Rack Power Density Trend: 8 kW Average to 132 kW GB200 Racks

NVIDIA’s own reference architecture puts the GB200 NVL72 rack at 120 kW. HPE’s data sheet for the same rack lists 132 kW nominal, with a peak electrical design provision near 192 kW. Meanwhile the global average rack, across the installed base that Uptime Institute surveys every year, sat at 2.4 kW in 2011 and 8.4 kW in 2020, and recent surveys still put the operator-wide figure in the high single digits. Those are not competing claims about the same number. They describe different things, and the gap between them is the actual story behind “rack power density.”

Two definitions of rack density produce two different curves

The confusion in most coverage of this topic comes from treating “average rack density” and “flagship rack specification” as the same metric measured at different points in time. They are not. One tracks the installed base of racks already running in data halls worldwide. The other tracks a single, top-of-line product a vendor is currently shipping.

Isazadeh, Ziviani, and Claridge, writing in Renewable and Sustainable Energy Reviews (2023), report that global average rack power density rose from 2.4 kW per rack in 2011 to 8.4 kW per rack in 2020, drawing on industry survey data compiled through that period. Uptime Institute’s own Global Data Center Survey 2025 describes the same population moving more slowly still: average server rack densities are rising, driven mainly by wider adoption of racks in the 10 kW to 30 kW range, with few facilities reporting anything above 30 kW. Source: Isazadeh, Ziviani, and Claridge, “Global trends, performance metrics, and energy reduction measures in datacom facilities,” Renewable and Sustainable Energy Reviews (2023); Uptime Institute, Global Data Center Survey 2025

AFCOM’s 2025 State of the Data Center Report describes a different segment and gets a very different number: rack density for recent deployments climbing from 7 kW in 2021 to 16 kW, with most respondents expecting further growth. That figure is not wrong next to Uptime’s. It is measuring a self-selected group of operators reporting on their newest builds, which skew toward the racks people are choosing to install now rather than the stock already running. Table 1. Same year, different rack density numberskW per rack

* Installed-base averages and single flagship-rack specifications describe different populations; not directly comparable

Source: Uptime Institute, Global Data Center Survey 2024 and 2025; AFCOM, 2025 State of the Data Center Report; NVIDIA DGX GB200 NVL72 rack guide; HPE QuickSpecs, NVIDIA GB200 NVL72 by HPE

Reading “rack density is now X kW” without asking which population X describes is the single most common error in this field. An installed-base average and a recent-deployment figure will diverge by a factor of two even in the same year, from the same country, without either number being incorrect.

One rack, three vendors, three numbers

NVIDIA’s DGX GB200 NVL72 rack guide lists the reference architecture at approximately 120 kW. HPE’s QuickSpecs for its own GB200 NVL72 configuration state a 132 kW nominal thermal design power (TDP), with an electrical design power at peak (EDPp) of roughly 1.5 times TDP, or about 192 kW, which HPE recommends as the busway provisioning target. Vertiv’s reference design, developed jointly with NVIDIA, is described as supporting up to 132 kW per rack within a 7 MW facility architecture. Source: NVIDIA DGX GB200 NVL72 rack guide, as summarized in Pantheon, “GB200 & GB300 NVL72 Power and Cooling Requirements” (2026); HPE, NVIDIA GB200 NVL72 by HPE QuickSpecs; Vertiv, “Vertiv Codevelops with NVIDIA Complete Power and Cooling Blueprint for NVIDIA GB200 NVL72 Platform” (2024)

These three numbers are not a discrepancy to resolve. Reference power describes an architecture at a stated operating point. Nominal TDP describes what the OEM configuration is rated to draw under typical load. EDPp describes the peak the electrical plant must be sized for, including margin for transients. A facility team that provisions to the 120 kW reference figure and gets billed against the 192 kW EDPp figure has read the wrong column, not received a bad spec.

