Summary
- Freeze-candidate wins in Texas, Georgia, Ohio and Michigan did not change what OpenAI, Meta and Anthropic want to build — the wins restructured how data centers get authorized, promoting water, noise and community consent from administrative periphery to veto-grade criteria and widening the documented option set beyond greenfield U.S. construction.
- The election produced freeze-promising winners rather than documented freezes, but the recorded corporate response — Anthropic’s S-1 warning, Meta’s assumption that U.S. building is “going to be harder,” OpenAI’s transparency package — treats local consent as a veto-grade criterion whose prohibition no cheap land, fast interconnection or favorable tax treatment offsets.
- The November results delivered the delayed signal of a balancing feedback loop: construction’s local water, energy and noise impacts eroded community acceptability, the erosion ran through election cycles, and statehouse constraint arrived against a commitments pipeline documented at approximately $895.4 billion at Anthropic alone through 2045.
- Inside the non-moratorium tier, the ranking of alternatives is not robust at the top: domestic scale-up of existing sites leads when speed-to-capacity outweighs prohibition-avoidance, overseas siting leads when the weights reverse, and greenfield construction in the four freeze-won states sits dominated under any plausible weighting.
- Whether freeze platforms harden into moratoriums or soften into negotiated conditions is the open variable both sides are documented trying to learn: Powell McCormick’s account of a winning governor-elect requesting a meeting rather than a moratorium cuts in both directions at once.
Within a day of midterm wins for candidates promising to freeze data-center development in Texas, Georgia, Ohio and Michigan — “all major states for data-center sites,” as The Wall Street Journal reported — the AI industry’s largest builders were re-running their siting decisions. Sam Altman acknowledged public concern about water use during an employee question-and-answer session and said he understood why people might not want to live near data centers, according to people familiar with the matter; Dina Powell McCormick said Meta assumes building in the United States is “going to be harder” and is weighing sites abroad; investor Brad Gerstner sized the exposure at as much as 35% of the U.S. data-center construction planned for next year; and OpenAI said Thursday it would drop nondisclosure agreements from site negotiations and publish information on the “energy, water and noise impacts” of its projects. The midterms did not change what these companies want to build; they changed the criteria and the channel through which data centers are authorized — turning local consent into a veto-grade criterion, loading an approval interface past its tested capacity, and delivering the delayed signal of a feedback loop whose commitments stock, documented at approximately $895.4 billion at Anthropic alone through 2045, was accumulated without a matching political sensor.
What changed in a day
A “wave of data-center opposition” helped candidates promising to freeze development win statehouses in Texas, Georgia, Ohio and Michigan, The Wall Street Journal reported. Companies that have committed “hundreds of billions of dollars” to the artificial-intelligence race, and that had “cultivated relationships with AI-friendly officials at all levels of government,” found themselves facing the prospect of blanket moratoriums, tighter environmental regulations and longer review timelines, along with potential new AI-safety regulations that “could further bog them down” — and “rushed to plan for a changed reality.”
The reckoning extended inside OpenAI. During a question-and-answer session with employees, Chief Executive Sam Altman “acknowledged that the public is concerned about water use” and said “that whatever the reason, he understood why people might not want to live near data centers,” according to people familiar with the matter. At one point, he reportedly mused that the company should perhaps look to build data centers in the desert. Senior OpenAI leaders were discussing whether the company would need to reset expectations for its data-center capacity next year, the Journal previously reported. Employees involved in OpenAI’s data-center development were “exchanging intelligence on newly elected Democratic politicians” in key states and discussing how best to engage with them, according to messages viewed by the Journal.
Across the industry, executives in charge of government relations for the hyperscalers “moved quickly to open lines of communication with incoming governors and their transition teams, seeking an understanding of how their campaign-trail rhetoric would translate into policy actions.” The tech investor Brad Gerstner, speaking on the “All-In” podcast, put a number on the exposure: as much as 35% of the U.S. data-center construction planned for next year is now at risk. The figure is an investor-side estimate, and the reporting does not specify its denominator.
