TL;DR: The purchase cannot be commissioned until each rack's electrical, cooling, structural and connectivity requirements are satisfied. Compare the ordered configuration with the site, then present feasible retrofit, relocation or replacement options with verified costs and dates.
How to approach it
Obtain the exact rack configuration and site-planning requirements before using a headline power number. Ask the facility team to assess the proposed positions, including delivery and maintenance access. Explain which constraint blocks installation and what evidence would make an alternative acceptable.
A strong answer
An NVL72 rack holds 72 GPUs joined by one NVLink domain, so it is bought, powered and cooled as one unit rather than as separate servers. A hall may have enough total megawatts while its individual cabinet positions cannot supply a dense rack. That distinction is the starting point for the conversation. Leadership needs a deployment decision with an evidence-backed schedule, not reassurance based on unused building capacity.
The NVIDIA DGX GB rack-scale guide covers GB200 and GB300 systems. Use the site specification for the actual ordered rack and its integrator. Rack input power, cooling interfaces, weight and cable options are configuration-specific; a figure from another product or reference design may not apply.
For an illustrative electrical comparison, assume the ordered rack requires a 140 kW design budget while the proposed position supports 15 kW. Four such racks require 560 kW in total. A 2 MW building might cover that aggregate load, but the per-position budget is still short by a factor of 140/15 ≈ 9.33. Spare capacity elsewhere does not automatically reach the rack through the installed feeds.
Illustrative balanced three-phase load:
P = 140,000 W; line-to-line voltage = 415 V; power factor = 0.99
I = P / (sqrt(3) × V × PF) ≈ 197 A per phase
This is a load calculation, not a breaker or conductor specification. Qualified facility engineers must account for the actual distribution, protection, continuous-load rules, redundancy, transients and applicable requirements. If either A or B feed must carry the full load after failure, each path needs that capability. That does not mean the equipment consumes twice the power in normal operation.
| Check | Evidence required |
|---|---|
| Electrical delivery | Rated and available capacity through the entire path to each rack |
| Cooling | Required liquid and air interfaces, flow, temperature, heat rejection and failure behavior |
| Structure and access | Loaded weight, point loads, route from dock, clearances and service procedure |
| Connectivity | Qualified cable types, lengths, ports and routing for the selected topology |
Power telemetry shows what a rack draws; it does not establish a breaker's rating or spare capacity under a failure. Obtain the electrical drawings and equipment ratings as well as measurements. Similarly, a room cooling total does not prove that the rack receives the flow and temperatures its system requires. Direct-to-Chip Liquid Cooling and CDUs explains the heat-transfer path.
Do not offer a rear-door heat exchanger as a drop-in replacement for a system's required direct liquid cooling. It addresses heat in the air stream and does not itself supply the cold-plate loop. Any alternative cooling arrangement needs explicit vendor support and facility approval. Avoid universal cable reach numbers too: medium, speed, cable assembly and transceiver qualification determine reach.
The practical alternatives depend on the site assessment:
| Option | What to compare |
|---|---|
| Retrofit selected positions | Construction scope, outage needs, approvals and commissioning date |
| Use another qualified site | Available capacity, contract, transport and network dependencies |
| Change the ordered configuration | Supplier terms, workload fit and revised delivery date |
| Use temporary capacity | Available compatible service, migration work and full rental cost |
Obtain quotes rather than inventing standard lead times. For scale, four 72-GPU racks contain 288 GPUs. At an assumed comparison rate of $2.50 per GPU-hour, 182 days represents 288 × 24 × 182 × 2.50 = $3,144,960 of rental-equivalent capacity. This is not automatically an accounting loss from idle owned hardware, and it is not a quote for equivalent rental capacity. The value of earlier work, capital cost, contractual obligations and migration effort decide whether a bridge is worthwhile.
Recommend the option that reaches a supported, accepted deployment with the best fit to the workload and deadline. A previously prepared high-density zone may make a retrofit attractive. If the site needs major construction and a qualified alternative is available, relocation may be faster. Either recommendation changes when the engineering assessment or supplier commitments change.
What interviewers probe next
- "Isn't the building's 2 MW enough?" Check both aggregate capacity and the complete delivery path to each rack.
- "Can we split the rack?" Do not assume the qualified rack-scale system can be rearranged; ask the vendor for a supported design.
- "Can power telemetry certify the feed?" No. Ratings, protection and failure-state capacity require separate checks.
- "Who should have caught this?" Add a site-readiness acceptance gate before future hardware commitments.
Common mistakes
- Using a headline rack number instead of the ordered configuration.
- Treating redundant capacity as double normal energy consumption.
- Substituting a different cooling method without system qualification.
- Presenting hypothetical rental value as a measured financial loss.
Key takeaways
- Four hypothetical 140 kW racks need 560 kW, with each position able to deliver its share.
- The example current is about 197 A per phase before engineering sizing rules.
- Select an option from verified facility requirements, quotes and commissioning dates.