How Much Power Does a Data Center Use? From Rack to Campus

A single enterprise server draws 300-500 watts. A fully loaded AI rack draws up to 132 kilowatts (kW) today, and up to 142 kW with next-generation hardware. A hyperscale AI campus now under construction draws as much as 5 gigawatts (GW) — more electricity than a mid-sized country. Between those two ends of the scale sits every colocation deal, power bill, and grid connection queue in the industry. This guide walks through the actual numbers at every step: server, rack, hall, campus, and grid.

Key takeaways

• A 1U/2U enterprise server uses 300-500W; an 8-GPU AI server tray uses roughly 10-11 kW.
• The average rack across the whole installed base runs 7.8 kW (Uptime Institute, 2026); the most common (modal) figure is 11 kW; AI racks such as NVIDIA’s GB200 NVL72 draw 120-132 kW.
• Power Usage Effectiveness (PUE) — total facility power divided by IT power — averages 1.52 industry-wide, so a 10 MW IT load typically pulls about 15.2 MW at the meter.
• Wholesale colocation is sold in blocks starting at 250 kW; hyperscale AI campuses now range from 100 MW to multi-gigawatt scale.
• Global data center electricity use is projected to nearly double from about 485 TWh in 2025 to 945-950 TWh by 2030 (IEA), reaching roughly 3% of world electricity demand.
• US data centers used an estimated 4.4% of national electricity in 2023 and could reach 6.7-12% by 2028 (LBNL/DOE).

The power ladder: from one server to a gigawatt campus

Every power figure quoted about data centers sits somewhere on the same ladder. Confusing the levels is the single most common mistake in vendor pitches and press coverage — “our facility runs on 100 kW” means something completely different from “our campus is a 100 MW site.”

Level	Typical power draw	Example
Single server (1U/2U, no GPU)	300-500 W	General enterprise compute
8-GPU AI server tray	~10-11 kW	8x NVIDIA H100 at 700W each, plus CPU/networking overhead
Standard colocation rack (average)	7.8 kW	Uptime Institute 2026 survey, excluding high-density outliers
Standard colocation rack (modal/typical)	11 kW	Same survey, most common reported figure
AI rack (GB200 NVL72)	120-132 kW	72 GPUs, direct-to-chip liquid cooling
AI rack (GB300 NVL72, next-gen)	~135-142 kW	Estimated, per integrator guidance
Wholesale colocation block (minimum)	250 kW	CBRE market convention for primary US markets
Mid-size data hall	1-5 MW	Typical enterprise or edge colocation facility
Large hyperscale building	100-300 MW	Single building on a multi-phase campus
Gigawatt-class AI campus	1-5 GW	Meta Hyperion (Louisiana), Microsoft Fairwater (Wisconsin)

How much power a server and a GPU actually use

At the bottom of the ladder, a standard enterprise server without accelerators draws 300-500 watts — the figure that dominated data center planning for two decades. AI hardware broke that assumption. An NVIDIA H100 GPU has a thermal design power around 700 watts on its own, before counting the CPU, memory, networking interfaces, and power-supply losses that sit around it in a server tray. Stack eight of them into one tray, as most AI training servers do, and the tray alone draws roughly 5.6 kW — with the full assembly, including CPUs and NICs, commonly landing near 10-11 kW.

The GB200 Grace Blackwell superchip changes the unit of account again: it pairs one Grace CPU with two Blackwell GPUs on a single module. A complete GB200 NVL72 rack links 72 GPUs across 18 of these compute trays inside one NVLink domain, and the whole rack — GPUs, switches, cooling manifolds — draws roughly 120 kW nominal, with deployed clusters measured at 130-132 kW, according to integrator data. Divide that by 72 GPUs and the effective draw is close to 1.8 kW per GPU once shared infrastructure is included, versus 700W for the chip alone. That gap between per-chip and per-rack figures is why quoted “GPU power” numbers from different sources rarely match — see our guide on rack power density for the full breakdown by hardware generation.

Rack density in 2026: two very different averages

Uptime Institute’s 16th annual Global Data Center Survey, published in 2026, reports two figures that both matter and are often conflated. Excluding a small number of facilities designed for extreme densities, the average rack draws 7.8 kW — barely changed from 7.5 kW the year before, and dominated by the enterprise and legacy colocation installed base that has not been retrofitted for AI. The modal (most common single) figure, however, passed 11 kW for the first time in 2026, up from 9 kW in 2025, as a growing share of new deployments target higher densities from day one.

Neither number describes a dedicated AI hall. Those run at 120-142 kW per rack with mandatory direct-to-chip or immersion liquid cooling — a jump of more than 10x over the broad average in a single hardware generation. Buyers negotiating colocation contracts need to know which of these three figures a quoted “kW per rack” price actually refers to, since the cooling and power-distribution infrastructure required scales non-linearly past roughly 30-40 kW, where air cooling stops being viable (see our liquid cooling guide).

