If an alien landed on Earth and tried to figure out what a data center was, E.T. might arrive at the concept of a big box where zillions of gallons of gasoline gets burned so humans can look at screens.
Data centers and gargantuan energy use have become so synonymous, that, while it’s true data centers consume a lot of energy and demand is rapidly increasing because of AI, it’s easy to lose sight of quantifying how much energy data centers use and why.
These are important questions because data center growth has major implications for the communities that house them, especially in a state with an independent grid and low public trust in the grid’s reliability.
How much power do data centers use?
Data centers vary wildly in size and power use. Prime Data Centers operates two adjacent data centers south of Dallas Love Field. From the outside, they look no different than a small warehouse, nondescript additions to an unremarkable industrial park. They have a combined capacity of 20 megawatts, equivalent to what’s needed to power about 15,000 homes or the total power output of four freight trains.
Another data center campus across Stemmons Freeway from the Hilton Anatole is being developed by CleanArc Data Centers and is planned to have 245 megawatts of capacity, more than ten times larger than the Prime Data Centers duo. This is in the realm of what counts as a hyperscale data center, an unofficial term usually denoting a data center with at least 5,000 servers, 10,000 square feet of floor space and 100 megawatts of power.
Still, data centers can get far larger and more powerful. Microsoft is building a 2 gigawatt facility in Reeves County and Meta is building a 1 gigawatt data center in El Paso. At the extreme end, Fermi’s Project Matador plans up to 17 gigawatts of capacity near Amarillo, about half what Oncor serves all of Dallas, North Texas and beyond.
What’s the difference between power and energy?
Think of energy as a finite substance and power as the rate at which that substance is spent or delivered.
If you have a faucet over a bucket, power, measured in watts, is the amount of water coming out of the faucet at any given moment. Energy is the total volume of water in the bucket collected over time, measured in joules for small amounts and watt-hours for larger amounts. When someone says that an appliance or data center needs X watts of power, imagine a water wheel that won’t move until the water is flowing at a particular rate.
Why do data centers need so much power and how do they affect the grid?
According to a U.S. Congressional report, about half of the power demand of data centers comes from actually operating the servers. Much of the rest comes from cooling the servers; just like your laptop heats up then stalls when too many tabs are open, the servers in a data center must be continuously cooled to undertake complex tasks like training AI models.
Data centers aren't just power-intensive because they’re big buildings, though. On a Dallas Cowboys game day, AT&T Stadium can demand 10 megawatts of power to air-condition 80,000 seats, light up a 600-ton jumbotron and shine floodlights on a football field. That’s only half of what Prime Data Centers’ small Dallas campus needs and only some of the time, which gets at one of the problems with data centers: they need constant power and a lot of it.
In August 2025, electricity provider Oncor, which has a peak power capacity of 33 gigawatts, said it had a queue of 200 gigawatts of power’s worth of interconnection requests in its queue, of which 186 gigawatts were from data centers. In the first quarter of this year, the data center queue alone had increased to 271 gigawatts.
The other problem with data centers is that the power they actually draw from the grid fluctuates in ways the grid isn’t built to facilitate, said Hao Zhu, an associate professor of computer and electrical engineering at the University of Texas at Austin.
“Data centers have a lot of power fluctuations, and this is a destabilizer to the power grid because they have very fast cycles, and very fast power fluctuation disrupts the grid,” Zhu said. “The impact of data centers, especially hyperscalers, on the grid is multifaceted. There is the demand aspect, but they are also a special type of load.”
How much energy does AI use?
Data centers need more and more power because AI needs more and more energy.
Many of the most popular AI products — like OpenAI’s ChatGPT, Anthropic’s Claude and Google’s Gemini — are closed-source models, meaning the public doesn’t know what happens once a query is routed to a data center, making actual energy use impossible to calculate. However, an MIT Technology Review team tested the energy demands of querying various open-source AI models to provide a rough estimate of what it actually takes to complete AI tasks.
For text-generating answers, the team used Meta’s Llama. The version of Llama 3.1 with 8 billion parameters, all the information the AI “knows” and uses to generate answers, used on average 57 joules of energy to respond to a prompt. The largest Llama model with more than 400 billion parameters used about 3,400 joules.
Generating a 1024 x 1024-pixel image with Stable Diffusion 3 Medium, a leading open-source image generator with 2 billion parameters, used about 1,100 joules of energy, with better image quality scaling linearly (a doubly more detailed image required double the energy, etc.).
These numbers only measure the energy used by the GPU chips where a model is housed; accounting for other energy usage, like cooling and the CPUs that tell the GPUs what to do, usage estimates are about double. For context, one AA battery holds about 10,000 joules of energy. With 10,000 joules of energy, you can power a 1,000-watt appliance, like a microwave, for 10 seconds.
