Cancelled and Delayed Data Center Projects 2025–2026

The 2025–2026 data-center backlash is no longer a zoning nuisance. It is now a capacity, power, and political constraint that has already killed multi-gigawatt campuses and, after the November 3, 2026 midterms, put a large share of the still-unbuilt 2028 pipeline inside states whose new governors ran on pauses. The market’s November 4 reaction split the AI trade in two: hyperscalers rallied on a slower-capex story, while power producers and data-center equipment names sold off hard—especially anything tied to Texas.

Named projects delayed, cancelled, or withdrawn

Trackers disagree on totals because they use different size cutoffs and mix full cancellations with developer exits. What is consistent is the mechanism: residents and local governments are blocking annexations, zoning, and on-site gas, and developers are walking away when permitting and community support look worse than at an alternate site.

Large, sourced cases:

• Amazon / AWS, near Trenton, Ohio — about 1,000 MW. On November 3, 2026, Trenton voters banned data centers above 25 MW with roughly 80% of the vote. That closes the city’s annexation path for an 18-building campus on about 600 acres of township land; the city had already withdrawn the annexation petition in September. St. Clair Township has also moved toward a zoning ban. The city-limits ban does not automatically kill a project that stays outside the city, but the political route the developer was using is closed. A separate Prologis facility already under construction in Trenton’s industrial park (about 250 MW in some local reporting) faces an unclear limit under the new charter rule. [1] [2] [3]

• Green Data (Green Data Centers), Socorro, New Mexico — 2 GW. On June 2, 2026, New Mexico Tech President Michael Jackson said the university would not proceed or extend its agreement, citing insufficient contiguous acreage, renewable-integration challenges, and community concerns. Coverage also described widespread local backlash over water and land. Socorro County then adopted a one-year moratorium. [4] [5] [6]

• A1 Data Center, Millville, New Jersey — 1.4 GW. On May 19, 2026, the Board of Commissioners banned data centers citywide, blocking what local coverage called the largest data-center proposal in New Jersey. Organized opposition (Climate Revolution Action Network and residents) cited grid draw, water, and incompatibility with local infrastructure. [4] [7] [8]

• Prometheus Hyperscale, east of Casper, Natrona/Converse counties, Wyoming — 1.25 GW (local figures also cite 1.2–1.5 GW). On October 28, 2026, President Trevor Neilson said the company would no longer pursue a data center in either county and pointed to a “clearer path to permitting and community support” at its Uinta County (Evanston) site and elsewhere. Neighbors had packed meetings over noise, emissions, and traffic. The landowner (Falls Ranch) said development would continue under a new company, so this is a developer withdrawal, not a confirmed death of the site. [9] [10] [11] [12]

• Skylar Capital Opportunities (Skylar Opportunities), Henderson, Nevada, BLM land near Paradise Hills — 1,000 MW plus a natural-gas plant, estimated build cost $1.5–2 billion. Withdrawn September 24, 2026. The company’s stated reason was the City of Henderson’s planned urban development on the same area. Environmental groups had also opposed data centers on public land; the same sponsor’s Boulder City project remained contested. [13] [14] [15]

• Stillwater Development, Lovejoy, Georgia — about 1.25–1.3 GW (Lovejoy / Stillwater Technology Park). Formally withdrawn August 6, 2026, after opposition; the mayor indicated the developers intended to pursue residential use instead. [13] [4]

• EdgeConneX, Ashville, Ohio — on-site gas plant of up to about 800 MW tied to two data centers. Voters repealed the term summary that set the village’s deal, about 82–18, on November 3, 2026, after the Ohio Supreme Court forced the question onto the ballot. Behind-the-meter generation did not neutralize the political objection. [2]

• Meta, Piqua, Ohio — 1.4 GW campus. About 67% of voters backed a charter amendment aimed at the campus and recalled the mayor and two commissioners (Cleanview tally of unofficial returns). [16]

• Geronimo Power, Nobles County, Minnesota — 400 MW “data park.” County officials rejected the proposal in spring 2026 despite outreach. The company’s CEO described opposition as more intense than anything in two decades of development. [17]

• Project Jupiter (STACK Infrastructure / BorderPlex; Oracle as tenant for OpenAI / Stargate), Doña Ana County, New Mexico — up to about 2.45–2.5 GW. Not cancelled. Buildings are under construction (first structure topped out in late October 2026). Power is the constraint: the New Mexico State Land Office rejected a gas-pipeline route across trust land; the Green Chile lateral’s in-service target slipped from August 2026 to February 2027, and an air permit for the Bloom fuel-cell microgrid remained contested into November. In September 2026 Oracle sent a force-majeure notice to the Blue Owl / STACK side to preserve the right to defer rent if 2028 energization slips. Oracle has said the schedule is unchanged. This is a live power-and-politics delay risk on a flagship campus, not a withdrawal. [17] [18] [19] [20]

