Executive Summary

Figures in this article are current as of November 4, 2026 (KST). Quinnipiac numbers were read off the original release PDFs and the official trend tables issued with them.

What Axios reported on November 1, 2026 was a proposal. The relevant passage says OpenAI "will argue that the infrastructure it's building, in contrast to legacy data centers, should be called 'AI compute campuses.'" There are two verbs stacked there. It will argue, and the facilities should be called. The report does not say an internal style guide was rewritten or an official name was swapped out.

The proposal did not arrive on its own. The policy push Axios described has four strands: legislation requiring mandatory safeguards on the most capable AI systems, a data center bundle, support for workers in transition, and international AI standards. The naming argument is the third item inside the second strand. The two items ahead of it in that same strand are community benefit agreements and state legislation giving residents a voice in how projects get built and what they receive in return. A demand to rename the buildings and a demand to widen residents' authority travel in one envelope. Either demand read on its own misrepresents the package.

An OpenAI spokesperson explained the reasoning in a statement to Business Insider: "Compute campuses better describes the scale and role of the infrastructure we're building for AI. These are long-term investments in computing, energy infrastructure and skilled work." This statement, too, stops short of saying a decision was made. It says the phrase describes things better. So far the modality of the report and the modality of the statement agree.

A day later, though, one notch moved. On November 2, 2026, OpenAI used the phrase in its own release about the electricity grid, the Grid Protectors Program. One sentence in that release reads:

Where the new phrase first landed in a company document is worth recording: a sentence about electricity bills. When Quinnipiac read three options to respondents who said they opposed such a facility in March 2026 and asked why, electricity costs drew 72%, water use 64%, and noise 41%. Multiple answers were allowed, so the total runs past 100. There is no evidence the sentence in the release was written with those options in view. What can be observed is that the new phrase and the hottest reason for opposition ended up in the same sentence.

The narrowest claim this article can make therefore has two steps. What was reported on November 1 is a plan to argue for the name, and on November 2 OpenAI began using it in a release of its own. No official name for all data centers has changed, and the subject is the company's own new construction.

The words OpenAI wants to drop are also words OpenAI has been using. Sachin Katti, who leads compute infrastructure at the company, wrote on an executive account in March 2026 that "Our flagship Stargate site is one of the largest AI data center campuses in the United States," referring to the site in Abilene, Texas. "Campus" was already there. So was "data center" in front of it. What this proposal would change is one word in that name.

How the verb for a single event wobbles as it crosses from outlet to outlet also deserves recording. In an article about the force of wording, counting the reporting verbs comes first. Below are four renderings of the same event.

One event, and the verb slides from intends to argue, to plans, to is considering, and finally to settled. The bottom row is not somebody else's problem. That is how the one-line brief handed to this report was written, and it was corrected only after the Axios original was opened. The body of this article uses the top row, the narrowest of the four.

The timing is stated too. Chris Lehane, who runs global affairs at OpenAI, said the company would press these proposals starting with the lame-duck session that convenes on November 9, 2026. Of the current opposition he said, "We think the conversation is good and healthy. It's democracy in action." He added, "You're going to see us talk more and more about the concept of speed limits," and of the next presidential cycle, "AI is going to be the issue in 2028."

The question Quinnipiac used throughout 2026 is a single line: "Would you support or oppose the building of an AI data center in your community?" That one line contains the noun phrase "an AI data center" whole. It is exactly the phrase OpenAI wants changed. To be precise, the part OpenAI would strip is "data center," not "AI," and "AI" survives in the proposed name.

The house asked twice nationally and twice in Pennsylvania. Within each series the sentence stayed put, and only between the two series is there a slight difference. That is why there are two lines rather than one. The two tables below take their values from the trend tables Quinnipiac publishes itself. Sample sizes and the margins of error that follow from them differ a little between waves, but the sentence asked, the mode of interview, and the weighting procedure do not. This is not an online panel: it is random-digit dialing across landlines and cell phones, with live interviewers asking the question.

2.1Line one: national adults

The two rows below are March and September 2026, six months apart. The population at both points is national adults, and the sentence asked is the single line quoted above. What changed is the sample size and the margin of error that comes with it. Three things have to hold before the word trendline means anything: the same house, the same sentence, the same population. This table is a case where all three hold.

