Gigawatt, gigawatt, everywhere a gigawatt. We’ll often hear something like, OpenAI’s Stargate is “9 gigawatts by 2029.” Meta is building a “5-gigawatt” campus. Or on a smaller scale, some startup you’ve never heard of just signed a “500-megawatt” deal.
But when you buy a computer, only the power supply is sold by the watt. You’re probably more interested in the Clock speed (gigahertz- GHz), how many processing units are on the chip, or memory (gigabytes), so why is AI measured in gigawatts?
Here’s what those stories often skip: a gigawatt of AI isn’t a measure of computing power. It’s a measure of electricity. When a company says it’s building a 1-gigawatt AI data center, it means the facility is designed to draw that much power from the grid when it’s running flat out.
That may sound like a small detail. It isn’t. Once you know what the number means, and what it doesn’t, these headlines become easier to read. You also get a better feel for where the money is really going.
Why Power Became the Yardstick
For a while, AI capacity was measured in chips, or in some kind of “computing speed.” The media talked about how many GPUs a company owned.
The trouble is that chips keep changing. A rack of Nvidia’s newest gear does far more work than a rack from three years ago, so “100,000 GPUs” means something different every year. Counting chips is like measuring a trucking company by the number of trucks without asking how big they are.
Power doesn’t have that problem. A gigawatt is a gigawatt. And, more and more, power is the thing that’s actually hard to get. You can order chips. You can pour concrete. But getting a utility to connect a gigawatt of new load can take years. 1 gigawatt (GW) = 1,000 megawatts (MW)= 1,000,000 kilowatts (kW). So the industry has settled on a number that’s easy to compare and tied to the real bottleneck: how much electricity the site can draw.
Two Numbers Hiding Behind One Headline
Here’s where the confusion comes in. When the news says “1 gigawatt,” they can mean two different things, and the press release rarely says which.
IT load (sometimes called “critical IT power”) is the electricity that goes to the computers themselves: the chips, servers, network switches, and storage.
Total facility power is the IT load plus everything needed to keep those computers alive. That includes cooling, as well as the losses from converting grid voltage down to chip voltage, backup systems, lights, and so on. This is roughly what the utility has to supply.
The ratio between the two has a name: PUE, or Power Usage Effectiveness. If a site has a PUE of 1.2, then for every 1 megawatt going to computers, it pulls another 0.2 megawatts for overhead. A perfect site would score 1.0. Newer AI buildings that pipe liquid straight to the chips can get down to around 1.1, while older air-cooled data centers often run 1.5 or worse.
So a “1-gigawatt” campus might have about 900 megawatts of actual computing inside. Or it might have a full gigawatt of computing and need 1.1 to 1.2 gigawatts from the grid. That’s a 10% to 20% swing hiding in one word. When you see a big number, it’s worth asking: a gigawatt of what?
The Speedometer and the Odometer
One more mix-up shows up all the time, even in good reporting: confusing power with energy. A gigawatt is a rate. Think of it as the speedometer. It tells you how fast electricity is being consumed at a given moment.
A gigawatt-hour is an amount. That’s the odometer. It tells you how much electricity was used over a stretch of time, and it’s what actually shows up on a bill.
The math is simple. Run 1 gigawatt for one hour, and you’ve used 1 gigawatt-hour. Run it every hour of the year (8,760 hours), and you’ve used about 8.76 terawatt-hours. For a sense of scale, U.S. data centers of all kinds used roughly 176 to 183 terawatt-hours in 2023, according to Lawrence Berkeley National Laboratory figures. A single 1-gigawatt campus running near full tilt would add about 5% to that.
In real life, no site runs at 100% around the clock, so actual yearly use comes in below the full-tilt figure. But the gigawatt rating tells you the ceiling, and that’s the ceiling utilities have to plan and build for.
So, How Big is a Gigawatt of AI, Really?
Numbers this large are hard to picture, so let’s tie them to a few things we know.
In chips. A single top-end AI chip, including its share of the server, cooling, and networking around it, takes something like 1.5 to 2 kilowatts. A full rack of Nvidia’s current systems runs well over 100 kilowatts. So, in one hour, that rack uses about as much electricity as a typical American home uses in four days. The best real-world example is OpenAI’s flagship Stargate site in Abilene, Texas. Builder Crusoe has said each of its eight buildings holds about 50,000 Nvidia Blackwell chips, per Epoch AI’s site tracker, and the full campus is rated at 1.2 gigawatts. That works out to roughly 400,000 chips, or about one-third of a million chips per gigawatt.