Dina Powell McCormick, who as president of Meta Platforms oversees its AI-infrastructure operations, said at the WSJ Tech conference Wednesday night that Meta assumes building data centers in the U.S. is “going to be harder” and that getting new ones done from scratch will be “very challenging.” She said Meta and other tech companies are increasingly considering building new data centers outside the U.S. as “one way around the new hurdles,” and that another avenue is scaling up existing data centers, where surrounding communities that have seen “benefits in job creation and tax revenue” “don’t require as much persuasion.” She also relayed one early opening: “We did have a governor that won last night who did make data centers a big part of their platform say, ‘As soon as we get past the inauguration, let’s meet, because we don’t want to lose this investment in our town. We don’t want these jobs to go elsewhere,’ ” she said. On the industry’s public posture, Powell McCormick said, “I think we have to double and triple down on telling the story.” OpenAI President Greg Brockman, also speaking at the conference, echoed the point: “We need to do a better job of really communicating to local communities about why data centers are something that can be beneficial to them.”
OpenAI became the latest AI giant to unveil a broad commitment on Thursday: the company said it would not require public officials to sign nondisclosure agreements to engage in data-center negotiations, would solicit community input, and would provide clear public information about the “energy, water and noise impacts” of its infrastructure investments. Anthropic had flagged a similar risk to investors a week earlier. In an S-1 filing for a planned initial public offering, the company previewed the risk that local and state governments might adopt data-center moratoriums: “Such moratoriums, or similar actions, could delay or prevent the construction or expansion of facilities on which we depend for compute.” The filing detailed approximately $895.4 billion of AI computing and infrastructure commitments through 2045. Anthropic said it is pursuing projects in several “non-U.S. jurisdictions” as part of a plan to “reduce concentration risks associated with operating a significant portion of our compute infrastructure in the United States.” The company has separately announced plans to build data centers in Texas and New York, with sites “coming online this year.”
A siting decision rebuilt around a veto
The documented option set now has four entries: greenfield construction on new U.S. sites; scale-up of existing facilities — Powell McCormick’s account, quoted above, holds that host communities with accumulated job and tax-revenue benefits require less persuasion; construction in “non-U.S. jurisdictions,” per Anthropic’s S-1; and desert siting, which exists in the record only as Altman’s reported musing. Grouped by response type, the record shows siting options (new U.S. builds, overseas construction, the desert variant), scale options (expansion of existing sites versus new builds from scratch, which Powell McCormick called “very challenging”), and communications (Powell McCormick’s “double and triple down” prescription and Brockman’s local-community appeal). One gap is visible in the option set as described: the reporting records no design-side response to the water and noise concerns Altman acknowledged — no account of facilities redesigned to impose less of the burden enters the record.
Mapped against the criteria a siting decision weighs, the record supplies:
- Outcome quality: compute capacity delivered, and how soon — OpenAI’s leadership reckoned with whether the outcome would make it harder to secure the computing power the company needs to keep growing.
- Cost and commitment: “hundreds of billions of dollars” committed industry-wide; approximately $895.4 billion through 2045 documented at Anthropic.
- Risk: blanket moratoriums, tighter environmental regulations, longer review timelines, potential new AI-safety regulations.
- Fit: water use and neighborhood impacts.
- Reversibility: overseas construction, expansion of existing facilities, desert siting.
Some criteria are under-named on both sides of the table. On the companies’ side, water and noise were peripheral enough that OpenAI’s pledge to publish impact information is itself news. On the states’ side, the revealed criterion set is broader than the campaign platforms: Powell McCormick described a governor-elect whose winning platform made data centers “a big part of their platform” — on which side of the question the account does not say — and who asked to meet to keep the investment and jobs in town (quoted above). Retention, not prohibition, is the criterion her account records.
The criteria also overlap. “Approvals” and “persuasion” — Powell McCormick’s framing — appear to measure the same underlying quantity, local political consent, with formal review timelines and the persuasion a site requires operating as separate proxies; the proxies diverge where goodwill has accumulated.
Because one criterion behaves as a veto, the decision fits an outranking structure of the ELECTRE family rather than an additive weighted score: an additive ranking would let strong scores elsewhere mask a site that cannot be approved at all. The veto’s scope needs precision. The election produced winners who promised freezes, not documented freezes. But a moratorium, once enacted, is non-compensatory — no amount of cheap land, fast interconnection or favorable tax treatment offsets a construction prohibition — and Anthropic’s hedged filing language reads as a prohibition threshold, not a cost line. That yields two tiers. In the first tier — any jurisdiction with an enacted moratorium — the veto binds absolutely and no weights compensate it; it is never converted into a weight. In the second tier — jurisdictions where no enacted moratorium exists, which covers every U.S. case documented in the article — prohibition-avoidance operates as the dominant weighted risk criterion, and greenfield projects in the four freeze-won states are outranked for near-term delivery under any plausible weighting. The scale-up-versus-overseas contest runs entirely inside the second tier.
