How much power does an AI data center use? Racks, halls and campuses

Power numbers for AI racks (about 120 kW now, 600 kW planned), for sites in megawatts and gigawatts, and for the US and the world, with sources.

Figure 1. Power and grid. Note: illustration generated for this story; not a chart of real data.

The short version

  • A current NVIDIA GB200 rack draws approximately 120 kW, about 16 times the 7.6 kW mean average rack for 2025 reported by Vertiv citing Uptime Institute data (our division), and newer racks are planned at 190 kW to 600 kW.
  • Data centers used about 4.4 percent of US electricity in 2023 (176 TWh) according to Lawrence Berkeley National Laboratory, and its projection for 2028 is 6.7 to 12 percent.
  • About 58 GW of US data center capacity is under construction, against about 65 GW operating today, so the buildings now going up would need almost as much power as the existing fleet.

An AI data center uses from a few megawatts to more than a gigawatt, and a single AI rack now draws about 120 kW. Those two numbers are linked: the rack sets how much power each square meter needs, and the number of racks sets the size of the grid connection. This guide gives the figures at each level, with the source for each.

Rack level: from 8 kW to 600 kW

A conventional enterprise rack is a small load. Uptime Institute’s 2025 global survey says average rack power density is rising slowly, driven by racks of 10 kW to 30 kW, and that few facilities exceed 30 kW. A separate Vertiv article citing Uptime data puts the mean average at 7.6 kW for 2025; we have not seen that number in Uptime’s own text.

An AI rack is a different class. NVIDIA’s documentation for the DGX GB200 system says the rack power consumption is approximately 120 kW. That rack holds 18 compute trays with four Blackwell GPUs each, which is 72 GPUs in all. The compute trays are liquid-cooled through cold plates on the CPUs and GPUs, while networking and storage are cooled by air. Power comes through up to eight power shelves, each needing 33 kW of input.

Against the 7.6 kW average, 120 kW is about 16 times higher (our division). Against the “up to 30 kW” ceiling that most facilities stay under, it is four times.

Newer racks go higher. On 16 November 2026, Ecolab and CoolIT introduced a cooling unit for racks of about 190 kW, for platforms such as NVIDIA Vera Rubin. NVIDIA’s own plans go further: at GTC 2025, its CEO Jensen Huang said a Kyber rack with 576 Rubin Ultra GPUs could draw up to 600 kW, according to TechRadar, and that future racks might need full megawatts. NVIDIA’s technical blog says it is moving to 800 V DC power distribution to support racks of 1 MW and beyond, starting in 2027, because the current 54 V design hits limits above 200 kW per rack.

Eliovp’s ModFlex product page gives an example configuration of GB300 NVL72 racks at 150 kW each. The range of figures here, 120 kW for GB200, 150 kW in that example, 190 kW for the next generation, shows that “AI rack” is not one number. Quote the platform and the year.

Hall and module level: megawatts

Multiply racks by power. Eliovp’s example of 16 racks at 150 kW is 2.4 MW of IT load. A 3 MW cooling unit sits at a similar scale. Crusoe’s Spark Liquid module supports up to 3.5 MW. At 120 kW per rack that is about 29 racks, if every watt were IT load (our arithmetic; Crusoe has not said how many racks the unit holds). Armada’s Orion is 10 MW per unit with up to 2,880 GPUs across 40 racks; 2,880 divided by 40 is 72 GPUs per rack, the NVL72 format, and 10 MW divided by 40 racks is 250 kW per rack of total unit power, which includes cooling and other overhead (our calculation).

That last number is a useful reminder that facility power is more than IT power. Cooling, power conversion and lighting add load on top of the racks. The ratio of total power to IT power is called PUE, and Uptime Institute reports that the industry average has barely moved for six consecutive years.

Campus level: gigawatts

A gigawatt is 1,000 megawatts. At 120 kW per rack, 1 GW of pure IT load would hold about 8,300 racks. Real campuses hold fewer because of overhead. Large AI campuses are now planned in gigawatts, and the numbers add up quickly. Cleanview counts 58 GW of US data center capacity under construction against 65 GW in operation, and 8.4 GW of new construction started in the third quarter of 2026 alone. The Texas grid operator is reviewing more than 474 GW of connection requests, about 90 percent of them from data centers, and that figure is more than five times the state’s record peak demand. Many of those requests will not be built; requests are not construction.

National and global level: terawatt-hours

Power draw (megawatts) is a rate. Energy use (terawatt-hours per year) is the total over time. Three estimates frame the totals.

For the US, Lawrence Berkeley National Laboratory’s report for the Department of Energy estimates data centers used 176 TWh in 2023, about 4.4 percent of US electricity, up from 58 TWh in 2014. It projects 325 to 580 TWh by 2028, or 6.7 to 12 percent. The spread between 6.7 and 12 percent reflects the uncertainty about how fast AI servers are deployed.

