Leadership bought GB300 NVL72 racks. Your datacenter is built for 15 kW cabinets. What do you tell them?

TL;DR: Four limits, checked separately. Power per rack: NVIDIA quotes roughly 120 kW for a GB300 NVL72 and integrator deployment guides put commissioned racks at 132 to 142 kW nominal with peaks near 155 kW, against a hall designed for 15 kW, so one rack needs the electrical capacity of eight to ten existing cabinets and a busway tap that does not exist. Heat rejection: at that density the cooling is liquid and mandatory, so the hall needs facility water to the row, which most air-cooled halls do not have and cannot easily be given while operating. Floor loading: these racks are heavy and concentrated, and raised floors frequently cannot carry them. Network reach: an NVLink domain is one rack, but the scale-out fabric still has to reach spines, and cable reach at 800G is short. The answer is not "no". It is three costed options: retrofit part of the hall, move to a site built for this, or change the hardware to something the facility can take, each with a lead time and a number.

How to approach it

Separate the four limits and check each, because they fail independently and a fix for one does not help another. Give the numbers per rack rather than for the building. Then convert into options with costs and lead times, because a refusal with no alternative gets escalated around you. Say which option you recommend and what would change it. Do not lead with the problem; lead with what is possible.

A strong answer

A typical situation: an order is placed for four GB300 NVL72 racks with a delivery date in eleven weeks. The company's hall is a well-run air-cooled facility with 2 MW of total capacity and cabinets provisioned at 15 kW. Total capacity looks sufficient, which is why nobody flagged it.

The four checks, each with its own arithmetic:

1. power per rack
   required: about 120 kW by NVIDIA's figure, 132 to 142 kW nominal per integrator guides,
     peaks near 155 kW
   available: 15 kW per cabinet
   the gap is per-rack, not total: 4 racks x 140 kW = 560 kW of a 2 MW hall, which the
     building can supply and no single cabinet position can deliver
   current at 415 V three-phase: 140,000 / (1.732 x 415 x 0.99) = 197 A per rack
     with A and B feeds, roughly 394 A reserved per rack position
   remedy: new busway with high-current taps to those positions. Months.

2. heat rejection
   140 kW per rack x 3,412 = 477,680 BTU/hr per rack
   air cooling at this density does not exist in a normal cabinet
   required: facility water to the row, a CDU per rack or per few racks sized 10 to 15%
     above load, manifolds, leak detection
   remedy: a facility water loop into an operating hall. Construction, months, and often
     restricted by the building's lease or design.

3. floor loading
   a populated rack of this class is well over a tonne concentrated in one footprint
   raised floors are commonly rated far below that as a point load
   remedy: slab placement, reinforced pedestals, or a different area of the building.
     Weeks to months, and sometimes impossible in the space available.

4. network reach
   the NVLink domain is inside the rack, so that part is fine
   the scale-out fabric still runs 800 Gb/s per GPU to leaf switches
   passive copper at 800G reaches about 2 m, where 400G reached 3 m
   remedy: place leaves in-row and use optics beyond, which adds transceivers and power
sanity: three of the four remedies are measured in months and the hardware arrives in eleven
        weeks, which is the sentence that has to be said in the first meeting rather than the
        third

The site assessment that should have happened, expressed as things you can actually read:

measurements to take before any order, not after
  metered PDU output per cabinet position, from the PDU's own SNMP or web interface
    gives the real breaker rating and the headroom, not the design document's number
  the busway tap rating stamped on the tap, per position
  facility water: is there a supply and return in the row, and at what temperature
    a hall with no pipe in the row has no liquid answer on any hardware timeline
  floor loading rating, from the building's structural documentation, as a point load
  distance from the intended rack position to the nearest switch position, in metres
    this is what decides passive copper against optics at 800 Gb/s

and once hardware is in, the per-GPU numbers that confirm the model
  nvidia-smi -q -d POWER            enforced limit and instantaneous draw per GPU
  DCGM_FI_DEV_POWER_USAGE           the same, as a fleet metric to sum per rack
  ipmitool sensor / redfish power   node-level draw including CPUs, NICs and fans, which is
                                    what the busway actually sees
sanity: summing DCGM_FI_DEV_POWER_USAGE across a rack and comparing it against the metered
        PDU reading is the check that the rack is drawing what the design said, and the gap
        between them is everything that is not a GPU

Colocation, Power Contracts and Site Selection covers why per-rack density eliminates sites that pass on total capacity. Direct-to-Chip Liquid Cooling and CDUs covers the second row, including why facility water runs warm at 30 to 40 degrees and what the loop needs.

The options, costed, which is what the meeting actually needs:

what the delay costs, which is the number that decides
  four NVL72 racks is 288 GPUs
  idle for six months at $2.5 per GPU-hour:
    288 x 24 x 182 x 2.5 = about $3.15M of unrealized capacity
  renting equivalent capacity for those six months costs more per GPU-hour and delivers work
sanity: the rental bridge is usually the recommendation, because it converts a six-month
        write-off into a premium on capacity that is actually producing, and it lets the
        retrofit proceed without a delivery date on top of it

The reversal condition: if the hall has a spare zone that was already built for higher density, or if the building has an unused water loop from a previous tenant, the retrofit collapses from a year to a couple of months and becomes the obvious answer. It is worth checking the facility's history before assuming construction, because halls are often built with more capability than their current provisioning uses. The other reversal is scale: for one rack rather than four, a self-contained rear-door heat exchanger and a dedicated feed can sometimes be arranged inside an existing hall, which is not viable for a row.

What interviewers probe next

• "Isn't 2 MW enough?" Total and per-rack are different numbers. The building can supply it and no cabinet position can deliver it, and only the second one matters.
• "Could you spread the load across more cabinets?" No. An NVL72 is one physical rack with an internal copper NVLink backplane; it cannot be split across cabinet positions.
• "What about air cooling with more airflow?" Not at this density in a standard cabinet. Rear-door heat exchangers reach roughly 40 to 70 kW, which is still far short.
• "Who should have caught this?" A pre-purchase site assessment. The useful answer names the process change rather than the person.

Common mistakes

• Comparing the order against total building capacity instead of per-rack capacity.
• Checking power and stopping, when cooling, floor loading and cable reach each fail independently.
• Answering with a refusal instead of costed options with lead times.
• Assuming a facility water loop can be added to an operating hall on a hardware delivery schedule.
• Forgetting that the rack cannot be split, since the NVLink backplane is physical.

Key takeaways

• Per-rack capacity is the binding limit: about 120 kW by NVIDIA's figure and 132 to 142 kW nominal per integrator guides, against 15 kW cabinets.
• Four independent checks: power per rack, heat rejection, floor loading, and network reach at 800G where passive copper stops near 2 m.
• A 140 kW rack draws about 197 A at 415 V three-phase and reserves roughly 394 A with redundant feeds.
• Three of the four remedies take months, so the schedule conversation happens in the first meeting.
• Six months of 288 idle GPUs is about $3.15M, which is why a rental bridge usually beats waiting.