Over a year, a U.S. home uses an average of about 1.2 kilowatts of electricity per hour, or about 10,500 kilowatt-hours in total. So, 1 gigawatt is in the ballpark of 750,000 to 850,000 homes. That lines up with R&D World’s estimate that the 1.2-gigawatt Abilene campus could, conservatively, use as much electricity as 750,000 homes. In other words, one AI campus can need as much power as a mid-sized city.
In power plants. A large nuclear reactor puts out about 1 gigawatt. So a “gigawatt-scale” AI campus needs roughly a whole nuclear reactor’s worth of steady power all to itself.
That last comparison helps explain why tech companies are talking about restarting old nuclear plants, buying natural gas turbines by the dozen, and funding small modular reactors. It’s not just for good PR. They need the juice.
OK, But How Much Computing Power Is That?
Computing power is measured in FLOPS, short for floating-point operations per second. That’s just a count of how many basic math problems the chips can solve each second. Nobody publishes a FLOPS figure for a “5-gigawatt” campus, but you can estimate one in three steps:
- Take out the overhead. At a PUE of about 1.2, a 5-gigawatt campus has roughly 4.2 gigawatts reaching the computers.
- Count the chips. Using the Abilene ratio of about 330,000 chips per gigawatt, that means about 1.6 to 1.7 million of today’s top Nvidia chips.
- Multiply by chip speed. Each current Nvidia Blackwell chip is rated at about 5 petaFLOPS (5 quadrillion operations a second) for the kind of math AI uses.
Put it together, and you get something like 8 to 9 zettaFLOPS. Written out, that’s about 8,500,000,000,000,000,000,000 calculations every second. (These are rough, back-of-the-envelope figures.)
Then you have to take that number with a big grain of salt, for a few reasons:
- It’s a peak, not reality. In real AI training, chips usually only deliver something like 30% to 50% of their rated speed. The rest of the time they’re waiting on memory or on each other.
- Precision changes the answer. The same chip can be rated twice as fast using lower-precision math, and far slower using the high-precision math of scientific supercomputers. That’s why you can’t fairly line up an AI campus against the world’s top supercomputers.
- The chips keep getting better. Fill the same 5 gigawatts with next-generation chips, and you could get several times more computing than with 2023-era chips. That’s exactly why the industry quotes gigawatts instead of FLOPS: the power number stays put while the computing number keeps moving.
- It isn’t one giant computer. A 5-gigawatt project is usually spread across many buildings, and often several sites, running separate jobs.
It’s also worth knowing that a big campus rarely runs on a single kind of chip. These projects take years to build, and chips are ordered only months ahead, so each new building gets whatever Nvidia is shipping when it’s ready to switch on. xAI’s first Colossus cluster in Memphis, for example, started with 100,000 of Nvidia’s older H100 chips in 2024. It now runs a mix of three different chip models spanning two generations, while the newer building next door uses only the latest ones. Since each generation does more math per watt, the last building on a campus can do far more work than the first on the same amount of power. So a “5-gigawatt” campus doesn’t have one fixed FLOPS number. Its computing power keeps climbing over time, even if its power rating never changes.
What Does a Gigawatt Cost?
This is where it gets interesting for investors, because “gigawatts” has quietly become Wall Street’s shorthand for dollars.
Nobody agrees on the exact price tag, and the range is wide:
| Estimate | Cost per gigawatt | What it covers |
|---|---|---|
| Bernstein | About $35 billion | Full build with Nvidia GB200 racks |
| Foxconn | About $47 billion | Full build with Nvidia’s newer Vera Rubin racks |
| Nvidia, 2025 earnings call | $50 to $60 billion | Nvidia’s own all-in figure |
Whatever number you pick, the chips are the biggest slice. Bernstein figures GPUs alone make up about 39% of the spending. And Nvidia CEO Jensen Huang said this month that the Nvidia gear going into each gigawatt has climbed from about $18 billion to more than $40 billion, as Nvidia now sells networking, processors, and storage along with its GPUs, according to a recap of his Bg2 Pod interview.
Then there’s the power bill. Bernstein estimates a gigawatt-scale site burns through roughly $1.3 billion of electricity a year at 15 cents per kilowatt-hour.