Even the flagship liquid-cooled rack keeps an air-cooled load at the classic ceiling

The GB200 NVL72 is a direct-to-chip liquid-cooled rack, but it is not an all-liquid rack. Uptime Institute’s own technical journal states that at full system load, cold plates on the GB200 NVL72 handle around 100 kW of the rack’s thermal output, while the air exhaust from fans, power shelves, and NVLink switches exceeds 25 kW. Source: Uptime Institute, “Density choices for AI training are increasingly complex,” Uptime Institute Blog (2025)

Twenty-five kilowatts is not a coincidental number here. It sits at the top of the range most engineers already use as the point where air cooling stops being straightforward, the same range covered in the next section. A rack that liquid-cools 100 kW of silicon still needs the same rear-door heat exchanger or in-row cooling infrastructure that a fully air-cooled rack would need at its own upper limit. Liquid cooling removes the dominant heat load. It does not remove the airflow design problem; it shrinks it down to exactly the size that was already considered hard.

The 20 to 25 kW air cooling limit is an airflow equation, not a wall

The commonly cited ceiling for air cooling, 20 kW to 25 kW per rack, is not a property of air or a regulatory limit. It falls out of the sensible-heat balance that governs how much airflow is needed to carry a given heat load across a given temperature rise: CFM = Q ÷ (1.08 × ΔT), where Q is the heat load in BTU per hour and ΔT is the intake-to-exhaust temperature rise in degrees Fahrenheit. Source: Data Center Frontier, “Understanding the Physics of Airflow in High Density Environments”

Assume a 20°F (11°C) rise, within the range most air-cooled equipment is designed around. A 25 kW rack generates 85,300 BTU per hour (25 kW × 3,412 BTU/hr per kW). Required airflow is 85,300 ÷ (1.08 × 20), or about 3,950 CFM. Scale the same equation to a fully air-cooled 120 kW load and the requirement is roughly 18,950 CFM, a figure that matches the approximately 19,000 CFM independently cited for a 120 kW to 150 kW rack at the same 20°F rise. Source: coolnetsystem.com, “120kW Racks Demand High-Density Data Center Cooling, Not Air” (2026)

This is a planning-level calculation, not a measured value; actual delivered airflow depends on containment, fan curves, and static pressure loss through populated server chassis, none of which the equation captures. What it shows is that the 20 to 25 kW figure describes what a raised floor, a perforated tile, and a standard containment aisle can practically deliver, not a thermodynamic boundary. Push the same arithmetic to 120 kW and the airflow requirement roughly quintuples, past the point where any floor-tile-based delivery scheme works at all. Raising fan speed to chase that airflow solves the equation on paper and produces a rack that is loud, hard to service, and burning a meaningful share of its own power budget on moving air rather than computing.

The generation curve, and where the confirmed numbers stop

Lay the confirmed figures end to end and the curve is not gradual. Average installed-base density took nine years to roughly triple, from 2.4 kW in 2011 to 8.4 kW in 2020. The flagship AI rack then went from that same single-digit range to 120 kW to 132 kW within the 2024 to 2025 product cycle, a jump of well over an order of magnitude concentrated into two to three years rather than a decade. Table 2. The generation curve, confirmed versus projectedkW per rack

* 2027 figure is a target stated at a keynote, not a published OEM data sheet

Source: Isazadeh, Ziviani, and Claridge, Renewable and Sustainable Energy Reviews (2023); NVIDIA DGX GB200 NVL72 rack guide; HPE QuickSpecs; Data Center Dynamics and Tom’s Hardware reporting on NVIDIA GTC 2025

NVIDIA has already named a number for the generation after that. At GTC 2025, CEO Jensen Huang described the 2027 Rubin Ultra NVL576 rack, built on the “Kyber” rack architecture, at roughly 600 kW, with an all-liquid cooling design that eliminates fans entirely. No shipping specification exists yet; the product is scheduled for the second half of 2027, and the figure comes from a keynote remark rather than a data sheet. Source: Data Center Dynamics, “Nvidia’s Rubin Ultra NVL576 rack expected to be 600kW, coming second half of 2027” (2026); Tom’s Hardware, “Nvidia shows off Rubin Ultra with 600,000-Watt Kyber racks and infrastructure, coming in 2027” (2025)

Treating 600 kW as confirmed the way 132 kW is confirmed would be a category error. One is a shipping OEM data sheet with a stated TDP and EDPp. The other is a public target stated ahead of any spec sheet, from a company with a direct incentive to signal the scale of what is coming. Both facts are worth tracking. Only one belongs in a facility’s current provisioning plan.

