Why 130 kW Racks Change Datacenter Design

For twenty years a datacenter rack drew somewhere between 5 and 15 kW, and the whole building was designed around that number: the raised floor, the CRAH units at the end of each row, the 208 V PDUs, the slab. A GB300 NVL72 rack draws around 132 kW. That is not an incremental change. It is a different building.

This note walks through what actually has to change when the rack density goes up by an order of magnitude, because the mistakes we see most often come from treating a liquid-cooled GPU row as a slightly hotter version of what the facility already had.

Air Stops Working at About 50 kW

Air cooling is limited by how much air you can move through a rack and how much heat each cubic metre can carry. Rear-door heat exchangers extend the range to roughly 40–60 kW per rack by cooling the exhaust at the door, but above that the fans alone would consume a meaningful share of the rack's power and the airflow would become impractical. Direct-to-chip liquid cooling moves the heat off the GPU, CPU and switch ASICs through cold plates into a secondary water loop, and that loop can carry 100 kW or more per rack without difficulty.

The practical consequence is that there is no 'air option' at NVL72 density. If the design calls for these racks, it calls for a coolant distribution unit, manifolds, quick-disconnects at each node and a facility water loop to reject the heat. Those have to be designed before the racks are ordered, not after they arrive.

The Power Train Changes Shape

A 132 kW rack at 415 V three-phase is roughly 185 A. Rack PDUs at that scale become power shelves — OCP-style units that feed a busbar running down the back of the rack — and the overhead busway feeding them has to be sized for a row of them, often 1,000–2,000 A per row. UPS sizing follows: a 16-rack row is over 2 MW of IT load, so N+1 UPS modules, generator capacity and the utility interconnect all move together.

This is also why 800 VDC distribution is coming. At 132 kW and up, converting AC to DC inside every rack wastes space and efficiency; the next generation of racks is designed to take high-voltage DC from a row-level rectifier. Building the busway and clearances to be 800 VDC-ready today costs little and avoids a second fit-out in two years.

The Floor Has to Hold It

A loaded NVL72 rack weighs around 1,600 kg, and liquid-cooled racks concentrate that weight on four points. Many raised floors are rated for 1,000–1,250 kg per tile and older slabs were never checked for point loads at this level. A structural review is a one-week piece of work that, skipped, becomes a six-month problem.

Commissioning Is a Different Discipline

Commissioning a liquid-cooled row means pressure-testing and flushing the loop, filling with the correct coolant chemistry, balancing flow to each rack, verifying leak detection and automatic isolation, and then thermal-mapping the row under synthetic load before a single real job runs. It overlaps with the electrical and network commissioning that teams already know, but it is new work, and it is the step most often compressed when a schedule slips.

Our rule: if the commissioning plan does not include flow balancing and a loaded thermal map, the facility is not ready, whatever the calendar says.

What to Do with This

If you are planning GPU rows in an existing facility, start with a retrofit assessment: structural, electrical and thermal capacity against the rack you actually intend to buy. If you are building new, design the cooling plant and power train around the rack density first and let the building follow. Either way, the rack has become the unit of design, and everything else is sized from it.