I’ve spent too many years staring at power bills, cooling loops, and capacity tickets to pretend every AI prompt is personally boiling the oceans. That kind of slogan is great for clicks and terrible for capacity planning. The real story is less theatrical and more operational: data-center load is growing fast enough that utilities, grids, and neighbors are noticingand the public opinion numbers are catching up to the megawatt math.
The old stub on this URL invented a cute “poll with neat percentages and a punchline about ChatGPT as your intern. I’m not recycling invented survey numbers. We have a real one.
What Americans Actually Told Pew
In January 2026, Pew Research Center surveyed 8,512 U.S. adults about data centers—buildings full of the gear that runs streaming, banking, cloud apps, and, yes, AI. Among people who had heard of them, the mood on the environment was not warm:
- 39% said data centers are mostly bad for the environment; only 4% said mostly good.
- 38% said they are mostly bad for home energy costs; 6% said mostly good.
- 30% said they hurt nearby quality of life; 6% said the opposite.
- Views flipped more positive on local jobs (25% mostly good vs. 15% mostly bad) and tax revenue (23% vs. 12%).
That is the poll nobody in a ribbon-cutting hard hat wants on a slide deck. The public will tolerate a warehouse full of servers if it brings payroll and property tax. They get less charitable when the same facility shows up in their summer power bill narrative—or when local media covers water for cooling next to a drought map.
The operational takeaway is simple: awareness is up, and the default story is load that someone else is paying for.”
Read the survey yourself: Pew — How Americans view data centers’ impact.
What the IEA Says the Load Actually Looks Like
Opinion is downstream of physics. The International Energy Agency’s Energy and AI report put hard numbers under the hype:
- Data centers used about 415 TWh in 2024—roughly 1.5% of global electricity.
- In the IEA Base Case, that roughly doubles to ~945 TWh by 2030, still under 3% of global electricity but growing ~15% a yearseveral times faster than everything else.
- A typical AI-focused site can draw as much power as ~100,000 households; the largest campuses under construction are described as roughly 20x that.
- Emissions from data-center electricity use rise from about 180 Mt today toward ~300 Mt by 2035 in the Base Case (higher in a Lift-Off case)still a small slice of energy-sector emissions, but one of the fastest-growing slices.
- In the United States, data centers account for nearly half of projected electricity demand growth to 2030 in IEA’s framing—enough that by decade’s end the country could use more power for data centers than for aluminum, steel, cement, chemicals, and other energy-intensive goods combined.
IEA also notes the uncomfortable local truth: globally the share looks manageable; in a handful of U.S. clusters it does not. Nearly half of U.S. data-center capacity sits in five regional clusters. Grid queues are long. Transformers and cables have multi-year lead times. The agency estimates that without mitigation, around 20% of planned projects could face delay risk.
Primary source: IEA Energy and AI, executive summary.
U.S. Power Demand Is Not Flat Anymore
You do not need a white paper to feel this if you watch Short-Term Energy Outlook headlines. The U.S. Energy Information Administration has been forecasting record electricity use into 2026-2027, with data-center development (plus manufacturing and electrification) as a named driver. Commercial load—where standalone data centers live in the accounting—is no longer a sleepy line item.
EIA’s longer Annual Energy Outlook work also flags server electricity as a rising share of commercial consumption. Flat end-use load shapes (servers that draw hard around the clock) are exactly the kind of customer that makes a utility’s planning spreadsheet sweat: you cannot “wait for the evening peak” if the evening peak never leaves.
For a plain-English EIA note on server energy in the commercial stock, see EIA Today in Energy on data center server electricity. For the short-term demand narrative, Reuters summarized EIA’s STEO framing that AI-driven data centers are helping push U.S. power use to new highs: Reuters on EIA power-demand outlook.
Water, Cooling, and the Part of the Ticket Most Dashboards Hide
Electricity gets the press. Water is the quieter constraint. Cooling towers turn megawatts into gallons. Analyses such as Brookings look at AI energy demand cite U.S. data-center water use on the order of ~17 billion gallons in 2023, mostly at hyperscale and colocation sites. Disclosures still vary by methodology—if you do not measure WUE the same way, you cannot compare vendors in a bake-off.
I used to review “green” claims like backup-generator test logs: politely, then with a red pen. Annual renewable certificates are not the same as sparing a local summer peak. Carbon-aware scheduling, workload shifting, and honest siting are the grown-up conversation.
Policy Is Catching Up to the Cable Tray
Regulators already treat interconnection like a scarce resource. Pause-and-study cycles happen when the load queue outruns generation and transmission—capacity management with a political overlay, not anti-tech theater.
For IT and cloud buyers, the homework is overdue:
- Ask for location and carbon intensity of inference, not just training PR. Training is the fireworks; inference is the always-on baseline.
- Prefer vendors who publish PUE, WUE, and hourly renewable matching—or admit they do not.
- Right-size models. A 70B-parameter model for rewriting a meeting invite is the energy equivalent of spinning up a mainframe to balance a checkbook.
- Track Scope 2 like an SLA. If your board cares about emissions, treat provider region choice like latency: measure it.
The Honest Middle
IEA is careful here, and so am I. Data-center emissions are not “cooking the planet by themselves relative to the whole energy system. Pretending AI will single-handedly solve climate change is equally silly. Existing AI applications could unlock material efficiency gains in industry, grids, and buildings if data, incentives, and cybersecurity allow adoption. Rebound effects are real. Silver bullets are marketing.
From Mason, Ohio, my driveway does not host a hyperscale campus. My power bill still lives in a regional market shaped by whoever is adding firm load next door. The Pew numbers say neighbors notice. The IEA numbers say the load is real. The useful response is not guilt about every prompt. It is treating compute like the industrial process it has become: site it where the grid can feed it, cool it without draining the wrong aquifer, measure it honestly, and stop confusing convenience with free energy.
Steve Miller — Mason, Ohio. Former programmer, sysadmin, and IT manager. Still asks for the PUE before the press release.