Turn a GPU enquiry into a comparable server quotation with workload targets, deployment constraints and acceptance tests.
Start with the workload
For AI inference, share the model family, expected input lengths, peak requests, concurrency and response-time target. For training, share dataset size, framework, checkpoint frequency and target time to complete a run. For engineering simulation, supply a representative case and the solver version.
Specify the complete system
Ask each bidder to itemise GPU memory and count, CPU and RAM, usable storage, network links, cooling, rack space, power draw and software or support costs. Confirm which components are included in the base quote and which are optional.
Agree on the acceptance test
Use the same representative workload for every proposal. Record throughput, latency and peak memory under sustained load. Include the delivery location, deployment window and required support coverage in the BOQ.
What to send with your GPU server BOQ
A useful BOQ starts with one real workload and a measurable target. If you need to serve an internal knowledge assistant, describe the model, the typical and maximum prompt lengths, the number of simultaneous users and the response time they can tolerate. If your team fine-tunes models overnight, include the framework, precision, model and dataset sizes, checkpointing requirements and the time window available for a run. These inputs help suppliers explain how they arrived at their recommendation.
Document the deployment environment as well. Include the site, rack space, usable power and cooling capacity, existing network speeds, and any restrictions on where data can be stored. If you have an existing cluster, describe how the new server must connect to storage, monitoring and identity systems. Ask bidders to list their assumptions where information is still missing.
How to compare two GPU server quotations
Where one vendor proposes a smaller system and another adds capacity for growth, ask both to show the workload and utilisation assumptions behind their proposal. Use the same contract period and support scope when comparing costs.
Worked example for an internal AI assistant
Suppose a team expects 40 employees to use an assistant, with eight active conversations at peak. Give each bidder the same test model and representative set of short and long prompts. Ask them to record time to first response, generation throughput, peak GPU memory and failed requests at the expected concurrency. Repeat the test with the longest realistic inputs and compare the results with the agreed response targets.
Check what happens when the request rate exceeds the planned peak. The team may need queuing, a second instance or tighter limits on input length. A short pilot makes those trade-offs visible before the hardware configuration is finalised.
What should the acceptance test cover?
Name the person who will validate the workload and agree on pass criteria before the purchase order. Include a sustained test, hardware inventory verification, monitoring checks and a handover of support contacts. Ask for a record of the software versions and configuration used during testing so the IT team can reproduce the results after deployment.
Common questions
Can I request a quote before choosing a GPU model? Yes. Provide the workload, expected concurrency, deployment location and success criteria, and ask bidders to explain their proposed configuration.
Can two smaller GPUs replace one larger GPU? It depends on how the model and software divide work and memory across devices. Ask for a demonstration using your actual workload before treating the configurations as equivalent.
Request a GPU BOQ review
Share your workload, project location and deployment window. Layots can review the requirements with your team before you request quotations.