Here’s a handy rule of thumb: every gigawatt you see in a headline represents somewhere around $35 to $60 billion of spending. When a company announces 10 gigawatts, it’s talking about a project in the hundreds of billions of dollars. That’s why these announcements move chip stocks, utility stocks, and power-equipment makers like GE Vernova and Siemens Energy.
Announced Is Not the Same as Plugged In
Now for the part the headlines tend to skip. Most of the gigawatt numbers you read are plans, not power that’s flowing today.
Big AI campuses are built in phases, building by building, over several years. Stargate is a good example. The overall project is expected to top 9 gigawatts by 2029. But as of this spring, research group Epoch AI estimated that only about 0.3 gigawatts were up and running at the flagship Abilene site, out of a planned 1.2. Epoch had actually cut that estimate in half from an earlier 0.6. It also noted that plans can change even after shovels hit the dirt, pointing to OpenAI backing out of an Abilene expansion.
None of that means the projects are fake. It means a gigawatt announcement can describe at least four very different things:
- Announced: a plan, a press release, maybe a signed letter of intent
- Contracted: power and land deals are signed
- Under construction: buildings are going up, and equipment is on order
- Energized: the power is actually on, and the chips are working
The gap between step 1 and step 4 can be years, and some projects never make it all the way. So when you see a gigawatt figure, ask three questions:
- Is this IT load or total grid power?
- Is it running now, or planned?
- Over what time frame?
If the story doesn’t say, assume it’s the biggest, most hopeful number available.
Why It Matters to You (Even If You Don’t Own Nvidia)
All those gigawatts have to come from somewhere. That’s where AI stops being a tech story and becomes an inflation story.
Data centers used about 4.4% of U.S. electricity in 2023. The Department of Energy projects that could reach 6.7% to 12% by 2028, as reported by CO/AI. Meeting that demand means new power plants, new transmission lines, and new substations. Utilities usually spread those costs across all their customers.
We’re already seeing it in some places. In the PJM grid that runs from Illinois to North Carolina, new data center demand added an estimated $9.3 billion to capacity costs for 2025 and 2026, according to the Institute for Energy Economics and Financial Analysis. That’s roughly $18 a month more on the average home bill in western Maryland and about $16 more in Ohio, per Yahoo Finance’s summary. Nationally, Consumer Reports found residential electricity prices rose about 7.1% in 2025, and more than 20% in some states.
However, USAFacts looked at the data and concluded that data centers probably weren’t behind price jumps at the national level through 2025, though there’s evidence of local increases. Fuel costs, aging grids, and storm repairs all play a part too. But the pressure is clear, and it’s growing.
Politicians are trying to make the tech companies pay their own way. In March, Amazon, Google, Meta, Microsoft, OpenAI, Oracle and xAI signed a White House “Ratepayer Protection Pledge” promising to build or buy the power their data centers need, along with the lines to connect it. More than 200 utilities, developers and states have since signed on. But it’s a promise, not a law, and the details on how anyone will check the math were never spelled out.
Congress has tried to put it into law. In September, the House passed the Ratepayer Protection Act by a vote of 417 to 3. But that bill only asks state utility regulators to consider making data centers over 100 megawatts cover their own grid costs. It doesn’t require it. The bill then stalled in the Senate, where Democrats want a hard mandate instead. A separate bipartisan Senate deal would require it for any data center of 20 megawatts or more, but nothing has passed yet.
For now, the real rules are being written by the states. Texas, Oregon, and Maryland have passed laws putting big power users in their own rate classes or making them pay to connect. Oklahoma followed this year with a law covering data centers of 75 megawatts or more.
The Bottom Line
A “gigawatt of AI” is a measure of electricity, not brainpower. It tells you how much power a data center is built to draw at full speed. For a quick translation:
- 1 gigawatt ~ one large nuclear reactor ~ about 750,000+ homes ~ roughly 300,000 to 400,000 of today’s top AI chips
- 1 gigawatt ~ $35 to $60 billion of spending
- 1 gigawatt for a full year ~ 8.76 terawatt-hours, or about $1.3 billion of electricity
The next time you see a headline about a company building “5 gigawatts” of AI, you’ll know it’s talking about a power draw the size of a few big cities, a price tag in the hundreds of billions, and a timeline measured in years. And you’ll know to ask whether any of it is actually plugged in yet.
You might also like:
