

SaaS Sentinel can supply market pricing, historical observations, provider breadth, benchmark evidence, and some availability signals. Your team supplies workload measurements, security requirements, contractual constraints, and final approval. Do not allow a missing data field in one layer to be silently invented by another.

Evidence

What the market currently shows

Source URL, observed rate, timestamp, region, purchase model, history, benchmark evidence, and qualifying availability signals.

Assumptions

What your workload model needs

Throughput, utilization, scaling, data volume, interruption recovery, time horizon, and non-compute cost estimates.

Approval

What a responsible owner must decide

Security, privacy, compliance, contractual risk, support expectations, budget ownership, and acceptable concentration risk.

Checklist section 1 of 9

1 · Workload definition

Define the useful work before comparing infrastructure. If the workload target is vague, every downstream price comparison becomes easier to manipulate.

The workload has a measurable completion or service target.Required precision, model size, batch shape, sequence length, or other material workload settings are recorded.A measured throughput value or a clearly labeled benchmark assumption is available.Target utilization is explicit rather than assumed to be 100%.Deadline, latency, uptime, or completion-window requirements are written down.The workload's tolerance for interruption, retries, and checkpoint recovery is known.

Checklist section 2 of 9

2 · Exact GPU product and topology

A GPU family name is not enough to prove equivalence. Preserve the hardware details that can change fit, throughput, or scaling.

Exact GPU model and memory capacity are confirmed.Variant or form factor such as PCIe, SXM, NVL, or another relevant configuration is known.Required GPU count per node is explicit.NVLink, InfiniBand, or other interconnect requirements are documented when multi-GPU or multi-node scaling matters.CPU, RAM, local storage, and host constraints are sufficient for the workload.The compared offers represent materially equivalent node shapes or the differences are carried into the decision.

Checklist section 3 of 9

3 · Pricing and billing

Normalize the compute rate, then add the material commercial rules instead of hiding them inside an unexplained multiplier.

The source-backed GPU rate and currency are recorded with observation time.The rate is normalized to a comparable per-GPU-hour unit when appropriate.Purchase model is identified: on-demand, spot/interruptible, reserved, community, serverless, or another model.Minimum billing duration, rounding, startup charges, or reservation commitments are understood.Storage, egress, public IP, support, taxes, and other material non-compute costs have been checked.Any negotiated or volume discount is separated from the public market baseline.

Checklist section 4 of 9

4 · Availability and capacity evidence

Do not turn a price listing into a stock claim. Capacity evidence needs its own source, scope, and timestamp.

Current pricing evidence and current availability evidence are treated as separate claims.The availability signal applies to the required GPU, region, and purchase model.The observation timestamp is recent enough for the volatility of the source.Required GPU count is supported; evidence for one GPU is not assumed to prove an eight-GPU node.Account quota, verification, credit, allow-list, or reservation prerequisites have been checked.A fallback region, provider, or GPU is defined when capacity risk is material.

Checklist section 5 of 9

5 · Storage, data and networking

Cheap compute can be a false economy if data movement or storage becomes the dominant operational cost or performance bottleneck.

Dataset location and expected ingress path are known.Persistent storage capacity, performance, snapshot, and lifecycle requirements are known.Expected egress volume and destination are estimated.Inter-region or cross-cloud data-transfer charges have been checked when relevant.Network bandwidth and latency are sufficient for the workload and distributed topology.Data staging, cache warm-up, and restart transfer time are included when they affect time-to-completion.

Checklist section 6 of 9

6 · Reliability and operations

The procurement decision includes the work required to keep the workload running, not only the price of the accelerator.

Provisioning and teardown can be automated or operated reliably by the team.Monitoring, logs, failure detection, and alerting meet the workload's operating requirements.Checkpointing and retry behavior are tested if interruptible capacity is in scope.Provider maintenance, eviction, or failure behavior has an operational response plan.A realistic recovery-time objective exists for production or deadline-sensitive workloads.Exit or migration effort is understood if the provider becomes unavailable or uneconomic.

Checklist section 7 of 9

7 · Security, privacy and compliance

Treat this section as an escalation boundary: the public pricing dataset cannot answer security or legal requirements for your organization.

Required data residency and geographic restrictions are defined.Encryption, identity, access-control, network-isolation, and secret-management requirements are checked against provider documentation.Required certifications, audit reports, DPAs, subprocessors, or contractual security terms have been reviewed by the responsible team.Sensitive model weights, customer data, or regulated datasets have an approved handling path.Logging and telemetry behavior is compatible with internal privacy and retention requirements.The person responsible for security/compliance approval is named before production commitment.

Checklist section 8 of 9

8 · Contract, support and commercial risk

A low rate is not a complete commercial offer. Record the terms that determine what happens when the workload, bill, or relationship goes wrong.

Support level and escalation path match the importance of the workload.Service terms, credits, cancellation, commitment, and renewal conditions have been reviewed where material.The organization understands who owns the provider account, billing relationship, and operational credentials.Price-change exposure is understood for non-committed usage.Budget guardrails and spend alerts are configured before large-scale production use.The decision has at least one credible alternative provider, architecture, or fallback strategy when concentration risk matters.

Checklist section 9 of 9

9 · Decision evidence and sign-off

A good infrastructure decision can be reviewed later. Preserve the inputs that came from evidence and the assumptions that came from your team.

Selected offer, provider, region, purchase model, and source URL are recorded.Pricing and availability observation timestamps are recorded separately.Workload assumptions and benchmark or measurement provenance are recorded.Base, downside, and alternative-provider scenarios have been compared.The reason the selected option wins is written in workload terms, not just hourly price.Owner, review date, and conditions that should trigger a re-evaluation are documented.

Final decision record

Write the recommendation so another person can challenge it

A procurement record should explain what was selected, what evidence was current at the time, which assumptions drive the result, what alternative was considered, and what change would cause the team to revisit the decision.

Selected option

Provider, exact GPU/node, region, purchase model, expected scale, and source-backed rate.

Why it wins

Cost per completed workload, technical fit, availability evidence, operational fit, and material commercial constraints.

What could invalidate it

Price movement, capacity loss, workload change, throughput miss, contract change, security requirement, or a better evidenced alternative.

Fill pricing evidence Review capacity evidence Check price history Check benchmarks