Data Center Rack Density Trends: Average kW per Rack From 2011 to AI Racks

Rack density is the IT power drawn by the equipment in one cabinet, in kilowatts. It sounds like one number, but surveys report several, and mixing them up is the most common source of confusion in density discussions.

• Average density: total IT load divided by rack count. Uptime calculates its headline mean after removing respondents above 30 kW as high-performance outliers.⁶
• Modal density: the density most racks in a facility actually run at. Uptime reports this separately because it describes the typical rack better than a skewed mean.⁶
• Peak density: the single densest rack on site. A facility can have a modest average and a handful of very dense cabinets.¹⁰
• Design density: what the power and cooling were built to support. Operators usually provision above what racks draw today.⁶

For site and building planning, design density is what matters, because it sets the electrical and cooling plant. Survey averages describe the installed base, which changes slowly because most data halls run for many years before they are refitted. That is why the headline averages move by tenths of a kilowatt a year while new AI halls are designed for 100 kW or more per rack.

Uptime Institute has asked operators about rack density in its annual global survey for more than a decade, and the long-run picture is gradual. As compiled by Raritan, a rack power equipment maker, from Uptime’s data, the average was about 2.4 kW per rack in 2011 and 5.6 kW in 2017.⁵ In 2020, Uptime reported a mean of 8.4 kW after excluding respondents above 30 kW, and said that was safely within the provisioned range of most facilities.⁶ In the same survey, 71% reported a modal density below 10 kW and the most common band was 5–9 kW.⁶

The dip from 2020 to the mid-2020s is a reminder that each year’s figure depends on who answers. Uptime’s 2024 survey report said average densities were increasing but remained below 8 kW, and that most facilities had no racks above 30 kW, while those that did had only a few.¹⁰

The 2026 survey shows the tail starting to move. The modal density passed 11 kW for the first time, and a growing number of operators reported peak densities of 30 kW or more for advanced computing.¹ Data Center Knowledge’s coverage framed the same results as AI driving uncertainty in planning, with operators serving AI training clusters and CPU-based enterprise workloads at the same time.¹¹

GPU systems sit far above the survey averages. SemiAnalysis compares a general-purpose CPU rack at up to about 12 kW with an air-cooled H100 rack at about 40 kW, and puts NVIDIA’s GB200 NVL72 at roughly 120 kW per rack. It identifies exceeding about 40 kW per rack as the main reason GB200 requires liquid cooling.²

The roadmap keeps climbing. NVIDIA plans its Rubin Ultra NVL576 rack for the second half of 2027, and CEO Jensen Huang said the liquid-cooled Kyber rack would consume about 600 kW, five times a 120 kW Blackwell rack.³ Through the Open Compute Project, Google has described moving to ±400 VDC power delivery so that IT racks can scale from 100 kW up to 1 MW.¹² Forecasts that extrapolate these figures to the whole market should be read with care: they describe the top of the market, not the installed base.

Our guide to high-density GPU racks and site power covers what a 120 kW rack asks of the building and the utility in detail. This guide focuses on the trend and what it means for how much density to plan for.

AFCOM’s State of the Data Center report tells a much faster story than Uptime. Data Center Knowledge reported AFCOM’s 2026 average at 27 kW per rack, up from 16 kW the year before, the largest year-over-year increase in the report’s decade-long history, and from 7 kW in 2021.⁷ Capacity, reporting the Uptime figures, noted that the two surveys measure different populations.⁸

None of these numbers is wrong; they answer different questions. Uptime’s exclusion of high-density outliers keeps its mean focused on mainstream enterprise and colocation space. Market forecasts from brokers and consultancies, such as JLL’s 2024 projection that the average would reach 50 kW per rack by 2027 from 36 kW, lean toward new construction, where AI demand is concentrated.¹³ When someone quotes an average density, ask which survey, which year and whether outliers were excluded.

Air cooling still dominates. In Uptime’s 2025 cooling systems survey, 75% of operators answering the question used perimeter air cooling and 22% used some direct liquid cooling (DLC), a share Uptime described as slow and steady growth. High rack density remained the top reason for adopting DLC, and operators judged liquid systems mainly by how readily they integrate into existing infrastructure.⁹

That integration question is where the density trend becomes a real estate problem. JLL reported North American vacancy at 1% in its midyear 2026 report, and its research lead said only a very small percentage of that space can support high-density deployments. A broker seeking 140–160 kW per rack approached more than 10 operators and found that almost none could support that density with direct liquid cooling.⁴ Available space is not the same as usable space for AI.

For the site, liquid cooling changes the outdoor plant as well as the hall: more heat per building, warmer water loops and different heat rejection equipment. See air vs. liquid cooling and cooling plant and equipment yard space.

Density multiplies the load a given floor area carries. As illustrative arithmetic only: a hall of 1,000 racks draws about 8 MW of IT load at 8 kW per rack, 40 MW at 40 kW and 120 MW at 120 kW. The building is roughly the same size in each case; the utility service, substation, electrical rooms, UPS and distribution and heat rejection are not. That is why dense campuses are increasingly limited by megawatts rather than acres, as covered in how much land a data center needs and data center power requirements.

The hardware cycle is the planning risk. Rack power has roughly tripled from air-cooled H100 racks to GB200, and NVIDIA’s roadmap points another five times higher within a few years.²³ A building designed tightly around today’s rack may be undersized for the next generation, while one designed for megawatt racks may sit underused if the tenant runs enterprise workloads. Uptime’s finding that operators are planning for AI clusters and CPU-based enterprise workloads at once is a reason to keep electrical capacity, floor loading and piping routes flexible.¹

For existing buildings, density is often the deciding factor in whether a retrofit beats a new build, because few existing halls can take the densest racks.⁴

1. 01Ask the likely end user for its design density in kW per rack, its cooling approach and how many racks per hall. Do not assume a survey average.
2. 02Convert density to megawatts per building and per phase, then compare with what the utility can deliver and when; see load ramp schedules.
3. 03Check whether the layout leaves room to grow heat rejection and electrical rooms if density rises mid-life.
4. 04For an existing building, confirm whether it can take liquid cooling distribution and the heavier racks that come with it; most current space cannot support 140 kW+ racks.⁴
5. 05When a pitch cites an “average” density, ask which survey, which year and whether outliers were excluded.

Rack density decides how much power, cooling and structure each square foot needs, so it belongs at the start of AI data center site screening, not at design development. BlackForge tests power and layout against the density a site is being marketed for; you can get a site reviewed.