Inside a procurement comparison

Make the tradeoffs visible.

An illustrative example of how gpu.pm structures a requirement and compares the costs, evidence and open questions behind provider options.

Discuss your requirement ↗
This is a worked example. The options and figures below are hypothetical. They do not represent suppliers, quotes, customer work, live capacity or an offer to book.

01 / The requirement

Define what has
to be true.

Workload

An AI team needs a dedicated single node for a 90-day deployment. Performance and model compatibility still require a representative benchmark.

Configuration

8 × H100 SXM, 80 GB per GPU, intra-node NVLink. Confirm host RAM, CPU, usable disk, access method and software environment.

Constraints

Europe, continuous rental, 2,160 hours per GPU. Exact country, start date, data residency and support requirements to be agreed.

Illustrative procurement brief

8 GPUs · Europe · 90 days

Example only
Hypothetical configurations and figures, not supplier quotes or confirmed availability.
What mattersOption A Lower base costOption B More included
Hardware8 × H100 SXMSingle node · 80 GB / GPU8 × H100 SXMSingle node · 80 GB / GPU
Persistent storage1 TB · $320 / 30 days1 TB · included
Compute / GPU-hour$2.80$3.10
90-day known subtotal$49,344Compute + specified storage$53,568Compute; specified storage included
Before decidingConfirm egress, support and start dateConfirm storage limits, support and start date

Same GPU count. Different inclusions. Subtotals exclude taxes, any sourcing fee and unquoted charges. Availability must be confirmed for your dates.

02 / The arithmetic

Show the full
cost basis.

Option A

8 GPUs × $2.80 × 2,160 hours = $48,384 compute. Add $320 of specified storage per 30-day period × 3 = $960. Known subtotal: $49,344.

Option B

8 GPUs × $3.10 × 2,160 hours = $53,568, including the specified storage allowance.

The $4,224 difference is a difference between known subtotals. Egress, extra storage, support upgrades, taxes and any gpu.pm fee must be priced before comparing total commitments. Neither option is established as the better choice.

03 / The evidence

A price is only one part of the decision.

TECHNICAL FIT

What will actually run?

Confirm topology, network and storage performance, usable memory, access and benchmark conditions. Single-node NVLink does not establish inter-node fabric capability.

CAPACITY & OPERATOR

What is confirmed?

Identify the operator and the seller’s authority. Record site, available dates, verification source, quote validity and any allocation dependencies.

COMMERCIAL TERMS

What are you committing to?

Compare billing units, prepayment, minimum term, SLA, support, cancellation, renewal, data handling and exit costs.

04 / The next step

Resolve the gaps
before committing.

A real comparison identifies confirmed facts, supplier statements and unknowns separately. If a configuration cannot meet the workload, a lower rate does not make it suitable.

gpu.pm’s sourcing deliverables and verification scope are agreed before the search. A comparison is decision support; the provider confirms capacity and supplies the contract.

Start your own brief ↗