AI Training Data Centers — NVL72-Class Rack Systems

A GB200/GB300 NVL72 cabinet draws about 132 kW continuously. The hall it lands in was most likely designed for 10–30 kW per rack — older floors for 10–12 kW. That is not a gap, it is an order of magnitude, and it breaks power architecture, structure and cooling at the same time. Air cooling gives up near 40 kW a cabinet; beyond that, direct-to-chip liquid is mandatory. Retrofitting the utility side takes 12–24 months at over $2M per MW in major markets, and each CDU adds $20,000–35,000. RackForge builds the rack boundary so the cluster is not waiting on it: liquid-ready frames, busbar-slot columns and contained aisles, specified against the NVL72 envelope.

Request a Quotation See the Gap in Numbers

132 kW per cabinet · 1,600 kg frames at 4× · liquid interface ready in the crate

The gap, side by side

Generic hall, 10–30 kW era vs AI-ready rack row

Eight requirements decide whether a GPU cluster reaches first token on schedule. The left column is how the installed base was built; the right column is the RackForge row specification.

Requirement	Generic hall, 10–30 kW era	AI-ready rack row
Per-cabinet power	No Designed around 10–30 kW racks; older floors at 10–12 kW	Yes 132 kW NVL72-class liquid builds; air-cooled GPU builds to 40 kW
Frame static load	No 1,000–1,363 kg catalog class, sample-tested at best	Yes 1,600 kg static, verified at 4× with a 72-hour creep recheck
Power feed	No Rack PDUs on branch circuits, no growth path per position	Yes Busbar-slot columns to a 400 kW per rack-position design ceiling
Heat removal	No Room air and CRAC — the path saturates near 40 kW a cabinet	Yes Manifolds and UQD pairs pre-installed, proven at 1.5× for 24 h
Aisle boundary	No Open aisles — 25–50% of cold air bypasses the load	Yes CAC on the 600 mm module; leakage 3–10%, under 1% with full sealing
Upgrade path	No Utility uprating 12–24 months, over $2M per MW in major markets	Yes Provision built into the columns — add capacity per row, not per hall
Liquid plant entry	No CDU sourced late at $20–35k per unit, interface improvised on site	Yes Rack-side UQD geometry ready before the CDU arrives on the dock
Delivery form	No Assemble, integrate and cable on a live floor for 2–3 weeks	Yes Pre-integrated and crated — the row lands as one SKU, one test record

Row of liquid-cooled GPU training cabinets behind glass cold-aisle doors with overhead busbar, cold white hall

The package

One row, three systems, delivered as one drawing

The rack is only the anchor point. What makes a GPU row converge is that the power path, the coolant path and the airflow boundary arrive as a single engineered set, aligned to the NVIDIA GB200/GB300 NVL72 reference cabinet of 600 × 1068 × 2236 mm.

Liquid boundary: manifolds above both door lines, UQD pairs at fixed coordinates, drip pans under every union — the loop holds 1.5× working pressure for 24 hours before dispatch.
Busbar feed: both uprights carry machined tap-off slots to the 400 kW per rack-position ceiling, so feed tracks the cluster instead of the floor plan.
NVLink bracing, optional: a diagonal brace set stiffens the bay against NVL72-class tray and switch masses — ordered with the rack, fitted at assembly.
Contained aisles: CAC doors, roof panels and sealing kits share the rack's 600 mm module, holding contained leakage to 3–10%.

Specifications

The row specification, air-cooled and liquid side by side

Example SKUs on a common 48U, 800 × 1200 mm frame: the air-cooled build carries GPU rows to the 40 kW air limit; the liquid build spans the 40–132 kW NVL72-class envelope. Both share columns, load rating and containment interfaces.

Parameter	RF-GP-48U · air-cooled	RF-LC-48U · liquid
Static load	1,600 kg	1,600 kg
Load verification	4× safety factor · static hold 72 h · creep recheck before release
Power envelope per cabinet	air-cooled to 40 kW	40–132 kW, NVL72-class
Width × depth	800 × 1200 mm · 600 mm builds quoted for NVL72 alignment
Height	2265 mm / 89.2 in (48U)
Columns	SPCC 2.0 mm, five-bend reinforcement, ±0.1 mm forming
Busbar slot	both columns · ≤400 kW per rack-position design ceiling
Power-shelf rails	pre-punched to GB200/GB300 recommended positions
NVLink braced frame	optional diagonal brace set, factory-fitted
Manifold & UQD	—	above-door routing · 25–80 L/min · PG25
Pressure proof	—	1.5× working pressure · 24 h · zero-drop record
Doors	front mesh ≥80% open area · glass front option
Grounding continuity	≤0.1 Ω frame-to-frame, jumpers included
Standards	EIA-310-E · IEC 60297 · CE (EN IEC 62368-1) · RoHS · OCP ORv3 compatible

The line items behind the row

Each system in the package has its own engineering page with full specifications and factory test records.

GPU High-Density Racks

The 1,600 kg frame with busbar-slot columns and pre-punched power-shelf rails, engineered against the GB200/GB300 NVL72 footprint and OCP ORv3.

40–132 kW per cabinet envelope
4× load test, 72-hour creep recheck
NVLink braced frame option

GPU high-density racks →

Liquid-Cooling Ready Racks

Manifolds, UQD quick disconnects and drip pans mounted at the factory, pressure-held at 1.5× working pressure for 24 hours before the crate closes.

25–80 L/min design flow on PG25
Zero-drop pressure record ships inside
CDU docking per OCP-aligned practice

Liquid-cooling ready racks →

Busway-Integrated Racks

Uprights engineered as busway tap-off stations with plug-in box brackets pre-mounted, on a 250–800 A track — power arrives per position, not per electrician.

250–800 A tap-off track classes
Grounding continuity ≤0.1 Ω
Top feed or under-floor feed

Busway-integrated racks →

The cluster ships on schedule. The row is already proven.

132 kW NVL72-class · 1,600 kg at 4× · 1.5×/24 h pressure record · MOQ 5

Request a Quotation sales@ems-rack.com