Quinnipiac national survey. The population at both waves is adults. Figures are percentages.

2.2Line two: Pennsylvania registered voters

The same house asked twice in Pennsylvania as well. The population here is registered voters rather than adults, and the two points are February and July. Opposition rose from 68 to 74 and support fell from 20 to 16. The direction matches the national line above. Because the thing being measured differs, though, the two lines must not be put on one axis and subtracted. There are two lines because there were two groups of people asked.

Quinnipiac Pennsylvania survey. The population at both waves is registered voters. The state question differs slightly from the national one: nationally "the building of an AI data center," at state level "building an AI data center." No evidence suggests that difference moved the numbers. What it does show is that even one house cannot hold its wording perfectly still.

2.372 and 73 are two populations from one interview

Coverage of this story circulates 72% and 73% side by side. Both are real values, and both come from the same interviews, conducted September 24 to 27, 2026. What differs is which column was read. Count all 1,202 adults and opposition is 72%; count only the 1,032 self-identified registered voters inside that sample and it is 73%. These are not two parallel surveys but one questionnaire split across two releases. The question numbers in the registered-voter release skip from 18 to 22 to 28, and every missing number appears in the adult release. The data center question is Q22 in both.

Reconstructed from Quinnipiac releases 3968 and 3969 and the crosstabs attached to each. Before any word is changed, changing who was asked already moves the figure by a point or two.

This distinction gets expensive because of the trendline. The trend table Quinnipiac builds itself carries two values on the adult base, 65 and 72. The 73% is not in it. So writing "opposition rose from 65% in March to 73% in September" produces a line that switches populations between its two points. Both numbers are right and the line is wrong. There is also a second 73% in Quinnipiac's output that same week. The September 30 release is headlined "73% Concerned About Potential Threat To Human Survival," which is a different question entirely and has nothing to do with data centers.

Why different surveys report different opposition rates is a question an earlier report worked through. That article looked at the scatter produced when many houses ask in many ways; this one looks at the line a single house has drawn with a single phrase. Cross-sections and time series are fragile in different ways. A cross-section wobbles when questions differ. A time series breaks when a question is revised.

2.4What the rules ask of whoever holds the question still

The house that held the wording fixed knows the word itself has shifted in public debate. Tim Malloy, a Quinnipiac polling analyst, put it this way in Bisnow on November 2, 2026.

What stayed fixed is the surface of the word, not what the word points at. That makes the rules governing question wording stricter, not looser. AAPOR's best practices for survey research put the principle in one line: do not change the ruler if you want to measure change. Knowing whether a gap between two waves reflects a real shift in attitude requires holding question wording, surrounding context, and mode of interview as nearly identical as possible. When the wording is altered, the result either looks like movement where there was none or hides movement that did occur.

Nobody usually reads the "don't know" column. In the national adult survey, though, that is the column that emptied most over six months. The three columns sum to 100, so the arithmetic closes. Support 24, oppose 65, don't know 11 became support 21, oppose 72, don't know 7, which means the 7 points opposition gained equal the 4 points "don't know" lost plus the 3 points support lost. Of the two columns that shrank, "don't know" is the larger.

Values from Quinnipiac's official trend table, redrawn as 100% stacked bars. Whatever the middle block gains, the block at the right end gives up.

Restraint is required here. These figures are marginal distributions from two separate samples, not the same people followed twice. The fact is that the "don't know" column fell from 11 to 7 while opposition rose from 65 to 72. The interpretation is that newly formed opinion went toward opposition. Writing "people who had no answer turned against it" produces a sentence the data cannot carry. Individual movement is not in this table.

Pennsylvania is composed differently. Opposition gained 6 points there, of which only 2 came from "don't know" and 4 came from support. National and state figures move in the same direction along different routes. That is precisely why one state's numbers cannot stand in for the country, or the reverse.

