Start with the actual hardware
Request versioned manufacturer documentation for the proposed server and rack configuration, including a GB200 deployment when relevant. Do not infer temperature, flow, pressure, coolant or connector requirements from the GPU family name alone. Preserve the document revision and the intended configuration.
Different architectures require different evidence
Direct-to-chip, immersion, rear-door and air systems have different interfaces and operating arrangements. Ask how heat moves from the hardware through the facility and out to heat rejection. Confirm residual air loads, water requirements, redundancy and ownership of each interface with the appointed engineer.
Make retrofit uncertainty visible
A facility may require additional equipment, structural works or commissioning changes. Until a scoped design and quotation exist, keep retrofit costs unresolved. Missing evidence is not proof of incompatibility, and a matching rack-density number is not system certification.
Take this checklist into the next meeting
- Obtain versioned rack and server requirements.
- Match architecture and physical interfaces.
- Confirm heat rejection, residual air, water and resilience.
- Assign retrofit scope and integrated commissioning ownership.
Evidence and assumptions
This guide develops the supplied buyer-side research framework. It makes no site-specific capacity claim. Obtain current utility documents, manufacturer specifications and appointed professional advice for the actual project. Any numerical examples are worked examples, not project data.
Source starting points: NVIDIA documentation and Vertiv cooling information. General pages do not replace a versioned project specification.