Generally, the better the model the more parameters, which means more energy spent per task. Frontier models like OpenAI’s GPT-6 Astra and Anthropic’s Claude Fable 5.1 are reported to use well into the trillions of parameters. Parameter estimates for consumer models like Claude Sonnet 5, Anthropic’s default model for free prompting, are around 1 trillion.
By far the most energy intensive AI task is generating video. High-definition videos are generated by expensive, closed-source models, but even just a five-second video at 16 frames per second (the frame rate of old silent movies) spent about 3.4 million joules using the open-source CogVideoX model, per MIT Technology Review. Converting to larger units, that is almost exactly one kilowatt-hour, AKA the unit energy companies charge you by.
Why does AI need so many more data centers than the internet, which we’ve had for decades?
Put simply, AI needs more from data centers than most other internet-based tasks, and demand is exploding.
Sending and receiving an email takes about 300-3,000 joules of energy, not even the amount of energy stored in one Tic Tac. A Google search can be about the same. One of the most energy intensive non-AI tasks one can do with the internet is stream video. An hour of Netflix uses about 300,000 joules or 0.08 kilowatt-hours of energy, according to analysis published by the International Energy Agency, with a lot of variation based on video quality, what device is streaming the video and internet source.
So, one hour of Netflix is roughly as energy-intensive as asking ChatGPT 100 basic questions. However, compared to AI, where most of the energy is spent at the data center by servers, most of the energy on average spent streaming Netflix is spent by the screen, usually your TV.
And while Netflix watching increases by a steady, but a few percentage points each year, AI usage has increased 10-fold every year since 2024 to now at least 300 trillion tokens a day (tokens being the chunks of text AI processes).
Much of this is because of the proliferation of enterprise use and agentic AI, which uses orders of magnitude more tokens than asking ChatGPT to make travel itineraries or Dall-E to make a cat meme. The legal AI startup Harvey, which fine tunes models from providers like OpenAI for use by major law firms, used 12 trillion tokens in May 2026, up from 1 trillion in January, and many other similar products have cropped up in recent years.
Another reason AI models need more power and more data centers: we keep making more of them.
“Frontier model training is a tremendous computational task, which is why GPUs are hot merchandise now,” professor Zhu said. “We have a shortage of GPUs, but we also have even more shortage of electric power. AI infrastructure is growing at an exponential rate, and all the top companies are looking for bigger, better models.”
It’s estimated that training OpenAI’s GPT-4, which was released in March 2023, consumed 50 gigawatt-hours of energy, roughly equivalent to powering 50,000 American homes for a month. Since the first quarter of 2025, major AI labs — including OpenAI, Anthropic, Google, Meta and more — have combined to release about 20 new AI models a quarter, according to AI Release Tracker.
How should the Texas grid respond to data centers and AI?
The queue of just data centers hoping to hook into Oncor is more than double the available power for the entire state, but the AI industry shows no signs of slowing down, and that’s where the problem for the grid is, Zhu said.
“If Texas’ power demand continues to grow because of population increase, because of more manufacturing coming to Texas, I used to tell my students, it's okay, because we can plan for that,” she said. “I stopped saying that to my students this year, because that is not the case with AI data centers. They're looking to connect to the grid tomorrow, not in two years, because they don't want to lose the competition with their competitors.”
Data centers bring a host of potential power-related side effects to the communities they’re in, Zhu explained. When they connect to the grid, they become competitors in the marketplace for energy, which could raise prices as has already been seen in data center-heavy states like Virginia and Maryland. Costs could fall on taxpayers through the price of building additional transmission lines and other infrastructure, even as the state takes steps to alleviate those costs. If a data center causes a shock to the grid, such as through suddenly disconnecting to protect its servers, its section of the grid may need to be isolated, which could block power to the homes and buildings nearby.
The energy and power needs of data centers and AI are evolving though, with chips getting more efficient and data centers innovating on how to cool servers with less energy.
Massive data center projects are also exploring behind-the-meter power, meaning building their own plants to provide a direct source of electricity, much like an off-the-grid home might have their own solar panels and a diesel generator. Fermi’s Project Matador, among other sources, even plans to have on-site nuclear reactors to power up to 18 million square feet of server space.
These innovations still aren’t happening quick enough to meet today’s AI demand, so other strategies are needed, Zhu said. This includes research on the unique ways data centers impact the grid beyond simple power needs and coordination of how data centers can reduce their power demands at times when the grid is strained.
The most important thing, though, is to get more power generation online.
“We need resources that can be deployed quickly. That includes gas turbines, that could also include renewables, because they're much faster than putting in a nuclear plant or adding a new hydropower plant,” Zhu said. “There is a huge need to grow our generation capacity. It's much slower, it's outpaced by AI.”