A Programs.com tracker of large U.S. projects (at least 100 MW or $1 billion) that were withdrawn, denied, or lost their land deal lists additional gigawatt-scale cancellations with less complete public reasons in the summary table, including Blakely, Pennsylvania (1,500 MW), Project Washington in New Castle County, Delaware (1,200 MW), Meridian Technology Park in Christian County, Illinois (1,000 MW), and LaOsa in Pinal County, Arizona (1,000 MW). [13]

How large the killed and stalled pile already is

Different research shops are counting different things, and the dollar figures should not be added together.

• Programs.com, updated through November 4, 2026: 69 large projects cancelled since January 2024 across 26 states; 15,910 MW of planned capacity cancelled on the 26 projects that reported a size; $165 billion of announced investment on the 32 that reported a dollar amount. Of that, 2025 was 23 projects, 4,700 MW, and $72.55 billion; 2026 through November 4 was 43 projects, 11,210 MW, and $76 billion. Local opposition was the cited reason in 33 of 69 cases (48%); zoning or permitting in 15; power or grid in only 1 under that coding. Power often shows up inside “opposition” rather than as its own label. [13]

• Sue Data Centers registry, as of November 4, 2026: 164 cancelled or withdrawn U.S. sites, 14.1 GW, $69.9 billion, 102 after documented opposition. The lower dollar total versus other trackers implies a tighter investment screen and a wider site count. [4]

• Heatmap Pro: at least $260 billion of data-center investment cancelled in 2026 after sustained local opposition, about half of that ($130 billion) in the third quarter alone, and roughly $1 trillion now facing some form of sustained local opposition. Heatmap also counted at least 20 projects killed by local opposition in Q1 2026, the most of any quarter on its record. [21] [17]

• Data Center Watch (10a Labs), as reported by Capacity and Bloomberg: at least 75 projects worth about $130 billion blocked or delayed by local opposition in Q1 2026, and a further 45 worth $68 billion in Q2. [1]

• S&P Global: 146 delayed or cancelled projects, with roughly three in four linked to local pushback; a separate summary said more than 145 were delayed or cancelled through early August, mostly from public opposition. S&P’s credit point is that moratoriums, zoning changes, and fights over noise, water, and power can pressure project financing, not just timelines. [22] [23]

The non-obvious pattern: many of the projects that lost had already offered the industry’s standard fixes—own generation, closed-loop cooling, tax payments, community benefits. Ashville’s EdgeConneX gas plant and Millville’s power-and-water objections show the fight has moved from “who pays for the substation” to “we do not want the load at all.”

How much 2027–2028 AI capacity is at risk

These are analyst estimates, not utility-approved load. They answer different questions—power shortfall, credibility of announcements, and political exposure—and should not be collapsed into one number.

2027 looks mostly built. 2028 does not.

SemiAnalysis’s November 1, 2026 model treats 2027 as largely locked: roughly three-quarters of the 33.8 GW due across its top 20 states was already under construction, which is why it still expected a global year of about +50 GW. The fight, in that model, is 2028: 62.2 GW of U.S. net additions, most of it not yet started. About 30 GW of planned 2028 capacity sat in states where pause candidates were running (versus about 5.7 GW if only “conditions, not a pause” candidates won). Texas alone was modeled at close to 30% of U.S. gigawatts in 2028. The single largest forecast risk SemiAnalysis named was a prolonged freeze of Texas Commission on Environmental Quality permits, which Abbott had already paused on September 21, 2026, on top of an ERCOT interconnection-queue audit begun August 3. [24] [25]

Political exposure after the election is pipeline share, not cancelled megawatts.

• Goldman Sachs economists, as reported November 4: states likely to tighten rules hold about 45% of planned U.S. data-center capacity. They expect temporary pauses in Texas, Georgia, and Michigan while legislatures write rules on power, water, and tax breaks, and they describe those curbs as likely short-lived, with some spending possibly pulled into 2027. Texas alone is roughly a quarter of planned capacity. This is exposure, not a cancellation forecast. [26] [27]

• Cleanview: nine states whose governor or governor-elect has promised or already imposed a pause hold 158.6 GW, or 41% of the U.S. pipeline. Another 69.2 GW (18%) is in states where the winner wants tax or rule changes but has not backed a pause. Cleanview also notes 56.7 GW already under construction nationwide is less exposed, and that the megawatts directly hit by November 3 local ballots were small—at least 1.6–1.9 GW among projects with a reported size. [16]

• Wells Fargo: Texas holds about 27% of planned U.S. data-center gigawatts; a separate Wells note put Texas at about 20% of the cumulative pipeline and close to 30% of planned additions through 2028. [28] [29]

Power and delivery risk was already large before any ballot.