Reading the no-opinion column as a signal is an old habit in the survey literature. Since Philip Converse raised the idea of non-attitudes in 1964, researchers have confirmed again and again that many respondents hold no real opinion on a policy question and manufacture an answer on the spot. Work in the Jon Krosnick tradition reports that no-opinion responses rise as a question grows harder or more remote from the respondent. Turned around, a stretch in which "don't know" shrinks is a stretch in which the issue has come closer and grown easier. It is the period when opinion is being formed, not the period when it hardens.

That changing a word in a question changes the answers is among the most thoroughly verified findings in survey methodology. How much it changes them, however, varies case by case, and even the textbook examples split on whether they replicate. The table below sets out four cases this article verified, putting effect size and replication status on the same row. Reading the effect sizes alone is how this literature gets misused.

Rows two and three must not be read apart. The 6.3-point gap is widely cited, and the fact that it vanished in a preregistered replication belongs to the same body of literature.

Row four matters most here. Its claim is that the label is not all that changes: the object people picture changes with it. A soup kitchen comes to mind under "assistance to the poor" and largely fails to under "welfare." If so, the two questions are not one thing asked two ways but two different things asked. The critique meshes oddly well with OpenAI's argument. OpenAI says its facilities belong to a different class from the legacy data center, and the wording-effect literature says a different name really does put a different object in people's heads. Is the name false, then, or is it more accurate? That question is the door into the next section.

One further condition attaches. Once an issue has already grown hot, renaming works as mobilization rather than persuasion. Field experiments run by Bechtel, Hainmueller and colleagues during actual referendum campaigns report that on salient, contested issues, frames and cues pushed respondents further toward their existing partisanship regardless of which direction the frame pointed. Data centers in the United States are already such an issue. By AdImpact's count, $173 million went into advertising built around data centers in the 2026 elections, with $95 million of that concentrated in Ohio, Michigan, Wisconsin and Pennsylvania. Nearly all of it was negative.

4.1The measurements in this case point the other way

Swapping inference for measurement makes this sharper. Two time series covered the same subject over the same months, one with "AI" in the question and one without. If the word AI inflates opposition, the series that carries it should climb more steeply. The result is the reverse.

Annenberg wave one sampled 1,330 and wave two 1,320 (field June 16 to July 19). The two houses differ in sample design and mode, so the two increases cannot simply be subtracted from one another. What can be said is that the series without "AI" did not rise less.

The internal composition of the Annenberg result also cuts against conventional wisdom. In wave two, opposition peaked at 70% among respondents under 30 and bottomed out at 57% among those 65 and over. That inverts the usual pattern in which younger people are the more permissive toward new technology, and Annenberg's researchers flagged the point themselves. Frequent AI users opposed at 60%, barely different from those who hardly use it at all. The hypothesis that familiarity with AI softens data center opposition does not stand on this evidence. By party, 69% of Democrats, 54% of Republicans and 53% of independents opposed, so a majority opposes on every side.

This blog covered another instance of a split name making things hard to count only a month ago, in a piece on the National Science Foundation offering to change "AI" to "SI" in grant abstracts. That case involved government documents and research funding tallies; this one involves a company and a public opinion series. What they share is a structure in which whoever revises the name and whoever has been counting with it are not the same people.

OpenAI's argument names its comparison explicitly. The sentence Axios quoted contains the phrase "in contrast to legacy data centers." Before it is a proposal about what to call something, this is a claim about classification. What the company builds, it says, is a different kind of building from what has been called a data center until now. The claim could be true. So the right move is to split it by axis and test it.

The left column of the table below gives industry benchmark ranges for existing hyperscale cloud campuses; the right column gives published figures for three AI training campuses built between 2025 and 2026: Stargate Abilene in Texas, Meta's Hyperion in Louisiana, and xAI's Colossus in Tennessee. Cells with nothing to report say not disclosed rather than sit blank. The empty cells are results of this table too.

Verified November 4, 2026 (KST). Of the three campuses, only Abilene is an OpenAI project; Hyperion belongs to Meta and Colossus to xAI. The 40–132 kW in the density row is a general industry range, and none of the three campuses has published rack density figures of its own.