Why the density figures do not line up

This piece does not resolve whether AFCOM’s 16 kW recent-deployment figure and Uptime’s sub-10 kW average figure are measuring comparable facility types, since neither report publishes a matched sample; the difference could reflect population skew, self-reporting bias, or a genuine and recent inflection, and the sources available do not distinguish between those explanations. It does not cover actual measured airflow or fan power draw for a populated GB200 NVL72 rack under real containment conditions; the CFM figures above are a planning-level calculation, not a site measurement. It does not cover the cost or lead time of retrofitting an existing air-cooled hall for direct-to-chip liquid cooling, and it treats the Rubin Ultra NVL576 figure strictly as an unconfirmed, pre-release target rather than a design input.

Labeling a density figure before plotting it

• Ask which population a quoted rack density figure describes: installed-base average, a recent-deployment segment, or a single flagship SKU, before comparing it to any other number.
• When two OEMs publish different kW figures for the same rack model, check whether each is reference power, nominal TDP, or EDPp peak; they are not the same quantity and should never be plotted on one axis without a label.
• Treat any density figure tied to a product that has not shipped, including the 600 kW Rubin Ultra figure, as a stated target with a date attached, not a design input.
• If a rack’s residual air-cooled load approaches 20 kW to 25 kW, size the airflow with the sensible-heat equation rather than assuming it scales linearly with the rack’s total power draw.
• Confirm whether a vendor’s CFM or ΔT figure assumes hot-aisle or cold-aisle containment; uncontained delivery requires materially more airflow for the same nameplate heat load.
• Route any binding cooling design decision through mechanical engineering review specific to the site; none of the figures in this piece substitute for a load calculation done against the actual facility.

References

• Isazadeh, A., Ziviani, D., and Claridge, D.E., “Global trends, performance metrics, and energy reduction measures in datacom facilities,” Renewable and Sustainable Energy Reviews, 174 (2023). https://ideas.repec.org/a/eee/rensus/v174y2023ics1364032123000059.html
• Uptime Institute, Global Data Center Survey 2025. https://datacenter.uptimeinstitute.com/rs/711-RIA-145/images/2025.Annual.Survey.Report.pdf
• Uptime Institute, Global Data Center Survey 2024. https://datacenter.uptimeinstitute.com/rs/711-RIA-145/images/2024.GlobalDataCenterSurvey.Report.pdf
• Uptime Institute, “Density choices for AI training are increasingly complex,” Uptime Institute Blog (2025). https://journal.uptimeinstitute.com/density-choices-for-ai-training-are-increasingly-complex/
• CoreSite, “How Colocation Data Centers Are Helping Solve the Power Density Challenge” (2025), citing AFCOM 2025 State of the Data Center Report. https://www.coresite.com/blog/how-colocation-data-centers-are-helping-solve-the-power-density-challenge
• Pantheon, “GB200 & GB300 NVL72 Power and Cooling Requirements” (2026). https://pantheon.run/learn/nvidia-gb200-nvl72-power-and-cooling
• HPE, NVIDIA GB200 NVL72 by HPE QuickSpecs. https://www.hpe.com/us/en/collaterals/collateral.a50009224enw.html
• Business Wire, “Vertiv Codevelops with NVIDIA Complete Power and Cooling Blueprint for NVIDIA GB200 NVL72 Platform” (2024). https://www.businesswire.com/news/home/20241014494536/en
• Data Center Frontier, “Understanding the Physics of Airflow in High Density Environments.” https://www.datacenterfrontier.com/special-reports/article/11427261/understanding-the-physics-of-airflow-in-high-density-environments
• coolnetsystem.com, “120kW Racks Demand High-Density Data Center Cooling, Not Air” (2026). https://www.coolnetsystem.com/120kw-racks-demand-high-density-data-center-cooling-not-air
• Data Center Dynamics, “Nvidia’s Rubin Ultra NVL576 rack expected to be 600kW, coming second half of 2027” (2026). https://www.datacenterdynamics.com/en/news/nvidias-rubin-ultra-nvl576-rack-expected-to-be-600kw-coming-second-half-of-2027/
• Tom’s Hardware, “Nvidia shows off Rubin Ultra with 600,000-Watt Kyber racks and infrastructure, coming in 2027” (2025). https://www.tomshardware.com/pc-components/gpus/nvidia-shows-off-rubin-ultra-with-600-000-watt-kyber-racks-and-infrastructure-coming-in-2027