• Morgan Stanley: a U.S. data-center power deficit of about 32–33 GW through 2028, roughly 34% of projected needs even after behind-the-meter turbines, fuel cells, and nuclear offtake. A later October 2026 vintage, as summarized by secondary outlets, put the pre-fix shortfall near 70 GW for 2026–2028 and the residual gap near 43 GW. An earlier September note had the cumulative midpoint shortfall rising from 5 GW in 2026 to 12 GW in 2027 and 33 GW in 2028. These are model estimates of energization risk, not politics. [30] [31] [32] [33]

• Goldman Sachs, May 2026 (via Utility Dive): only 50–60% of planned data-center capacity would come online as expected over the next two years, citing delays and cancellations. In an October 5 note, Goldman still called the growth outlook through 2027 “largely intact,” raising its 2026 year-end U.S. capacity forecast by 5 GW to 64 GW and cutting 2027 year-end by 5 GW to 90 GW, with power-demand growth of 12 GW in 2026 and 17 GW in 2027. [17] [34]

• Bernstein, as summarized from the Financial Times: only about 35% of the planned U.S. pipeline is “credible” given grid-connection timelines, transformer shortages, and liquid-cooling bottlenecks. A separate Bernstein construction note estimated 35–40% of announced capacity globally at risk of delay or cancellation, with power—not capital—as the gating factor. [35] [36]

• Seaport Research’s Jay Goldberg (Barron’s, as reported): only 30 GW of 55 GW announced for 2027 materializes—about 45% fails—because of electricians, gas turbines, and power connections. That is a bearish single-analyst estimate, not a consensus. [37]

What “at risk” means for 2027–2028, labeled as estimates. A reasonable reading of the public figures, not a new forecast: most of the capacity that can actually energize in 2027 is already in steel and is not what the ballots stopped. The capacity that can slip or die is the unbuilt 2028 slice. SemiAnalysis’s ~30 GW of 2028 additions in pause-candidate states, Goldman’s ~45% of the planned U.S. pipeline now inside tighter-rule states, and Cleanview’s 158.6 GW / 41% of the full pipeline in pause states are the political bounds. Morgan Stanley’s ~32–43 GW through 2028 is the power bound, and it overlaps the political one wherever a moratorium also freezes behind-the-meter gas. Goldman’s own economists expect pauses to be temporary. The downside case is a Texas permit freeze that lasts through the 2027 legislative session and pushes a high-teens-of-GW Texas 2028 cohort to the right.

How the stocks actually traded on November 4

Election Day was November 3. The price discovery was November 4, with Democrats taking both chambers and Gina Hinojosa defeating Greg Abbott in Texas on a data-center moratorium. The tape was a split, not a uniform AI selloff. The Dow rose on the order of 1,300–1,500 points (about 2.6–3%) and the S&P 500 about 1.6–2%, the classic gridlock bid. Data-center builders, merchant generators, and cooling/power equipment did the opposite. [38] [39] [40]

Hyperscalers — up, on a scarcity-and-capex story.

• Amazon about +6.1% to +7.4%
• Microsoft about +6.7%
• Apple about +5.3% to nearly +7%, to a record in some accounts
• Alphabet about +3.2% to +4.6%
• Meta about +1.4% to +2.7%

Reuters said that if those gains held, the five largest tech names were on course to add more than $750 billion of combined market value. The mechanism investors bought: the companies that already own capacity get a scarcer asset, and a slower buildout eases the cash-flow fear that had been punishing them. Reuters put hyperscaler capex plans above $740 billion this year, up from roughly $400 billion last year. Neoclouds did not get that benefit—CoreWeave was down roughly 4–10% depending on the print time. [38] [41] [42]

Nvidia — down, but not a crash. Closed at $200.13, −3.17% from $206.69. That followed a +2.29% session on November 3 and sat on top of an already weak month (about −16% over 30 days into November 4). The election was a incremental negative for 2027–28 chip pull-through, not the start of the drawdown. [43] [41]

Power and equipment — the real selloff.