The result differs by axis. On power density and footprint, OpenAI's claim is close to true. The abatement structure reuses the existing playbook for attracting data centers outright. Permanent employment will not support a comparison at all: benchmarks span a tenfold range from 20 to 200 per 100 MW, and the campus-side numbers differ by three or four times between the contractual minimum and the company's own publicity. Water, the most contested axis of all, has no official figure from any of the three, so there is nothing to judge with. The statement "different from legacy data centers" is true on some axes, unsupported on others, and untestable on the axis that matters most.

5.1Where two readings came out opposed

The permanent jobs row is where preparing this article produced opposite conclusions twice. One calculation set 500 workers against Hyperion's 5,000 MW, read that as 0.1 per MW, and wrote that it falls an order of magnitude below the usual 1 to 2 per MW. The other converted the same figures to a per-100-MW basis and wrote that they land inside the benchmark range. Each is correct on its own arithmetic. They parted over whether the benchmark is 20–50 per 100 MW or 100–200. When the ruler differs, the same fact yields opposite conclusions. The problem this article is about turned up once more inside its own source material.

5.2One number moves no matter what the buildings are called

Section 1 noted that electricity costs were the reason opponents cited most in Quinnipiac's March survey. That number is already moving, and the name has nothing to do with it. Two figures are the firmest. First, PJM capacity market auction prices rose 833% year over year, and Monitoring Analytics, the independent market monitor, attributed $9.3 billion of the $14.7 billion total to data center demand, or 63%. That is an accounting allocation of market costs rather than an estimate. Second, residential electricity rates in the six states where data centers cluster most have risen 25.6% since 2022, against 23.7% elsewhere. The gap is about 2 points.

Nothing like "sharply wider" should be attached to that second figure. Two points is not a large gap. What the number says is that dense areas rose somewhat more, not that data centers drove rates up. Natural gas prices and the cost of replacing aging transmission were at work over the same years. The direction is clear regardless. Renaming the buildings AI compute campuses changes neither the PJM auction price nor anyone's utility bill. The power-side background is laid out separately in a report on the power wall.

5.3The example OpenAI chose is somewhere else

One addition is needed. The project OpenAI offered in the Axios report as its instance of "different from legacy data centers" is not among the three campuses above. It is a site in Pike County, Ohio: the PORTS-Pike Technology Campus, announced August 17, 2026, which SB Energy builds, owns and operates with OpenAI as a 20-year tenant. The first 800 MW arrives in 2028 and completion is set for 2032, and OpenAI has committed $164 million to community and education programs in Ohio. Abilene's figures must not be attached to the Pike County story. On this subject, writing down which site a number belongs to every single time is a matter of accuracy rather than style.

Data center measures on U.S. ballots for November 3, 2026 number 41 across 10 states. That is Ballotpedia's count, which Bloomberg recounted and reported as "dozens." If passed, 28 would restrict development and 13 would make it easier. Across the full year the figure becomes 50 measures in 13 states. Six more were already decided before November 3 and three are scheduled after it. Mixing these counts is best avoided. Ballotpedia records that all six of the earlier ones came out on the side of tighter regulation. Forecasting forty-one from six, though, is another matter.

How fast the count grew shows when it is set beside this blog's own record from four months earlier. The Monterey Park piece from July 2026 recorded the first U.S. city to ban data centers by ballot. There was one of them then, carried with 86% in favor. Four months on there are 41 measures of the same kind.

Where the 41 sit reveals the geography of this issue. More than half are in a single state.

Ballotpedia's count for the November 3, 2026 ballot. Sorted by character, 28 would restrict development and 13 would ease it. Of the 28 restrictive ones, 17 are bans, and of those 17, twelve block only projects above a stated size while five ban outright. The remainder are three six-month moratoria and eight requiring resident approval, expert review or similar.

Here is where this article turns. What these ordinances count is not words. The Kansas measure bans facilities above 25 MW, and Urbana, Ohio bans those above 7.5 MW, the lowest published threshold of the set. The 25 MW threshold reached nine separate ballots within Ohio alone in identical language. Granville, Sunbury, Piqua and six other municipalities copied the phrasing a statewide campaign had been using, word for word. Language travels that easily when its core is a number. One Wisconsin measure requires resident approval for projects above $450 million, and the North Dakota measure requires independent expert review and a public hearing for projects over its threshold. Two of the Illinois measures are non-binding expressions of opposition. Every limit in this list is written in megawatts or dollars.

