The cheapest problems to fix are the ones you catch before the rack arrives. This domain is about confirming the room can actually take a 120 kW, 1.4-tonne liquid-cooled AI rack — power, cooling, floor, space, and pathways — before anything ships.
An AI rack is unforgiving of a bad site. Discover mid-install that the floor can't bear the load, the cooling loop can't reject the heat, or the door is too narrow, and you've stranded millions of dollars of hardware. Domain 2 tests whether you can verify the environment is ready before commitment — the highest-leverage work in the whole deployment.
Site assessment answers one question five ways: can this room actually host this rack? The five readiness pillars are power, cooling, structural (floor loading), space/access, and network pathways. Miss any one and the deployment stalls.
A single GB200 NVL72 draws roughly 120 kW — more than an entire traditional rack row. Planning power means confirming the facility can deliver that continuously, with redundancy, and with headroom.
Confirm the feed and upstream distribution can carry the full rack load plus margin. Never plan to run at 100% of a breaker's rating — leave headroom for inrush and growth.
Dual independent feeds so loss of one path doesn't drop the rack. Verify the redundancy scheme (N+1, 2N) the facility actually provides.
Size to real power (kW) and apparent power (kVA), and balance load across phases. Mismatches trip breakers and waste capacity.
Confirm the rack's DC busbar/power-shelf configuration matches the delivered voltage and the whip/feed type at the rack position.
Rated capacity ≠ usable capacity. Continuous loads are typically planned to a fraction of the breaker rating, and redundant (2N) designs mean each path must independently carry the full load. Plan to the usable, redundant figure, not the nameplate.
At 120 kW you're rejecting heat through liquid, so the site's facility water loop — not just its CRAC units — is what you assess.
| Check | What to verify | Why |
|---|---|---|
| Heat rejection capacity | Facility loop can absorb the rack's full thermal load (plus other racks) | The CDU only moves heat to the facility loop; that loop must actually reject it |
| Supply water temperature | Facility supply-water temp fits the equipment's class (ASHRAE water classes W17–W45) | Too-warm water starves the cold plates; too-cold risks condensation |
| Flow rate & pressure | Adequate GPM and pressure at the manifold connection | Insufficient flow means the cold plates can't carry heat away fast enough |
| Water quality & connections | Fluid type/treatment and quick-disconnect compatibility | Wrong fluid or fittings cause fouling, leaks, or corrosion |
| Residual air heat | Rear-door heat exchangers / room cooling for the non-liquid portion | Not 100% of heat goes to liquid; the balance still needs air handling |
"The CDU moves heat; the facility rejects it." Your job on assessment is to confirm the facility side can actually take the heat the CDU will hand it.
A fully populated NVL72 weighs well over 1,300 kg (roughly 3,000 lb) concentrated on a small footprint. Structural and access checks are where deployments physically get stuck.
Verify the floor's rated load (point load and distributed load) supports the rack. Raised floors especially have limits; you may need load-spreading or a slab position. Include seismic bracing where required.
Trace the whole route: dock, door widths, elevator capacity, ramp slopes, aisle turns, ceiling height. A rack that won't fit through the door is a classic, avoidable failure.
Confirm front/rear service clearances for tray insertion, cabling, and CDU maintenance, plus hot/cold aisle spacing.
Floor loading is the most underestimated check. These racks are far heavier than legacy gear; a raised-access floor rated for older equipment may not carry a modern AI rack's point load. Verify the number — don't assume.
Finally, confirm the cabling pathways: fiber routes to the spine/MDA, tray/ladder capacity for the bundle volume, and where scale-out uplinks land. Under-planned pathways force messy, out-of-standard cabling later (which hurts you in Domains 4 and 5).
Boil the five pillars into a go/no-go checklist you can run per rack position:
Site-assessment questions reward precise thinking about capacity, redundancy, and limits. Rehearse them with explained answers.
Practice Pre-Deployment Planning →NCP-ARI-style questions on Domain 2. Choose an answer to see the explanation.
Pick the readiness area you're least sure about for a targeted study focus.
Tap a card to flip it.
For educational use only · Not affiliated with or endorsed by NVIDIA Corporation. Rack power/weight and ASHRAE water-class figures are drawn from published NVIDIA and industry materials and can change; confirm current values on the GB200 NVL72 page and applicable facility standards before relying on them.
