How Much Power Do AI Data Centers Use? – From One Rack to a Campus
Ten years ago a busy server rack drew about as much power as a couple of household ovens. Today a single rack built for training large models can pull more than a small neighborhood. That jump changes everything downstream, from the cooling to the conversation with the utility, and anyone sizing a site or reading the headlines needs a feel for the real numbers. How much power do AI data centers use, from one rack up to a full campus?
AI data center power consumption starts at the rack: a dense accelerator rack can draw well over 100 kilowatts (kW), compared with roughly 5 to 10 kW for a typical enterprise rack. A hall of a thousand such racks reaches 100 MW of IT load before cooling and losses, and the largest campuses now plan for a gigawatt (GW) or more, built in phases. Actual use depends on chip type, utilization, cooling and whether the site trains models, serves them or both.
Why Do AI Racks Draw So Much More Power?
Because the chips themselves draw far more, and because AI systems pack as many of them as possible into one rack so they can talk to each other quickly. A conventional server might hold two general-purpose processors. An AI training server holds several accelerators, each one a power-hungry piece of silicon, plus the processors, memory and networking that feed them.
The vendor numbers make the point. NVIDIA lists the maximum thermal design power of its H100 SXM at up to 700 watts, and the company's GB200 NVL72 system packs 72 Blackwell GPUs and 36 Grace CPUs into one liquid-cooled rack. For a deeper look at the chip-level side, our guide to GPU power consumption follows the watts from a single chip up through the server.
Here is a rough model of why rack numbers climb so fast. It is illustrative, not a vendor spec:
- Assume 72 accelerators at about 1,000 W each: 72 × 1,000 W = 72 kW.
- Add host processors, memory, network switches, power conversion and pumps or fans. Call it another 40 to 60 percent.
- Rack total: roughly 100 to 115 kW.
Every watt you send to a server comes back as heat, and at those densities air simply cannot carry it away fast enough. That is why most new AI racks use direct-to-chip liquid cooling, with coolant running through cold plates mounted on the chips.
How Fast Has Rack Density Changed?
Quickly, and the average is still climbing. AFCOM's annual survey, as reported by Data Center Knowledge in March 2026, found that average rack density reached 27 kW, up from 16 kW a year earlier and about 7 kW in 2021. The same survey found 36 percent of respondents had liquid cooling deployed.
Keep in mind that figure is an average across all kinds of facilities, so it hides a wide spread: plenty of enterprise rooms still run racks under 10 kW, while purpose-built AI halls run many times the average.
From an operator's chair, that spread is the real story. A site designed for 10 kW racks cannot simply swap in 100 kW racks. The busway, the cooling plant, the floor loading and even the fire protection plan were all sized for something else.
AI Data Center Power Consumption From Rack to Campus
The easiest way to understand AI data center power consumption is to scale it up in steps. The table compares a traditional enterprise build with an AI build using the same round numbers. These are illustrative inputs, chosen to show the math rather than to describe any one project.
| Step | Traditional enterprise | AI training build |
|---|---|---|
| Power per rack | 8 kW | 120 kW |
| IT load, 1,000 racks | 8 MW | 120 MW |
| Assumed PUE | 1.5 | 1.2 |
| Facility load | 12 MW | 144 MW |
| Yearly energy at 80% average load | about 84,000 MWh | about 1,009,000 MWh |
Enterprise build vs AI training build, MW at the fence
Model: illustrative inputs from the table above (1,000 racks; rack power and PUE as labelled)
View the data
| Facility load | |
|---|---|
| Traditional enterprise (8 kW racks, PUE 1.5) | 12 MW |
| AI training (120 kW racks, PUE 1.2) | 144 MW |
Showing the Math
For the AI build: 1,000 racks × 120 kW = 120,000 kW, or 120 MW of IT load. Power usage effectiveness (PUE) of 1.2 means every IT kilowatt carries 0.2 kW of cooling and electrical overhead, so 120 MW × 1.2 = 144 MW at the fence. At an 80 percent average load over a year, 144 MW × 0.8 × 8,760 hours = 1,009,152 megawatt-hours (MWh), or about 1 terawatt-hour.
Now scale to a campus planned for 1,000 MW at the fence. Divide by the PUE to get IT load (1,000 ÷ 1.2 ≈ 833 MW), then divide by rack power (833,000 kW ÷ 120 kW ≈ 6,900 racks). Under 7,000 racks, each about the size of a refrigerator, can ask the grid for as much power as a large nuclear reactor produces.
Notice that the AI build has the better PUE in this example. Liquid cooling moves heat more efficiently than air, so overhead per IT watt can drop even as total power soars. Efficiency and total demand are separate questions.
How Is an AI Campus Different From a Traditional Data Center?
An AI campus is built around a few very large, very dense loads, while a traditional data center is built around many smaller tenants and mixed workloads. That difference shows up in almost every system on the site, not just the electric bill.
- Density: AI halls concentrate 100 kW or more per rack, so a building of the same size holds many times the load of an enterprise hall.
- Cooling: liquid loops, coolant distribution units and dry coolers or towers replace much of the room air handling that older halls depend on.
- Electrical design: bigger busways, more transformers per hall and, on the largest campuses, medium-voltage distribution pushed closer to the racks.