Two notes keep the scope honest. First, all 21 Ohio measures are local. A statewide constitutional amendment banning anything above 25 MW was confirmed on June 19, 2026 as failing to qualify for this year's ballot. Second, the actual text of these ordinances was not obtained in the original. What appears above reproduces the summaries given by the tally sources and the coverage, which is why no clause here is placed in quotation marks.

This is an article about measurement rather than about public opinion. There is a question Pebblous asks first whenever we open a customer's data: when did the definition of this column change? Without knowing the date of a definition change, a trend is not a trend but two fragments, and a chart that welds the fragments together either shows a change that never happened or erases one that did. This article put that question to a single item on a national survey.

7.1Three quality lessons this story hands over

First, one word in the name of a population changes the number. A single set of interviews produces 72 on the adult base and 73 on the registered-voter base, and the same Q28 splits into 81 and 79. In data work the equivalent is one filter line defining what gets counted. Without knowing where that filter is written, nobody can explain why two reports disagree.

Second, "don't know" is a signal rather than a residual. While opposition gained 7 points, that column lost 4, more than the 3 points support gave up. It plays the same role as the share of deferred judgments in a labeling job. A stretch where deferrals fall is usually a stretch where annotators are settling on a criterion, and where the decision boundary moved in that stretch stays invisible unless the deferral rate is watched on its own.

Third, a new name can be the more correct one and still not join up with the values recorded before it. The same thing happens on the day a schema column gets a name that fits it better, and the dashboard raises no warning that day. Two numbers becoming incomparable because one company counted "revenue" one way and another counted it differently is something we watched play out at these very companies. There the definitions differed between firms; here a definition is about to differ before and after a date.

7.2Four questions for renaming a metric

Four questions fall straight out of this story. They are not specific to polling: they apply just as directly to KPIs, label definitions and schema column names.

• What changed that prompts the rename? Has the object itself changed, or is only its appearance being adjusted? For OpenAI, the object really did change on power density and footprint, and did not on tax structure.
• What breaks if it changes? Measure the length of the line that breaks, and write down who had been making decisions from it.
• Can there be a period of overlap? Survey research already has a name for this: the split-ballot. One sample is divided so the old wording and the new are each put to a half, and the gap gets measured first. And once a change is made, there are only three ways to report it. The Pew Research Center published the framework when it moved from telephone to online, and it sets out all three: break the trend and show only the new values, place the new values inside the old trend with the change point marked, or wait until several years of new values accumulate and then join them gently.
• Does whoever proposes the change have a stake in the result? If so, record that fact alongside the change. Recording it is enough, and the rest belongs to whoever is reading.

7.3A record made in advance, not an audit after the fact

Articles of this kind on this blog have mostly been audits after the fact. A bias test that was actually measuring reasoning ability, revenue defined differently from company to company, "AI" offered up for replacement by "SI" in grant abstracts: each of those is a record of opening up a metric once the thing had already happened. This article stands somewhere slightly different. It writes down the moment just before the measuring word changes. Written down afterward, the account has to begin by reconstructing what the values were before.

Quinnipiac figures were taken directly from the release PDFs and from the trend tables and sample documents the house publishes alongside them. Originals behind an access block were verified through syndicated copies, and each such entry says so.

Primary — polling releases

• 1.Quinnipiac University Poll, release 3955 (2026-03-30). National adult AI survey; the data center item is Q38. poll.qu.edu
• 2.Quinnipiac University Poll, release 3969 (2026-09-30). National adult AI survey with the official trend table and sample documents; the data center item is Q22. poll.qu.edu
• 3.Quinnipiac University Poll, release 3968 (2026-09-29). National registered-voter political survey, a subset of the same interviews, with the data center item also numbered Q22. poll.qu.edu
• 4.Quinnipiac University Poll, release 3962 (2026-07-15). Pennsylvania registered-voter survey with trend table. poll.qu.edu
• 5.Quinnipiac University Poll, release 3948 (2026-02-25). Pennsylvania registered-voter survey. pa02252026_pdkl34.pdf
• 6.Annenberg Public Policy Center. AI survey topline (W38). annenbergpublicpolicycenter.org

