The Power Bill Behind the Prompt

Every AI prompt has a physical cost. In the Tennessee Valley, that cost runs through the same rivers, dams, and power lines that have kept our lights on for ninety years. Here is what it means for your power bill, your water, and your community.

It started with our own power company

In May 2025, BrightRidge, the public utility that powers much of Johnson City and Washington County, did something quiet but telling. It paused all new data center projects. Not because of one bad actor, but because the people running our grid wanted to understand the cost before they signed up for it.

That pause is a good place to start, because it captures something most of us miss. The boom in artificial intelligence is not just happening in California or somewhere in the cloud. It is reaching for the power and water right here in the Tennessee Valley. The Tri-Cities started asking hard questions before a single large AI data center was ever built here.

This is my attempt to lay out what that physical cost actually is, in plain terms, without hype and without hiding the parts that worry me.

The cost you cannot see

When you type a prompt into an AI tool, nothing about it suggests a physical cost. There is no meter, no gauge, no sound. But every answer is generated in a building full of specialized computers that draw real electricity and, in many cases, real water to stay cool.

Here is the honest part: nobody outside the AI companies can tell you the exact cost of a single prompt. The companies do not publish the details. A 2025 review by Stanford scored the industry 40 out of 100 on environmental transparency, and 10 of 13 major AI companies disclosed nothing at all about their energy or water use. So the numbers below are careful outside estimates, and they vary widely. That uncertainty is part of the story.

For a simple text prompt, the estimates land in a tight range. A standard Google search uses about 0.3 watt-hours. A simple AI prompt uses roughly 0.24 to 0.34 watt-hours, about what a high-efficiency lightbulb uses in a couple of minutes. Small. Almost nothing.

The picture changes fast with newer, more capable models. Researchers measuring GPT-5 in 2025 found it averaged around 18 watt-hours for a longer response, and could reach 40 watt-hours for a heavy, multi-step task. That is the same single prompt costing roughly fifty times more, simply because the model is bigger and works harder.

Water follows the same pattern, and it has a hidden layer. The water you can measure is what evaporates on-site to cool the servers, which OpenAI puts at a tiny fraction of a teaspoon per prompt. But there is a second, larger cost most people never see: the water used at the power plant to generate the electricity in the first place. That indirect water use is often two to four times the on-site amount. How you make the power decides how much water the prompt really costs.

One prompt is nothing. A billion is a power plant.

The reason any of this matters is scale. One prompt is trivial. Billions of them per day is a different kind of number.

• Global data center electricity use is projected to more than double to around 945 terawatt-hours by 2030, close to the entire electricity use of Japan, with the IEA naming AI as the most important driver of that growth.
• One recent estimate puts AI-related data center water use in the hundreds of billions of liters a year, though estimates vary widely depending on whether they count only on-site cooling water or also the water used to generate the electricity.
• A single large data center can use up to 5 million gallons of water a day, about the same as a town of 10,000 to 50,000 people.

Those are the kinds of numbers that turn an invisible click into a physical footprint with a place on the map.

Why the Tennessee Valley is on the map

This is where our region comes in. Data center operators want power that is cheap, reliable, and available around the clock. The Tennessee Valley Authority happens to run exactly that kind of grid.

TVA's largest single source of electricity is nuclear, followed by natural gas, coal, and the hydroelectric dams along our rivers. Nuclear and hydro together form a large part of TVA's firm, low-carbon backbone: power that can run day and night without depending on wind or sunlight.

That combination of nuclear reactors and the dams along our rivers is a big part of why the Valley is attractive to AI builders. As TVA's own spokesman put it, they do not recruit these data centers. The companies come anyway, because the reliable, low-cost power is already here. The rivers and dams that built this region in the 1930s are now part of why the AI industry is looking our direction in the 2020s.

The bill is starting to land here

The demand is not theoretical. Data centers grew to 18 to 20 percent of TVA's industrial power load in 2025, and TVA expects that to roughly double by 2030.

To keep up, TVA is building about 6,200 megawatts of new generation, the largest construction program in its history. Some of it is clean: a new small modular nuclear reactor is moving forward at the Clinch River site in Oak Ridge, backed by a federal grant. But some of it is not the clean story we might hope for. In February 2026, citing this fast-growing demand, TVA reversed plans to retire two coal plants and decided to keep them running. Growth has a cost, and right now part of that cost is more fossil fuel.

There is also a fairness question, and both TVA and the state have started to answer it. TVA is considering a separate rate classification and upfront capacity-related charges for data centers, so the cost of serving them is not shifted onto the 10 million people who already pay TVA bills. A new Tennessee law moves in the same direction. For covered data center agreements entered into, amended, or renewed on or after January 1, 2027, municipalities and electric utilities generally cannot pay or absorb the electrical infrastructure costs for data centers projected to draw 50 megawatts or more during their first three years.

We do not have to look far for a warning. In the broader TVA region, Memphis and nearby Southaven offer the cautionary case. xAI's Colossus projects have faced lawsuits and regulatory scrutiny over the gas turbines used to power their AI operations, with the NAACP alleging unpermitted turbine operations and Clean Air Act violations affecting nearby historically Black communities. Water concerns have grown too, after a promised water-reuse project was reportedly paused. That is the same TVA region we live in, and it is the cautionary tale our local officials are clearly trying to avoid.

To be fair, the picture is not all alarming. At least one analysis found no clear link between data center growth and higher statewide power rates, and some economists argue that large new customers can actually lower everyone's rates over time by spreading fixed costs. The honest answer is that it depends entirely on the rules we set now.

My honest take

I build my business around AI. I use these tools every day, and I believe they save people real time and do real good. So when I raise concerns, it is not to scare anyone away from the technology. It is because I want to be as transparent about AI as I can be, including the parts that are uncomfortable.

And here is the uncomfortable part. The power and water demands of these data centers are a real problem, not a talking point. I do not think we should pretend otherwise. My honest view is that we rushed this technology out faster than we built the hardware and infrastructure to run it responsibly. We scaled the demand before we scaled the clean power, the efficient chips, and the cooling systems that would let us meet it without straining local grids and aquifers.

The hopeful part is that better technology can ease a lot of this. More efficient chips, smarter cooling, and a grid that leans harder on nuclear and hydro can shrink the footprint of each prompt over time. But that only happens if we are honest about the problem first, and if our region keeps asking the questions BrightRidge started asking in 2025.