How to Write a GPU Server RFQ Without Creating an Unusable Quote

Use brand-name-or-equal language only within the chosen acquisition strategy and with contracting-officer review. This guide is an engineering and market-research structure, not legal advice.

Write a workload paragraph: model/use case, training/fine-tuning/inference, precision, context/concurrency, dataset and target performance. State the security domain and operational date. Attach a benchmark or define how the government will create one.

Name the buying unit: “four factory-integrated eight-GPU servers,” “one deployable rack” or “one scalable unit with compute, fabric and storage.” List inclusions and exclusions. If the rack depends on government-furnished spine switches or storage, identify the handoff and compatibility evidence.

Set quantities, options and growth. Separate base award from optional nodes, switches, storage, spares or support years. Explain whether options must preserve fabric/rack compatibility and price basis.

Avoid conflicting requirements. A specification may combine the memory of B300, dimensions of an H200 server and air cooling in a way no supported product can meet. Conduct market research and an engineering scrub before release.

FAR 52.211-6 tells offerors that an equal product must meet the solicitation's stated salient physical, functional or performance characteristics and identifies descriptive materials they may need to provide. The buyer must therefore state what characteristics make the reference product suitable.

Do not list every brochure attribute as salient. Select characteristics connected to mission, compatibility, facility, support or schedule. Mark each as mandatory, threshold, objective or informational. If only the named brand can satisfy the requirement, follow applicable justification and approval procedures rather than writing impossible “equal” criteria.

For orders under IDIQ contracts, FAR 16.505 addresses fair opportunity and brand-name requirements. Vehicle-specific ordering guides can add procedures. The contracting officer determines what applies.

Require offerors proposing an equal to identify manufacturer/model, map each salient characteristic, provide source documentation and disclose deviations. Do not make “equal” a post-award substitution process.

Organize characteristics by outcome and interface.

Accelerator/platform: minimum supported GPU generation or performance, accelerator count, usable memory, memory/error-protection features, precision support, interconnect topology and partitioning where required.

Host: CPU architecture/count, system memory, PCIe topology, boot and local scratch, secure boot/root of trust, management controller and serviceability.

Network: compute-fabric adapters, port speed/count, protocol/validated platform, storage/data ports, management ports, optics/cables and topology. State whether switches are included.

Storage: usable capacity, endurance, throughput/metadata target, filesystem/protocol, data protection and GDS compatibility if required.

Software: operating system, driver/CUDA range, container/orchestration, NVIDIA AI Enterprise or equivalent entitlement, management/monitoring integration and license/support term.

Reliability/support: redundant components, hot-swap requirements, warranty, response, repair unit, firmware/vulnerability process and spares.

Tie values to workload or system interfaces. If a minimum can be met through multiple architectures, allow the offeror to explain and prove its method.

Include maximum dimensions, rack units, weight, rails and service clearances. State rack standard and delivery path. For a populated rack, add total weight, floor loading, center-of-gravity/transport and seismic requirements.

Electrical requirements should state available voltage/phase/frequency, feed count, connector/receptacle, redundancy and maximum allowable rack load. Ask for typical and maximum configured draw and a device power schedule.

Cooling should state available air or liquid interface, inlet/environmental limits and site responsibility. For liquid cooling, include technology-water temperature, flow, pressure, quality, CDU/manifold ownership, quick disconnects, leak detection and commissioning. Reference the facility readiness guide.

Define data-center handoffs: spine ports, storage endpoints, management VLAN/VRF, identity, time, DNS, logging and security tools. Require a port/cable matrix and rack elevation. “Compatible with existing network” is not testable without named interfaces.

List solicitation-specific clauses and representations through the contracting package. Require configuration-level evidence appropriate to TAA applicability, Section 889, prohibited sources, country of origin, OEM/support channel and software licenses. Do not ask offerors to certify a vague “NDAA compliance” concept without identifying the relevant requirement.

Request manufacturer data sheets, support-matrix links and an exact BOM. Require notification and approval before substitution. A change in drive, NIC, switch or embedded management component can affect performance, firmware, origin evidence and delivery.

Specify security deliverables: firmware/software manifest, secure configuration, vulnerability notification, update/rollback, administrative access, log integration, media handling and sanitization. Identify any personnel, site-access or tool/media restrictions for installation and support.

State data rights and licensing. Confirm government use rights, transferability, offline operation if needed, renewal pricing basis and behavior after expiration. Separate perpetual components from subscriptions.

Set required delivery and operational dates, destination, shipping condition, secure logistics, staging, asset tagging, installation, configuration, documentation and training. Ask for lead time by critical component and quote validity.

Factory acceptance can include BOM/serial inspection, firmware baseline, burn-in, GPU health, fabric and storage tests, redundant power, telemetry and representative workload. Site acceptance adds shipping inspection, facility connections, uplinks, identity/logging, security configuration, failover, workload performance and documentation handoff.

Every test needs method, tool/version, duration, threshold, evidence and remediation. Distinguish “system powers on” from “cluster meets mission benchmark.” Define who owns a failed interface and how retest affects payment/acceptance.

Require as-built BOM, rack elevation, power schedule, cable matrix, firmware/software manifest, licenses, test results, approved deviations, runbooks and support contacts. Tie payment milestones to meaningful deliverables where the acquisition permits.

Give respondents a template:

Normalize quotes at the same boundary. Clarify ambiguities equally. Do not reward a source for omitting required fabric, optics or services.

Watch for requirements that specify a GPU memory number but omit usable system memory, name a server but omit rails and rack depth, require redundant power without defining independent feeds, or call for “liquid cooling” without a CDU and facility-water boundary. Another frequent error is requiring a fabric speed while leaving switch topology and cables to “be determined.”

Commercial terms can also undermine the technical work. A quote-validity period shorter than the evaluation schedule, undefined annual software renewals, unrestricted substitutions or delivery defined only as arrival at the dock can move major cost and risk after award. Put those items in the response template.

Finally, avoid untestable adjectives: state the threshold for “high performance,” the recovery objective for “resilient,” the evidence for “secure” and the exact clause/context for “compliant.” If the team cannot describe how receiving will verify a requirement, revise it before release.

Run a red-team review with engineering, facilities, security, users, receiving and contracting. Ask each to identify one failure that the draft would allow. Revise before release.

Uniqcli can review a draft line item against AI rack integration, current OEM configurations, the H300 nomenclature check, facility interfaces and acceptance. Request a GPU RFQ red-team review.

The useful output is a marked requirement with ambiguous units, missing interfaces, unsupported combinations, evidence gaps and proposed test language—not a rewrite designed to force one brand.

Acquisition note: Contracting officials must determine applicable competition, justification, clauses and evaluation. This is not legal advice.