The weights are elicited from documented conduct rather than stated preferences. Continuity and speed rank heaviest: “hundreds of billions of dollars” are committed, and Anthropic’s filing documents approximately $895.4 billion through 2045. Community acceptance carried enough weight by Thursday for OpenAI’s public commitments. Concentration risk moved from implicit to explicit when Powell McCormick called overseas sites “one way around the new hurdles.”
The desert variant is pruned from the ranking as unprogrammed: a reported musing, according to people familiar with the matter, carrying no announced site, timeline or commitment, it cannot be ranked against documented alternatives — and it is a different object from Anthropic’s documented pursuit of “non-U.S. jurisdictions.” As a substitution, desert siting would act on proximity and political tolerance but not on the water criterion, and it would move the water burden into regions where supply is characteristically scarcer — an apparent tension rather than a resolution.
One caveat guards against hardening that prune: Anthropic’s announced Texas and New York plans, with sites “coming online this year,” stand alongside the freeze-won states in the same record — and the filing’s “could delay or prevent” hedge is doing exactly the work its language intends.
The ranking’s sensitivity follows. Within the second tier, raise speed-to-capacity above prohibition-avoidance and domestic scale-up leads — Powell McCormick’s account of one winning governor opening with a meeting request indicates receptive hosts exist. Raise prohibition-avoidance above speed and overseas siting leads, though the source supplies no foreign timelines, so that branch rests on the least-documented criterion. The bottom of the ranking is robust: the unprogrammed desert musing and the greenfield entries in the freeze-won states, dominated by prohibition risk under any plausible weighting, do not move. The top is not robust, and the live contest turns on a single weight.
Where the load concentrates: the approval interface
If a freeze statute lands in one of the four states, the component that yields is legible in advance. It is not steel, chips or grid interconnection. It is the approval interface: the sequence of site control, permitting and authorization between a signed commitment and a shovel, together with the company–community negotiation channel. OpenAI’s Thursday package — no nondisclosure agreements for officials, solicited community input, published impact information — reads as a transparency redesign of that interface; the reporting does not describe what practice preceded it.
The load on the interface is documented from three directions. Gerstner’s investor-side estimate puts as much as 35% of next year’s planned U.S. construction at risk, denominator unspecified. Anthropic’s S-1 documents approximately $895.4 billion of commitments through 2045. And companies that “rushed to plan for a changed reality” faced the prospect as a multi-state moratorium wave became specifiable. Read structurally, roughly the next year’s construction flow ran through a channel controlled by state and local politics, and a single election reset that channel’s capacity — that channel reading is an analytical extension, not part of Gerstner’s statement.
One second failure channel named in the article is deliberately deprioritized: “potential new AI-safety regulations that could further bog them down.” No instance of that channel appears in the article. The moratorium channel, by contrast, carries four documented statehouse outcomes, an investor-facing S-1 disclosure, and executives’ on-record responses. The single-yielder diagnosis stands on that asymmetry of documentation, not on the second channel’s absence.
Two design dependencies explain why the load lands where it does. First, officials over communities: the companies had “cultivated relationships with AI-friendly officials at all levels of government,” and OpenAI’s NDA announcement confirms that closed-door channels were part of the prior operating model. Second, co-located benefits and impacts: jobs and tax revenue accrue to host towns, and so do the water, energy and noise burdens — the three impact categories OpenAI now commits to publishing. The water concern Altman reportedly acknowledged, according to people familiar with the matter, runs through residents, not transition teams.
None of this arrived unannounced. The leading indicators were visible before the vote: an opposition wave strong enough to decide statehouses; Altman’s reported acknowledgment of water concerns and understanding of why “people might not want to live near data centers”; Anthropic’s S-1, filed the prior week, warning of moratorium risk; and internal OpenAI traffic — according to messages viewed by the Journal — in which employees were exchanging intelligence on newly elected politicians in key states (each documented above).