For hyperscale sites, a Harvard study reported by The Register estimated that 403 hyperscale data centers used 68 to 99 TWh in May 2024 through April 2025, about 1.8 percent of US electricity. The researchers worked from each site’s electrical capacity and assumptions about utilization, not from meters, and could not separate AI from general computing. They said demand has probably grown since.

Globally, the IEA estimates data centers used about 415 TWh in 2024, roughly 1.5 percent of world electricity, and projects about 945 TWh in 2030, just under 3 percent. It says AI-oriented accelerated servers are growing about 30 percent a year and conventional servers about 9 percent. The IEA expects the US to add about 240 TWh by 2030, an increase of about 130 percent. The IEA report dates from 2025, so recent construction data may already be ahead of its base case.

The sources use different years, scopes and methods, so their numbers are not directly comparable. LBNL covers all US data centers; Harvard covers only hyperscale ones; the IEA covers the world.

What is driving the totals

The IEA attributes almost half of the net rise in data center electricity demand to accelerated servers, the AI-oriented machines built around GPUs, and about 20 percent to conventional servers. Accelerated servers are growing faster, but they also draw more per rack, so a single new AI hall adds much more load than a conventional hall of the same floor area. The IEA also says data centers account for less than 10 percent of global electricity demand growth from 2024 to 2030, so the load is large for the local grids that serve a campus even though it is a small part of the global total. Seen from a utility, the unit that matters is the single site: one campus of several hundred megawatts is a request the local grid has to study on its own.

Using these numbers for planning

Start from the rack and work upward. Pick the platform, which fixes the kilowatts per rack. Multiply by the number of racks for IT load. Add an overhead allowance for cooling and electrical losses, using the PUE your design supports. Then ask the utility what it can deliver and when, because for most projects the grid connection, not the building, sets the timeline.

Plan for the next rack, too. A hall wired for 120 kW racks cannot take 190 kW racks without new cooling and power distribution. Designs that add capacity in stages, such as modular builds, let operators size each stage for the rack they will actually install.

ModFlex, Eliovp’s modular facility line for high-density AI and HPC, targets the rack classes above: AMD Instinct MI355X, NVIDIA B300 and NVIDIA GB300 NVL72. Capacity is added module by module, and cooling is air, direct liquid or a hybrid, chosen per project. Rack figures are on Eliovp’s ModFlex page, and Eliovp’s post on 1 to 2 MW NVL deployments covers rack power.

Questions readers ask

How many kilowatts does an AI rack use?

NVIDIA's DGX GB200 guide gives approximately 120 kW per rack. A CoolIT cooling unit launched in November 2026 is aimed at racks of about 190 kW, and NVIDIA's CEO said at GTC 2025 that a planned Kyber rack could draw up to 600 kW.

How much power does a typical data center rack use?

Far less. Uptime Institute's 2025 survey says average rack density is rising slowly, mostly through racks of 10 kW to 30 kW, and that few facilities exceed 30 kW. A figure of 7.6 kW for the 2025 mean average comes from a Vertiv article citing Uptime data, which we have not checked against the full report.

How many megawatts is a gigawatt AI campus?

One gigawatt is 1,000 MW. If all of it fed racks of 120 kW, that would be about 8,300 racks. In practice cooling and electrical losses take a share, so the real count is lower. This is our arithmetic, not a vendor figure.

How much of US electricity do data centers use?

Lawrence Berkeley National Laboratory estimated 4.4 percent in 2023, or 176 TWh, and projected 6.7 to 12 percent by 2028 (325 to 580 TWh). A Harvard study of 403 hyperscale sites estimated 1.8 percent for 2024-25 for that class of facility alone.

Why are AI racks so much denser than regular racks?

A rack like the GB200 NVL72 holds 72 GPUs wired together as one system and needs liquid cooling for its CPUs and GPUs. Conventional racks hold general servers that are cooled by air.

Sources

  1. NVIDIA DGX GB200 user guide: hardware, NVIDIA
  2. NVIDIA 800 VDC architecture will power the next generation of AI factories, NVIDIA Technical Blog
  3. Megawatt-class AI server racks may well become the norm before 2030 as Nvidia displays 600kW Kyber rack design, TechRadar
  4. CoolIT & Ecolab launch 3MW CDU, DatacenterDynamics
  5. Uptime Institute Global Data Center Survey 2025, Uptime Institute
  6. Energy and AI: energy demand from AI, International Energy Agency
  7. DOE releases new report evaluating increase in electricity demand from data centers, US Department of Energy
  8. America's hyperscale bit barns draw as much power as entire states, The Register
  9. Cleanview: US data center construction slows in Q3 2026 following record-breaking Q2, DatacenterDynamics
  10. Texas environmental regulator resumes data center permitting following ERCOT review, DatacenterDynamics
  11. Armada launches 10MW modular data center, Orion, DatacenterDynamics

From Eliovp

Posts by our publisher on the same topics, republished from eliovp.com.

Nvidia confirms the modular DC direction

NVIDIA DSX POD brings Vera Rubin AI factories to modular data centers. ElioVP explains site readiness, cooling and the thinking behind ModFlex.