- Utilization: training clusters tend to run near full output for long stretches, while many enterprise rooms run well below their nameplate capacity.
- Location: training is less sensitive to network delay than serving live users, so developers can chase cheap, available power in places a cloud region would never go.
Utilization deserves extra attention because it changes the energy math more than people expect. An enterprise room rated at 10 MW might average a fraction of that in practice, because servers idle and racks sit half full. An AI training hall rated at 100 MW may average much closer to its rating. Two facilities with the same nameplate can differ several times over in yearly energy use.
Redundancy is also being rethought. Traditional enterprise and colocation sites usually carry full generator and UPS backup for every rack, because a crashed banking system or retail site costs money by the minute. Some AI developers question whether every training rack needs the same protection, since training jobs can save checkpoints and restart. Others, especially those serving live users, keep the full backup stack. Neither choice is free: less backup lowers capital cost but raises the cost of each outage.
Floor loading and building height change too. A liquid-cooled AI rack full of coolant and copper is much heavier than an older air-cooled rack, and overhead space fills with pipe as well as cable tray. Retrofitting an old hall for AI often runs into the slab and the roof before it runs out of power.
Where Does the Power Go Inside an AI Data Center?
Most of it goes to the servers, and the share going to cooling depends heavily on the type of facility. The International Energy Agency's Energy and AI analysis estimates that servers use about 60 percent of data center electricity overall, and that cooling ranges from about 7 percent in efficient hyperscale facilities to more than 30 percent in less efficient enterprise sites.
The rest goes to storage, networking, power conversion losses in transformers and uninterruptible power supply (UPS) systems, lighting and offices. In an AI hall, the accelerators dominate the IT share, and the network fabric that ties thousands of them together is a meaningful load of its own.
For operators, this split matters during design reviews. Shaving cooling overhead is worth chasing, but on a modern AI site the IT load is so large that even a perfect cooling system leaves the utility request mostly unchanged.
Training vs Inference: Which Uses More Power?
Training and inference are two different jobs with two different power profiles, and both are large. Training builds a model by running enormous clusters of accelerators together, often around the clock for weeks or months. Inference is the serving side: answering prompts, generating images and running the model for users once it exists.
How Training Loads Behave
Training wants the biggest possible cluster in one place, tightly networked, because the chips must constantly share results. That pushes developers toward very large single campuses with huge, steady power draws. It also creates a quirk: Microsoft Research, with co-authors from OpenAI and NVIDIA, described in 2025 how synchronized training jobs swing between compute and communication phases, causing large, rapid power oscillations that utilities and on-site generators have to absorb.
How Inference Loads Behave
Inference is spread across many sites and follows user demand, so its load has daily and weekly patterns more like a traditional cloud. Each request is small. Published per-prompt estimates for text queries are typically a fraction of a watt-hour, though long inputs and image or video generation use far more. The total is large because the number of requests is enormous and still growing.
Which uses more over a year depends on the company and the year, and published splits vary. The fair answer is "both, and the mix keeps shifting." Training tends to set the biggest single-site loads, while inference sets the breadth of the buildout across many regions.
How Big Are AI Campuses Getting?
The planning unit has moved from tens of megawatts to hundreds, and in some cases to multiple gigawatts built in phases. A decade ago a 30 MW facility was big. Today, utility filings and company announcements routinely describe campuses of 300 MW, 1 GW and beyond.
Some of these are now well documented. Louisiana regulators approved three new Entergy gas plants, roughly 2.3 GW in total, to serve Meta's data center campus in Richland Parish. OpenAI and Oracle have described more than 5 GW of Stargate capacity under development, including the Abilene, Texas, site. These are announced plans, though, and phases can slip, shrink or grow.
When you read a big campus number, check which number it is. A campus "up to 1 GW" usually means the long-term grid request, not what runs on opening day. IT load, facility load and the interconnection request are three different figures, and announcements do not always say which one they mean.
What Does This Mean for the Grid and On-Site Power?
It means the power plan now drives the site plan. A load of several hundred megawatts usually needs new transmission or substation capacity, and that work runs on utility timelines measured in years. Many developers respond by pairing a grid connection with on-site generation, using gas turbines, reciprocating engines or fuel cells as bridge power or as a permanent supply.
The load profile matters too. Large AI sites can be sensitive to voltage dips, and training clusters can swing quickly. Batteries, UPS systems with fast controls and generators chosen for good load-step response all help keep those swings from reaching the grid or tripping a local plant.
For a site selector, the practical checklist starts with three numbers: the IT load the tenant wants, the PUE the cooling design can hold on the hottest day, and the date the power has to be live. Get those three right and the rest of the conversation with the utility gets much simpler.
From 700 Watts to a Gigawatt: Reading AI Power Numbers With Confidence
Think of AI data center power consumption as a stack of multipliers: watts per chip, chips per rack, racks per hall, halls per campus, then PUE for overhead and utilization for the yearly total. Every big number in the news is just those steps multiplied out, and any one of them can be fudged.
The next time you see a headline number, run the steps yourself: rack power, rack count, PUE and hours. You will quickly see whether the figure is IT load, facility load or a long-range grid request, and you will be asking the questions that matter for power, cooling and timing.