Four percent. That is the share of American adults who think a data center is mostly good for the environment. On home energy costs, the share seeing a mostly good impact is also 4%. On life next door, 4% again. Not one small group answering all three questions the same way. Four percent on each question.
Pew Research Center published those findings on September 22, 2026, from a survey of 10, 548 adults taken between July 20 and August 9. The gloom was already there in January. By late summer it had thickened, and it had thickened across demographics.
The cloud grew a street address. A lot of people do not want it on their block.
That is the political fact under the AI boom. In the 2026 midterms, data centers have become something candidates would rather not claim. The Guardian’s coverage has cast the buildings, the AI they serve, and public distaste for the billionaires behind them as the reason big tech turned toxic, while still treating the economy as the central fight and naming AI as one of six voter issues. Salience is not rank. For anyone selling AI for business, the practical distinction is sharper. A technology can lose the permit hearing without deciding the presidency.
The product grew a footprint
For a few years, commercial AI lived in a browser tab. A chat window. A promise that routine work would cost less, and that the cost of running the model in production would stay someone else’s problem. Abstract, portable, easy to demo.
Then the industry started pouring concrete. Pew’s April 13, 2026 short read noted that most data centers being planned or under development will sit in rural areas. That is where the argument stops being about model weights and starts being about substations, water, tax deals, and the view from a kitchen window. You cannot put a language model on a yard sign. You can put a data center on one.
Abundance is a lovely word until it shows up as a line item on a power bill.
What people actually believe
Pew asked about five impacts: the environment, home energy costs, quality of life nearby, local jobs, and local tax revenue. On the first three, “mostly bad” is now about half the country, and it got there in a single season.
The share calling data centers mostly bad for the environment rose from 39% in January to 54%. Mostly bad for home energy costs rose from 38% to 50%. Mostly bad for quality of life nearby rose from 30% to 49%. Only 4% see a mostly good impact in each of those three areas.
Jobs and tax revenue stayed split. Similar shares, between 21% and 24%, call data centers mostly good and mostly bad for each. “Mostly bad” for local jobs rose from 15% to 24%, and for local tax revenue from 12% to 21%. A real souring. Not a majority verdict.
On all five questions, between 18% and 32% are unsure. Twelve percent say they have not heard of data centers at all.
The shift shows up where the buildings are headed. Among rural adults, “mostly bad” for the environment moved from 32% to 50%; for home energy costs, from 32% to 46%; for nearby quality of life, from 28% to 45%.
Party did not freeze the numbers in place. Democrats and Democratic-leaning adults remain more negative. Republicans moved too. Among Republicans and Republican leaners, 42% said data centers are mostly bad for quality of life nearby, up from 24% in January.
Read that as a map, not a eulogy. Voters are far more convinced of environmental, energy, and quality-of-life harm than they are convinced the promised jobs and tax base will fail to show up. A large share still has no firm view. That leaves an open argument. Nobody has handed the industry a blank check.
A president and a governor can disagree
National rhetoric and state action have already split, and they should not be stapled together as cause and effect. Pew, citing Politico and CNBC on August 31, 2026, says President Trump has consistently supported data-center construction. Pew, citing Politico on August 23, reports that Texas governor Greg Abbott, previously a backer of expansion, issued a statewide moratorium on new data centers in August.
Those are parallel facts. The Pew survey was in the field from July 20 to August 9. A late-August pause cannot explain a shift already measured. Public opinion had turned. A Republican governor in a major state then hit pause, while the president was still for the build-out.
If a siting plan assumed red states were a safe harbor, August was a correction.
The resistance is not only a campaign mood. Pew points to protests against new construction, and to public trackers of state and local responses kept by the National Conference of State Legislatures and the National League of Cities. The rules will not be the same in every state. A single federal mood is not a permit.