Name	November 4 move	Close / context
Vistra (VST)	about −13.5% to −14.5%	Close about $119.31–$120.29 vs prior close $139.59. KeyBanc cut it to Sector Weight, citing Texas exposure as the largest among names it covers. [44] [45] [46]
Constellation (CEG)	about −6.7% to −9%	Close $245.72, −6.71% on one end-of-day tape; intraday headlines ran as low as −8% to −9.2%. Signed Google and Amazon deals were not torn up; the hit was future offtake. [46] [47] [48]
Talen (TLN)	−8.5%	Close about $326.78–$327.05 vs November 3 close $357.28. That November 3 session was +7.29% on earnings, so the election day gave back the earnings pop and more. [49] [50] [51]
GE Vernova (GEV)	−9.3%	Close $832.83 from $918.20. Morning headlines had it −8% to −11.5%. Only about 20% of its contracted gas-equipment gigawatts were for data centers as of midyear, so the move was sentiment more than booked backlog. [41] [52]
Vertiv (VRT)	−12.9%	Close $205.55 from $235.97. Same day it announced a three-year Infineon power-semiconductor supply deal—positive operations news that did not stop the tape. Q3 sales had just jumped 61% and organic orders more than doubled. [28] [41] [53]

Utilities — bifurcated, not uniformly down. Merchant independent power producers with data-center offtake (Vistra, Constellation, Talen, and NRG, the last down about 13.5% in one MarketWatch account) were the losers. Regulated names were a mixed political trade: Heatmap/Jefferies flagged Southern as the clearest commission-control risk after Democrats swept both Georgia PSC seats, because Georgia Power is about 60% of Southern’s earnings. Renewables-tilted utilities and pure solar did the opposite—NextEra about +2.3%, First Solar about +7% to +9.6%, Enphase about +3.6% to +10%. The market was repricing who gets to build gas for AI, not “utilities” as a sector. [54] [29] [51]

What else was moving the same tape

The election was the catalyst for the data-center complex. It was not the only driver, and treating every November 4 print as a pure referendum on AI power overstates it.

• Gridlock bid vs. AI-capex fear. The broad indices rose because divided government reduces the odds of new sweeping policy. That same logic is why hyperscalers rose: investors had spent the prior days selling them on capex and return-on-invested-capital fears. A slower buildout is good for the owners of existing compute and bad for the companies that sell the next gigawatt of turbines, switchgear, and liquid cooling. Citi’s Scott Chronert told Reuters a slower buildout changes the trajectory for many of those suppliers. [38] [39]

• The Texas surprise, not the House, was the incremental shock. A Democratic House was widely treated as partly priced. Hinojosa’s win was not. KeyBanc called it a negative surprise and downgraded Vistra the same day, which amplified a move that was already sector-wide. [44] [28]

• Company-specific noise cut both ways and did not explain the sector. Talen’s November 3 earnings pop makes the November 4 decline look larger versus the prior close than versus the pre-earnings level. Vertiv’s Infineon agreement and its just-raised outlook were ignored. Vistra had closed the Cogentrix gas-plant purchase on October 30; Constellation had a conditional Department of Energy loan commitment of up to $5 billion the prior week. Fundamentals did not break on November 4. The multiple on future large-load growth did. [50] [53] [52]

• Other election trades ran in parallel. Defense and prison operators fell on expectations of tighter oversight of Trump-era spending and immigration policy. Solar and healthcare rallied on the prospect of a friendlier federal backdrop. One midday compilation also had Brent crude down sharply (about 8.9% to $94.29); if that print held, the energy complex had a commodity overlay on top of the data-center de-rating, which would make a pure “AI power” read of every generator a bit too clean. [54] [40] [42]

• Analysts already called the move partly a knee-jerk. Raymond James described the sweep as a significant policy shock. Morningstar/MarketWatch quoted analysts warning of an “excessive knee-jerk,” noting Vistra had already hedged a large share of expected 2027 generation and that new officials may not slow the boom as much as the first print implied. Goldman’s economists, the same day, framed state pauses as temporary. Evercore called the setup “incrementally more negative sentiment” for AI-infrastructure names, not a broken order book. [27] [52] [29]

What this means if you are trying to build or underwrite the next campus

The binding constraint has moved from “can we get a chip allocation” to “can we get a permit, a gas molecule, and a community that will not put 25 MW into the charter.” Projects that still depend on annexation, a new air permit, or a state land crossing—Project Jupiter is the template—are the ones that slip even when the building is already topped out. Projects already under construction in 2027 are the ones the election did not stop.

