AI Data Center Power Explained: Megawatts, PUE and Why Grid Access Is the New Bottleneck
For most of the past two years, the scarce input for AI was chips. Increasingly it is electricity. Companies can order GPUs, but they cannot easily get a few hundred megawatts of grid power delivered to a site, and the wait for a new grid connection in many US regions now runs for years. That is why data center deals are announced in gigawatts, why tech companies are signing contracts to restart nuclear plants, and why some sites are building their own gas turbines.
This guide explains the units, the efficiency numbers and the bottlenecks, so the headlines make sense.
The units
Data center power gets reported in a confusing mix of units. Here is the short version.
| Unit | What it measures | Scale |
|---|---|---|
| Kilowatt (kW) | Power at a moment | One server rack used to draw 5 to 10 kW |
| Megawatt (MW) | 1,000 kW | A traditional enterprise data center: 1 to 20 MW |
| Gigawatt (GW) | 1,000 MW | Roughly one large nuclear reactor’s output; the size of the biggest AI campuses now planned |
| Megawatt-hour (MWh) | Energy: 1 MW running for one hour | Used for electricity bills and supply contracts |
| Terawatt-hour (TWh) | 1 million MWh | Used for national and global consumption |
Power (MW) is about capacity: how big a connection a site needs. Energy (MWh or TWh) is about how much electricity it uses over time. A 100 MW data center running flat out all year uses about 876,000 MWh, or 0.876 TWh.
Headline numbers like “a 1 GW campus” usually mean the connection or the IT load the site is planned for, not what it draws on day one.
How AI changed the rack
The biggest change is density, meaning how much power is packed into one rack.
| Era | Typical rack power |
|---|---|
| Traditional enterprise servers | 5 to 10 kW |
| Early GPU training clusters | 20 to 40 kW |
| Nvidia GB200 NVL72 | Around 120 to 130 kW |
| Next-generation designs planned for 2027 | Several hundred kW, with Nvidia showing designs near 600 kW |
A tenfold or larger jump in rack power breaks old assumptions. Air cannot carry that much heat away, so liquid cooling (cold plates on the chips, with coolant piped through the rack) has gone from niche to standard for AI. Power distribution inside buildings is also changing, with Nvidia and others pushing 800-volt DC distribution to cut losses and copper weight.
PUE: how efficient is the building
Power usage effectiveness (PUE) compares the total power a facility uses with the power that reaches the IT equipment.
PUE = total facility power รท IT equipment power
A PUE of 1.0 would mean every watt goes to computing. A PUE of 1.5 means half a watt is spent on cooling, power conversion and lighting for every watt of compute. Industry surveys put the average across all data centers at roughly 1.5 to 1.6 for years now. The best hyperscale sites run near 1.1, and Google reports a fleet average around that level.
PUE has limits. It says nothing about whether the computing itself is useful or efficient, and it can look better in cold climates. Liquid cooling tends to improve it, because pumping liquid takes much less energy than moving huge volumes of air.
How much power AI data centers use
A few reference points, from the most widely cited studies:
- A December 2024 report from Lawrence Berkeley National Laboratory found US data centers used about 4.4% of US electricity in 2023 and projected 6.7% to 12% by 2028.
- The International Energy Agency estimated global data center use at about 415 TWh in 2024, roughly 1.5% of world electricity, and projected it could more than double by 2030.
The range in these forecasts is wide, because nobody knows how many announced projects will actually get built. That uncertainty is part of the bottleneck: utilities have to plan generation and transmission for load that may or may not arrive.
Why grid access is the bottleneck
Getting a large new load connected to the grid takes much longer than building the data center itself.
Interconnection queues. In the US, new generation projects and large new loads wait in queues to be studied by grid operators. Lawrence Berkeley’s tracking showed well over 2,000 GW of generation and storage waiting in queues by the end of 2023, more than the entire existing US power fleet. Typical waits from request to operation have stretched to around five years.
Transmission. Even when generation exists somewhere, moving it to a data center can need new high-voltage lines, which take a decade in some places because of permits.
Equipment. Large power transformers and high-voltage switchgear have lead times of two to four years, and only a handful of factories make the biggest units.
Capacity prices. In PJM, the grid operator covering much of the mid-Atlantic and the heart of US data center country in northern Virginia, capacity auction prices jumped to record levels for the 2025/26 delivery year and hit the price cap for 2026/27. Data center demand was a major reason. That cost flows into ordinary household bills, which is turning data center power into a political issue.
The workarounds
Developers are trying several routes around the grid.
- Behind-the-meter gas. Build gas turbines or engines on or next to the site. xAI’s Colossus cluster in Memphis used mobile gas turbines, and large campuses in Texas and Louisiana are pairing data centers with new gas plants.
- Nuclear contracts. Microsoft signed a deal with Constellation to restart a reactor at Three Mile Island (now called the Crane Clean Energy Center). Amazon bought a data center campus next to Talen’s Susquehanna nuclear plant, and Meta signed a long-term deal for Constellation’s Clinton plant. Google has an agreement with Kairos Power for small modular reactors.
- Co-location. Putting data centers next to existing power plants avoids some transmission, but regulators have pushed back. In late 2024 the Federal Energy Regulatory Commission rejected an expanded co-location agreement at Susquehanna, citing grid reliability and cost fairness.
- Flexibility. Some operators are agreeing to cut load during grid emergencies in exchange for faster connections. Studies suggest that even modest flexibility could free up a lot of existing grid capacity.
- Going abroad. Countries with spare power, from the Nordics to the Gulf, are pitching themselves as AI hosts. This links power directly to AI sovereignty debates.
What it means for the AI buildout
Power is now a gating factor for how fast AI capacity can grow, and it shapes where data centers go. It also changes what matters inside the data center. Every watt saved on networking, through technologies like co-packaged optics, or on memory, through better HBM, is a watt that can go to GPUs.
For investors, power is one of the tiers in the AI capex value chain, and it is the one with the longest lead times. Watch interconnection reform at grid operators, the next PJM capacity auctions, how many nuclear restarts and small reactors actually reach operation on schedule, and whether gas turbine makers can keep up with orders.