Checks are not permits
While the polling moved, some of the money behind the industry kept moving too. Americans for Tax Fairness Action Fund, reported by Common Dreams on October 7, 2026, identified the Andreessen and Horowitz families as the top-spending billionaire families this cycle, at a combined $149 million, with checks going to both Republicans and Democrats. That is family political spending, not a clean tally of AI-industry donations, and it should not be stretched into one.
The same reporting says Marc Andreessen and Ben Horowitz have given major donations to Leading the Future, a super PAC that opposes strict artificial intelligence regulations. That shows interest, and a legal route for unlimited outside spending. It does not show a winner. This cycle has not proved that the checks decide the permit, the primary, or the map.
What changes if you buy or sell the technology
The board question is whether the building that trains and serves the product has become a political liability the vendor does not fully control, and that the buyer inherits anyway. Cable-news unpopularity is beside the point.
Power is the engineering constraint and the political exposure at once, and those are different facts. Pew’s energy-cost numbers show what adults believe about the bill at home. They do not, by themselves, prove what a given facility does to rates, or what it costs to run a model in production. Half of adults already think these buildings are mostly bad for that bill. Communities that expect to pay more will not applaud a clever demo. The political risk is real even if the cost assumption in the spreadsheet is still true.
Incentives age badly in public. A tax break that read as economic development two years ago can read as a transfer once electric bills and farmland enter the argument. Assume every abatement can become a campaign issue. And do not treat a headline about a pause as a finished policy. If the text is not in hand, it is not a plan.
The jobs story is still contestable. The people to contest it with are not the 4% who already see a mostly good impact. They are the unsure share, between 18% and 32% depending on the question, and the much larger group that expects harm on energy, land, and daily life. Named hires, named tax payments, named water and power terms, and a community agreement a county commissioner can defend are the evidence that might move them. A rendering and a promise of abundance will not.
Brand risk travels into the sales motion. The midterms framing has treated distaste for the billionaires behind the build-out as part of the souring. If the pitch for an AI agent is “trust the lab, ” it borrows a reputation the seller does not control. The safer face is operational: what the system does, what it costs to run, who is accountable when it fails, and where the electricity comes from. That is true whether the buyer is automating a back office or a sales team. The customer still has to live next to the story.
Plan for a split map. A president can want the buildings. A governor in his own party can freeze new ones. Local protests can multiply while national politics keeps the economy on top. A data center can still be the hardest item on a town agenda. Those facts can all be true in the same quarter.
Questions worth answering before the next site plan
-
Do Americans think data centers help or hurt?
On the environment, home energy costs, and nearby quality of life, the hurt side is ahead. Pew’s late-summer survey put “mostly bad” at 54%, 50%, and 49%, with only 4% seeing a mostly good impact in each area. On local jobs and local tax revenue, the public is split.
-
Is the backlash only Democratic?
No. Democratic leaners remain more negative, but 42% of Republicans and Republican leaners now say data centers are mostly bad for life nearby, up from 24% in January.
-
Are voters picking candidates mainly on AI?
Not on the evidence in hand. The economy remains the central fight of the cycle, and AI is visible enough to sit among the named issues without topping them. Data-center anger can decide a permit fight without deciding the presidency.
-
Does the spending prove the industry bought the election?
No. It shows a route, not a winner. Andreessen and Horowitz have backed Leading the Future, a super PAC opposed to strict AI rules, and their families’ combined $149 million in political spending this cycle went to both parties. That is not a verified AI-industry total, and it does not say who wins in November.
-
What should an operator change now?
Treat power, water, and local tax deals as political risk, not as a footnote to the product. Twelve percent of adults have not even heard of data centers yet, which means the reputation is still being written, and it will be written locally, one site at a time.
The industry spent years telling people the gains would be general and the costs would be invisible. The costs acquired an address. Some of the people writing the checks are still betting that money, and a friendly White House, can outrun that. The polling from July and August says the outrunning has not happened. If you sell the product, you inherit the building’s politics whether or not you own the building. Stop selling inevitability, and start showing the bill.