Three breakage narratives describe how the interface fails. The first is an obligation–approval mismatch: the transmission mechanism by which a statehouse win becomes a binding approval constraint — legislation, permitting, or utility intervention — is undetermined in the reporting, but Anthropic’s S-1 language describes capacity arriving later than obligations require, and the reset-expectations discussions reported inside OpenAI are a documented symptom of the mismatch. The second is fallback closure: Meta’s scale-up alternative depends on incumbency insulating expansions from the same constraints, and whether water and energy limits bind at existing sites as they do at new ones is not established by the reporting — the fallback narrows if they do. The third is substitution risk: each alternative leaves part of the constraint set untouched. Overseas siting substitutes regulatory counterparts whose approval processes the reporting does not examine for the U.S. approval dependency it does document — feasibility asserted by executives, not established in reporting. The desert variant does not act on the water criterion at all.
What to watch over the next twelve months follows from the structure:
- Permitting and review timelines lengthening.
- Quiet withdrawals or re-sitings.
- The growth rate of overseas announcements, as a proxy for domestic constraint.
- The spread of OpenAI-style transparency commitments.
- Further investor-facing risk disclosures of the S-1 kind.
- And whatever community acceptability does in response to being, for the first time, addressed directly.
The loop structure behind the freeze wins
Four stocks anchor the picture: operational U.S. compute capacity; the committed-but-unbuilt pipeline, of which Anthropic’s approximately $895.4 billion through 2045 is the documented instance; community acceptability of data-center siting — not directly measured in the source, but inferable from documented behavior, including Powell McCormick’s observation that existing host communities “don’t require as much persuasion”; and an emerging overseas pipeline. The flows run through new-site approvals into the pipeline, completions out of it, and diversion of planned capacity to foreign jurisdictions. Two auxiliaries matter: statehouse composition, itself a delayed function of acceptability, and the water, energy and noise impacts the system imposes. Those impacts are exogenous only conditionally, for a fixed geography — geography is itself a decision variable in the siting analysis, so impact levels are partially endogenous to the siting choice. The reported desert musing is precisely a proposal to change the water burden by changing the map.
One reinforcing loop (R1) drove the commitment boom: commitments drive construction, construction builds operational capacity, capacity funds products and revenue, and revenue drives further commitments — the loop behind the “hundreds of billions” now committed, still running on inertia as Anthropic’s sites come online.
One balancing loop (B1) has a long delay: construction volume produces local water, energy and noise impacts; acceptability falls; the delay runs through election cycles and multi-year review timelines; then statehouse constraint rises and the approval flow narrows. The November results are that delayed signal arriving as a discrete change — and the S-1 disclosure is the first documented wiring of the political variable into investor-facing accounting.
Two more loops are newly actuated but unproven. A reinforcing loop (R2): as domestic approvals narrow, companies divert planned capacity overseas; the job and tax-revenue benefits stop accruing to new host communities; acceptability erodes further. A balancing loop acting against B1 (R3): OpenAI’s transparency commitments and the communication push described by Powell McCormick and Brockman act on the same acceptability stock from the other direction, with unknown magnitude and the same long response time. Alongside R3 sits a second documented force on acceptability: governor receptivity. Powell McCormick’s account of a winning governor opening with a request to meet is a pro-siting signal carried by incoming executives — an auxiliary input on the acceptability stock whose magnitude is unknown, and the documented early counter-weight to B1 of unknown strength.
The dominance has shifted. R1 dominated the commitment boom; B1 now dominates the approval flow — that is the operational meaning of Gerstner’s 35% figure. R2 and R3 are actuated but unproven, and both transmit through the slowest stock.
The timescales are separated. Balancing loops dominate the short term, which is why the documented adjustments concern next year’s pipeline and incoming governors’ teams. The electoral loop runs on two- and four-year cycles while the obligation stock runs to 2045; the political loop can turn several times inside a single obligation cycle. That separation is the structural logic of Anthropic’s concentration-risk reduction: decoupling long-dated commitments from any single political boundary.
The wiring is asymmetric, and the delays are shared. The balancing loop acts on the flow, while the commitments stock was accumulated without a matching sensor — a structural reading of the design, not a prescription. Both sides face the same delays: new capacity cannot be completed inside a year, and at least one incoming governor, per Powell McCormick’s account, is opening with a meeting rather than a moratorium.