Primary — reporting and company releases

• 7.Jim VandeHei. "Behind the Curtain: OpenAI's post-election power play." Axios (2026-11-01). Direct access to axios.com was blocked, so this was verified through a syndicated copy. syndicated copy
• 8.Sasha Jones. "OpenAI Plans Data Center Rebrand To Make Projects Palatable To Voters." Bisnow (2026-11-02). Source for the OpenAI spokesperson statement (provided to Business Insider) and for the Tim Malloy remarks. bisnow.com
• 9.OpenAI. "Introducing the Grid Protectors Program" (2026-11-02). Direct access to openai.com was blocked, so this was verified through syndication and quotation.
• 10.Ballotpedia. "Data center-related ballot measures, 2026." ballotpedia.org
• 11.Miquéla V Thornton. Bloomberg (2026-11-03), BusinessMirror syndication. The article itself names Ballotpedia as the source of the count. businessmirror.com.ph
• 12.AdImpact tally, via Axios (2026-10-31). $173 million in advertising built around data centers. axios.com
• 13.Sachin Katti (head of compute infrastructure, OpenAI). Post calling the Abilene Stargate site an "AI data center campus" (2026-03-06). x.com

Academic — wording effects and non-response

• 14.Smith, Tom W. (1987). "That Which We Call Welfare by Any Other Name Would Smell Sweeter." Public Opinion Quarterly 51:75–83.
• 15.Schuldt, Jonathon P., Sara H. Konrath, & Norbert Schwarz (2011). "'Global Warming' or 'Climate Change'? Whether the Planet Is Warming Depends on Question Wording." Public Opinion Quarterly 75(1):115–124.
• 16.Soutter, Alistair R. B., & René Mõttus (2020). "'Global warming' versus 'climate change': A replication." Journal of Environmental Psychology.
• 17.Schuldt, Jonathon P., Peter K. Enns, Sara Konrath, & Norbert Schwarz (2020). The original authors' reply to that replication. Journal of Environmental Psychology.
• 18.Huber, Gregory A., & Celia Paris (2013). "Assessing the Programmatic Equivalence Assumption in Question Wording Experiments." Public Opinion Quarterly 77(1):385–397.
• 19.Bechtel, Michael M., Jens Hainmueller, Dominik Hangartner, & Marc Helbling (2015). "Reality Bites: The Limits of Framing Effects for Salient and Contested Policy Issues." Political Science Research and Methods 3(S3).
• 20.Krosnick, Jon A. et al. (2002). "The Causes of No-Opinion Responses to Attitude Measures in Surveys."
• 21.Converse, Philip E. (1964). Origin of the non-attitude concept. Cited indirectly.

Standards and methodology

• 22.AAPOR. Best Practices for Survey Research. Source for the principle of holding trend questions fixed and for the split-ballot prescription. aapor.org
• 23.Pew Research Center (2019-02-27). "What our transition to online polling means for decades of phone survey trends." The three ways of reporting a broken trendline.
• 24.U.S. Energy Information Administration. Today in Energy — PJM capacity market. eia.gov

Related Pebblous articles

• 25.Pebblous, "Did Data Center Opposition Grow, or Did the Question Change?" (2026-09-23). report/datacenter-opposition-reasons-pa-study
• 26.Pebblous, "The National Science Foundation Offers to Change 'AI' to 'SI' in Grant Abstracts" (2026-10-13). blog/nsf-grant-ai-to-si-renaming
• 27.Pebblous, "Monterey Park Bans Data Centers by Ballot" (2026-07-16). blog/monterey-park-data-center-ban
• 28.Pebblous, "AI Hit a Wall This Summer, and It Wasn't Chips or Money" (2026-08-01). report/ai-datacenter-power-wall-2026
• 29.Pebblous, "OpenAI and Anthropic Count Revenue Differently" (2026-10-10). blog/openai-anthropic-annualized-revenue-definition
• 30.Pebblous, "The bias check AI labs report ranks models like a reasoning test" (2026-09-16). report/benchmark-concept-validity-2026-09