PUE: why the meter reads more than the racks

None of the figures above are what a facility actually pulls from the grid. Power Usage Effectiveness (PUE) is the ratio of total facility power to IT equipment power, and it captures everything the racks do not: chillers, CRAC/CRAH units, UPS losses, power distribution losses, and lighting. Uptime Institute’s 2026 survey puts the industry-wide average at 1.52 — essentially flat for seven consecutive years despite widespread efficiency messaging. That means a facility with 10 MW of IT load typically draws about 15.2 MW at the utility meter.

The average masks a wide spread. Larger, newer hyperscale facilities report PUE closer to 1.1-1.2 because they operate at a scale that justifies free cooling, higher chilled-water setpoints, and custom power distribution — advantages a 2 MW retail colocation hall cannot replicate. When comparing quoted facility power figures across operators, always ask whether the number is IT load, total facility load, or somewhere in between; the difference compounds quickly at scale.

From megawatts to gigawatts: how facilities are actually sized

Commercial power commitments follow a fairly standard ladder. CBRE’s wholesale colocation convention starts contracts at 250 kW and scales in 250 kW-plus increments for primary North American markets. A typical mid-size enterprise or edge data hall runs 1-5 MW total. A single hyperscale building on a larger campus now runs 100-300 MW. Above that, AI-era campuses have broken into gigawatt territory: Microsoft’s Fairwater site in Wisconsin is tracked by Epoch AI at an end-state capacity of roughly 2,263 MW of IT power by 2028, and Meta’s Hyperion campus in Richland Parish, Louisiana was expanded in mid-2026 to a 5 GW target, with the company pledging more than $50 billion in total investment. Both are outliers even within the hyperscale tier — most new campus announcements in 2026 are still designed in the 100-300 MW range with room to expand, per industry tracking.

The practical implication for anyone evaluating a colocation or build-to-suit deal: ask for the committed kW figure, the PUE assumption baked into the pricing, and the timeline for grid interconnection — not just the headline campus capacity, which is frequently a multi-year, multi-phase end state rather than what is actually energized today. Grid connection lead times, not floor space, are now the binding constraint on how fast any of these numbers can be realized; see our guide on grid connection queues for typical wait times by market.

Turning kilowatts into a power bill

Electricity rates vary by a factor of 10 or more across the markets tracked in Coloprice’s power index, which changes the economics of a given kW commitment substantially depending on location.

Market	Commercial rate (USD/kWh)	Approx. monthly cost per 1 MW IT load*
US Virginia	0.117	~$85,400
US Texas	0.085	~$62,000
Singapore	0.27	~$196,900
Malaysia	0.135	~$98,500
Thailand	0.128	~$93,300
Sweden	0.10	~$72,900

*Illustrative only: assumes a 1.52 PUE (1.52 MW drawn at the meter per 1 MW of IT load) run continuously for 730 hours/month, at the listed commercial rate. Actual bills add demand charges, taxes, and market-specific surcharges — see power-index.json for the full rate table with sources and notes per market.

Even before demand charges and taxes, the spread between Singapore (highest among major APAC colocation hubs) and Texas (deregulated ERCOT market) is roughly 3x on the same 1 MW load — a gap larger than most negotiated colocation discounts, and the main reason power sourcing, not real estate, now drives site-selection decisions for large deployments. Our global data center index tracks wholesale colocation pricing by market for the same reason.

What this looks like at the national and global level

Aggregate all of the above across every facility in operation and the numbers move from company-specific to macroeconomic. The IEA’s Electricity 2026 report estimates global data center electricity consumption at roughly 485 TWh in 2025, projected to reach 945-950 TWh by 2030 — a near-doubling that would put data centers at about 3% of global electricity demand. AI-focused capacity grows far faster than the segment overall, roughly tripling in that same five-year window as the mix shifts from general-purpose compute toward accelerated hardware.

The United States shows the sharpest trajectory among individual countries. Lawrence Berkeley National Laboratory’s report for the US Department of Energy put data centers at 4.4% of national electricity consumption in 2023 (176 TWh), with a base-case forecast of 6.7-12% by 2028 (325-580 TWh) depending on how fast AI deployment continues — a compound annual growth rate of 13-27% through the forecast period, up sharply from roughly 18% between 2018 and 2023 and about 7% in the decade before that. The IEA separately estimates data centers will drive around half of all US electricity demand growth through 2030.

What to do with these numbers

For anyone actually sizing a deployment or evaluating a colocation quote, three checks follow directly from the figures above. First, confirm whether a quoted power figure is IT load, total facility load, or committed (contracted but not necessarily used) capacity — the gap between IT load and meter load is the facility’s PUE, and a 1.1 versus 1.5 PUE changes the real power bill by more than 35% on identical IT hardware. Second, match your rack-density assumption to your hardware roadmap: legacy enterprise gear fits comfortably in the 7-11 kW average rack, but any GPU-dense deployment needs to plan for 120 kW-plus racks and the liquid cooling infrastructure that comes with them well before signing a lease. Third, treat electricity cost per kWh as a primary site-selection variable, not an afterthought — the difference between a 0.085 and a 0.27 USD/kWh market compounds into millions of dollars a year on a multi-megawatt commitment, often exceeding the base rent itself. Coloprice’s colocation and AI data center sector pages and the global facility directory list committed power figures per site where operators disclose them, precisely so these comparisons can be made before a contract is signed rather than after.