The structure carries the signatures of several patterns the systems-dynamics literature has named. It has the shape of limits-to-growth: compute demand drives buildout while local absorption capacity — water, grid, political tolerance — sets the limiting factor, and the storytelling and transparency responses are attempts to raise that limit. It resembles fixes-that-fail: rapid large-scale commitments produced impacts that mobilized the opposition now constraining the fix, the arc running from cultivated relationships with AI-friendly officials to the overnight reversal. It shows shifting-the-burden: overseas relocation and desert siting shift load away from the fundamental constraint, while disclosure commitments and communications prescriptions address the constraint directly. It carries tragedy-of-the-commons in Garrett Hardin’s formulation: individually rational siting races drew down political tolerance and water in the same handful of states in unison — the structure underlying Gerstner’s aggregate exposure figure. And it exhibits success-to-the-successful: incumbency in a locality confers an advantage entrants lack, per Powell McCormick’s account.
What the frame leaves out — and the open variable
The edges of the picture are definable. Water authorities, grid operators and county-level zoning bodies sit at the analysis’s boundary — named implicitly through the impact categories OpenAI pledged to disclose — and are candidates for inclusion as policy responses form. Two domains are deliberately excluded: semiconductor supply and model-competition dynamics, because the constraint set the reporting identifies operates entirely at the siting-approval-impact layer, and nothing in the article indicates the election changed hardware supply or model-race conditions.
One variable remains open, and it is the one both sides are documented as trying to learn. On the companies’ side, government-relations executives sought “an understanding of how their campaign-trail rhetoric would translate into policy actions” (quoted above); on the states’ side, the governor anecdote records the ambiguity in both directions at once — a governor-elect whose winning platform made data centers “a big part of their platform,” on which side of the question the account does not say, who wants the investment kept. Whether freeze platforms harden into moratoriums or soften into negotiated conditions determines which loop dominates next. The reporting leaves it undetermined.
Verification status and sourcing notes
Supplied-source priority governs throughout: every fact supplied by The Wall Street Journal article is retained unchanged as the source of truth. External checks did not resolve the following claims, each of which is reported as the article states it and marked unresolved: the four-state midterm outcomes; Gerstner’s 35% figure; the approximately $895.4 billion S-1 figure; Altman’s anonymous-source acknowledgment; OpenAI’s Thursday commitments; and the S-1’s existence, timing and risk language.
Dina Powell McCormick’s role carries one unresolved tension. The article’s framing — “president of Meta Platforms” overseeing its AI-infrastructure operations — is retained here. A same-outlet (WSJ) excerpt reports Meta naming her president and describes her helping guide overall strategy amid the AI-infrastructure buildout — phrasing consistent with, though softer than, the article’s “oversees.” One reading treats that excerpt as content-level corroboration; another treats it as topic-level only, leaving the article’s framing unverifiable with partial corroboration. No resolution is available from the material supplied, and the analysis reads identically under either: the article’s framing stands, and the excerpt’s additional “Vice Chair” title detail, which exceeds the supplied material, is not adopted.
The governor quotation reaches the record only through Powell McCormick’s account of what a governor-elect told her; it is attributed as her account throughout.
The evidence base is executive-side: company statements, an executive podcast appearance, and anonymous employee accounts. No candidate-side or community-side statement beyond the election outcomes themselves appears in the supplied record, so even-handed sourcing across the two coalitions cannot be fully confirmed.
Two methodological qualifications attach to the analysis above. The veto-criterion and ranking findings rest on qualitative judgment; no enumerated criterion identifiers or numeric thresholds back them. And every evidential hedge in the source material — “according to,” “reportedly,” “per people familiar with the matter,” and the like — is preserved verbatim above; the matching was performed by hand rather than by automated alignment.
Analytical techniques used in this piece
This analysis applies the methods below. Each links to a short, plain-English explainer you can read and reuse.
- Multi-Criteria Decision Analysis
- Scores competing options against several weighted criteria at once.
- Pre-Mortem (Fragility)
- Imagines a system has already broken and traces the structural fragilities that let it.
- Systems Dynamics (Structural)
- Maps a system’s structure — stocks, flows, and the architecture that shapes its behavior.