High-Density GPU Racks: What 120 kW+ Racks Ask of a Data Center Site

Rack density is the electrical power drawn by the IT equipment in one cabinet. For most of the industry it is still modest. Uptime Institute’s 2025 global survey found densities rising, with more racks in the 10–30 kW range, but few facilities exceeding 30 kW.³ Its 2026 survey found a growing number of operators reporting peak racks of 30 kW or more.¹⁰ AFCOM’s 2026 State of the Data Center report put the average among its respondents at 27 kW, up from 16 kW a year earlier, though its members are mostly traditional facilities rather than AI campuses.¹¹

AI systems sit far above those averages. SemiAnalysis puts NVIDIA’s GB200 NVL72 rack, which links 72 GPUs in one liquid-cooled cabinet, at about 120 kW, and notes that air-cooled H100 racks typically top out near 40 kW.¹ Gigabyte lists its GB300 NVL72 rack at up to 140 kW, fed by six 33 kW power shelves.² NVIDIA has also told the industry to prepare for racks of up to 1 MW.¹²

For long-run trends and forecasts, see our guide to data center rack density trends. This guide focuses on what dense racks change on the land and in the utility request.

Air can only carry so much heat through a cabinet before fan power, noise and airflow become impractical. One engineering firm puts the practical limit at about 30 to 40 kW per rack, and SemiAnalysis ties the GB200’s liquid cooling requirement directly to its 120 kW density.⁴¹ In a dense GPU hall, liquid cooling is the baseline, not an option. That decision is made by the hardware, not the site, so a site that cannot support liquid-cooled buildings is effectively closed to current AI deployments.

Most GPU racks today use direct-to-chip cooling: cold plates on the processors, a coolant distribution unit (CDU) that separates the rack loop from the building’s facility water loop, and heat rejection outside. NVIDIA describes its Vera Rubin rack as using 45°C liquid cooling, and its CEO has called the rack 100% liquid-cooled.⁶¹³ Our guide to air vs. liquid cooling compares the systems in more detail; temperature limits for the equipment are set through ASHRAE TC 9.9 guidance and the IT vendor’s specifications.

Warm-water liquid cooling changes the outdoor plant more than the indoor one. At CES in January 2026, NVIDIA’s CEO said that with 45°C coolant a Vera Rubin data center does not need a chiller, which would remove equipment that costs power, space, noise and often water.¹³ That claim should be tested, not assumed: LiquidStack’s CEO, quoted in the same report, said chillers could still play a role in some halls because of secondary loop temperatures.¹³

For site screening, three consequences follow.

• Climate matters differently. With warm supply water, dry coolers can cover more hours of the year in more climates, which can reduce or remove evaporative water use. Hot design days still set the size of the plant.
• Yard space grows with density. More heat per building means more heat rejection equipment per building, often on the ground beside it or on the roof. Our guide to cooling plant and equipment yard space covers sizing.
• Water depends on design. A closed loop needs a fill and small top-ups; evaporative heat rejection uses water every day. Ask for the design’s water usage effectiveness (WUE), and see closed-loop and waterless cooling.

Dense racks push electrical design toward higher voltages closer to the IT equipment. NVIDIA is moving its partners from conventional 415 or 480 VAC three-phase distribution toward 800 VDC, citing efficiency, scalability and less material.⁶ DCD reported NVIDIA’s claim that 800 VDC can carry over 150% more power through the same copper, avoiding 200-kg copper busbars to feed a single rack.¹² More than 20 partners showed 800 VDC equipment at the 2025 OCP Global Summit.⁶

The change also reaches the grid. Large AI loads can change their power draw quickly, and NVIDIA says its Vera Rubin rack carries 20 times more energy storage to keep power steady.⁶ Grid operators are treating large-load behavior as a reliability issue: NERC’s board voted on April 16, 2026, to issue a rare Level 3 alert on large loads, which went out in May, citing customer-initiated load reductions and oscillations that happen within seconds.⁹ Expect utilities to ask about load swings and ride-through during interconnection; see NERC reliability and large loads and UPS and electrical distribution.

Higher density means the same IT load fits in far fewer racks and far less floor area. That does not shrink the site as much as it might seem, because substations, cooling yards, generators and setbacks do not shrink with the racks. What it does is raise the megawatts a utility must deliver to each building and each acre.

Demand is growing fastest exactly where density is highest. The IEA reports that global data center electricity use grew 17% in 2025 while use by AI-focused data centers grew about 50%, and it projects total data center consumption roughly doubling from about 485 TWh in 2025 to about 950 TWh in 2030.⁵¹⁴ For land, the practical result is that a compact parcel can now hold a building whose power need once required a campus. Whether the parcel works depends on how much power the data center needs and how fast the utility can deliver it. At the largest end, see what a gigawatt-scale AI campus requires.

Dense racks are heavy. TechRadar reported a DGX GB200 NVL72 at about 1.36 metric tons, or 3,000 lb.⁷ One engineering firm notes that a loaded GPU rack concentrates about 3,000 lb in under 7 square feet, nearly double the point load a standard raised-floor panel handles, and that a flooded CDU can reach 6,600 lb.⁴ AMD’s data center design guide says its 72-GPU and larger racks need special floor reinforcement and may not match the standard 19-inch footprint.⁸

That pushes AI halls toward slab-on-grade construction with overhead power busways and coolant manifolds. AMD notes that routing large coolant pipes under a raised floor can be cumbersome and leak-prone, and that slabs suit heavy racks and immersion tanks.⁸ For land, slab-on-grade single-story buildings favor flat sites with good bearing soils; a geotechnical investigation belongs early in diligence. Our guide to slope, soils and geotechnical review covers what to test.

Existing buildings are the hard case. A legacy raised-floor facility may have the utility service for a few dense racks but not the floor capacity, piping routes or heat rejection to convert at scale, which is why many AI projects favor new construction on open land.

1. 01Confirm the design density and rack generation the buyer or tenant expects, and whether the hall must support later, denser racks.
2. 02Translate density into megawatts per building, then test that against the utility’s deliverable capacity and schedule.
3. 03Model heat rejection for both dry and evaporative designs on local design-day temperatures, and confirm water service for the wetter case.
4. 04Reserve yard space for cooling plant beside each building, plus substations and backup power.
5. 05Order a geotechnical review early: slab-on-grade halls with heavy racks need good bearing soils.
6. 06Ask the utility what it will require on load ramping, swings and ride-through for an AI load.

None of these replaces engineering by the design team, but each can screen out a site before money is spent on options, studies and deposits. The order matters: power and cooling questions usually decide the outcome sooner than zoning or price, so ask them first. You can get a site reviewed if you want an outside screen of power, cooling and land together.