Yesterday I sat down with someone who works for one of the larger electric utilities here in the Midwest. Nothing to do with AI, just a personal conversation. When I mentioned what I do for a living, he said something that stuck with me. Everybody gets up in arms about AI data centers straining the power grid, he said, but almost nobody talks about the fact that manufacturing draws more electricity nationally than every data center in the country combined.
I wanted to check that before I repeated it to anyone. Here is what I found.
Manufacturing employs about 600,000 people in Michigan, more than any other single industry among the state's medium and large employers, according to the state's own quarterly employment data. When a data center gets proposed near a Michigan town, the pushback about the power grid gets loud fast. When a stamping plant or an injection molder that has run the same equipment for decades adds a shift, nobody blinks about its electric bill. Both draw from the same grid.
What the electricity numbers actually say
The US Energy Information Administration puts total US electricity consumption in 2025 at about 4.2 trillion kilowatt hours. The industrial sector, which the EIA defines as manufacturing, mining, construction, and agriculture, accounts for roughly a quarter of that. Manufacturing alone makes up about 71 percent of the industrial total, according to the EIA's own 2026 Annual Energy Outlook, which puts manufacturing's share of the entire country's electricity use at somewhere around 18 percent by itself.
All the data centers in the country combined used about 224 terawatt hours in 2025, or roughly 5 percent of total US electricity, according to the Department of Energy's Lawrence Berkeley National Laboratory.
Run the numbers and manufacturing alone draws somewhere around three and a half times what every data center in the country uses combined. Count the rest of the industrial sector too, mining, construction, and agriculture, and it is closer to five times.
The utility guy was right, and by a wider margin than either of us guessed sitting at that table.
So why does AI get all the attention
Part of it is visibility. A new data center campus announces itself. It shows up in headlines, in local zoning fights, in utility rate case filings. A stamping plant that has run in the same spot in Saginaw or Bay City for forty years does not generate the same kind of news. Part of it is growth rate. Data center electricity use is climbing fast, from about 224 terawatt hours in 2025 toward a projected 606 terawatt hours by 2030, and fast growth draws attention even from a smaller base. But growth rate and share of the grid are two different measurements, and most of the public conversation blurs them together.
That same blurring shows up in the specific number people repeat about individual AI use, so it is worth checking that one too.
Where the ten times number came from
The claim traces to a 2023 estimate from data scientist Alex de Vries, who calculated that a ChatGPT query cost about 3 watt hours of electricity. He built that number on OpenAI's earliest GPT-3.5 model, running near maximum power draw on older Nvidia A100 servers, and assumed a query length of roughly 4,000 input tokens and 2,000 output tokens, close to 1,500 words. Compare that against the cost of a Google search, a figure Google itself put at 0.3 watt hours in a 2009 blog post from senior vice president Urs Holzle, and you land on a ten to one ratio. Goldman Sachs repeated the figure in a 2024 research report. Alphabet's own chairman, John Hennessy, told Reuters an AI exchange probably costs about ten times what a standard search does. The comparison stuck, and it is still the number most people repeat today.
What a query actually costs now
Two model generations and a lot of hardware improvement later, that math does not hold up on its own terms. In 2025, the research group Epoch AIÂ rebuilt de Vries's calculation using GPT-4o's real architecture, a more realistic power draw assumption of 70 percent of peak capacity instead of near maximum, and token counts that reflect how people actually use the tool instead of a worst case scenario. Their result: a typical ChatGPT query costs about 0.3 watt hours, roughly ten times lower than the original 3 watt hour estimate. OpenAI's own CEO, Sam Altman, has put the number at about a third of a watt hour too, comparing it to the electricity an oven draws in a bit over a second.
Google went further than either of them. In August 2025 the company published the first technical paper of its kind, laying out exactly how it measured the energy, carbon, and water cost of a Gemini prompt from the AI chip through the cooling system. The result: a median Gemini Apps text prompt uses 0.24 watt hours, emits 0.03 grams of carbon dioxide equivalent, and consumes 0.26 milliliters of water, about five drops. That is roughly what a microwave draws running for one second, or a television running for nine. Google also reported that the energy cost of a median prompt dropped 33 times over the trailing twelve months as its models and hardware improved.
Not every estimate lands that low, and I am not going to pretend the science is fully settled just because the flattering numbers come from the companies selling the product. Independent researchers working from public hardware specs instead of company disclosed data still land higher. One widely cited 2025 analysis combining several independent studies averaged closer to 2 watt hours per query, which puts ChatGPT at roughly seven times the energy of a Google search rather than ten. Better than the original claim, but not the tidy near zero story some AI companies would prefer. The honest range today runs from about a quarter of a watt hour to two watt hours, depending on whose methodology you trust.
What that looks like in real terms
Even at the higher end of that range, a single AI query is not a meaningful draw against anything else you do in a day. At Epoch AI's 0.3 watt hour estimate, one query is roughly what it costs to charge your phone on a standard 5 watt charger for about four minutes, or to run a single LED bulb for a couple of minutes. Flip it around: to match the electricity a 60 watt bulb burns left on for one hour, you would need somewhere between 30 queries, using the higher 2 watt hour estimate, and 200 queries, using Epoch AI's lower one.
What this means for your business
If you run a manufacturer, a nonprofit, or a church office in Michigan and you are weighing whether using ChatGPT or Gemini for a first draft, a summary, or a training document is somehow environmentally reckless, it is not, not compared to what your own building or your own production line already pulls from the same grid. The scale question belongs at the level of national grid planning and data center construction, not at the level of whether you asked an AI tool to help write a policy memo. If anything, the equipment running your production floor is doing more of the heavy lifting on that meter than the chatbot open in the next tab.
Bottom line
Manufacturing is still the heavier load on the American power grid than artificial intelligence, and it is not close. Michigan's own economy, built on the assembly line more than the server rack, is proof of what that looks like at the state level. The specific claim about your ChatGPT query burning ten times what a Google search does was true in 2023, on a worst case estimate, and it has not kept pace with how efficient the models have gotten since. AI's energy footprint deserves a serious conversation. Serious conversations need the whole grid, not just the newest part of it.