For capital, the November 4 split is the underwriting rule. Existing hyperscale capacity was bid up because scarcity has political value. Uncontracted merchant power and the equipment that only gets ordered when a new campus is permitted were marked down because Texas, Georgia, and Ohio just demonstrated that a campaign can freeze the next increment. Goldman’s 45% and Cleanview’s 41% are exposure maps, not loss rates. The loss rate depends on whether those pauses are the short legislative timeout Goldman expects or the multi-quarter TCEQ freeze SemiAnalysis flagged as the single biggest risk to the 2028 number.

---

Recent Findings Supplement (November 2026)

Recent trackers show dozens of US data center projects cancelled or withdrawn in 2026, with local opposition as the dominant factor and power/grid issues secondary.[1][2]

Programs.com’s Nov. 4, 2026 update lists 69 cancelled projects (≥100 MW or $1B) since Jan. 2024 across 26 states, equating to 15,910 MW of reported planned capacity and $165B in announced investment. Local opposition accounts for 48% of reasons cited; zoning/permitting for 22%; power/grid for just 1%. Sue Data Centers tracks 164 cancelled sites totaling 14.1 GW and $69.9B as of Nov. 4, with 102 linked to documented opposition. Heatmap reported at least $260B cancelled in 2026 alone through local fights (as of mid-Oct.). S&P Global tallied 146 delayed/cancelled projects through ~Aug. 2026, with ~75% tied to local pushback.[3][4]

• Examples include Green Data’s 2 GW Socorro, NM campus (withdrawn June 2026 after opposition and moratorium); A1 Data Center’s 1.4 GW Millville, NJ project (blocked by citywide ban May 2026); Lovejoy, GA 1.3 GW campus (withdrawn Aug. 2026); and multiple smaller Oct. 2026 denials/withdrawals in PA, WI, IA, CO, NV, GA, and MI often citing opposition or zoning.[1][2]
• FT reported 47+ large projects delayed/withdrawn/blocked in 2026; 10a Labs/Data Center Watch counted 75 projects (~$130B) blocked/delayed in Q1 2026 and 45 (~$68B) in Q2.[5]

This scale indicates local resistance is now a structural filter, not isolated friction; developers are shifting to friendlier jurisdictions while hyperscalers face higher costs or delays for remaining sites.

Voters and newly elected officials imposed widespread curbs in the Nov. 2026 midterms, directly blocking or exposing specific large projects.[5][6]

Voters in at least 33–38 US jurisdictions passed bans, moratoriums, energy caps, or public-vote requirements. Only one pro-construction measure passed (North Dakota). Trenton, OH voters banned >25 MW data centers (80% support), blocking annexation for a planned 1,000 MW Amazon campus. Other examples: Piqua, OH measures targeting Meta’s 1.4 GW campus (mayor/commissioners recalled); Oregon, OH capped at 25 MW (affecting Vantage plans); Joliet Township, IL (~80% support for restrictions). Democratic gubernatorial wins in Texas (Gina Hinojosa, pledging moratorium), Ohio, and Georgia add state-level risk.[7]

Nine states now have governors/governors-elect promising pauses or stricter rules. Cleanview estimates these could pause 41% of planned US capacity; local ballot measures directly affect at least 1.6–1.9 GW with reported sizes. S&P and others note rising credit risk from these delays.[8]

For entrants or competitors, this raises the bar for community engagement, transparency (hyperscalers like Amazon, Meta, Google, and Microsoft have begun reducing NDAs or disclosing usage), and site selection toward states with stable or supportive policies.

Power constraints are causing targeted cancellations and delays, though less common than opposition.[9]

Oracle sought financial protections for its 2.5 GW Project Jupiter (Doña Ana County, NM) after regulators blocked a supporting gas pipeline. Microsoft terminated its anchor lease on Nscale’s Loughton, UK 50–90 MW AI campus (23,040 Nvidia GB300 GPUs planned) after UK Power Networks indicated energization not until the early/mid-2030s. Satellite imagery shows delays at Microsoft-partner IREN Texas sites (equipment shortages) and a New Jersey Nebius project (~140 MW at risk of slipping to 2027). Goldman Sachs (May 2026) estimated only 50–60% of planned capacity would come online on schedule amid such issues.[10][11]

These cases illustrate that even advanced or hyperscaler-backed projects can unravel when grid timelines slip beyond AI deployment cadences; behind-the-meter generation or co-location strategies are becoming more critical.

Analyst estimates indicate substantial 2027–2028 AI/data center capacity at risk, primarily from permitting, grid, and political factors, with 2028 more exposed than 2027.[12][13]

SemiAnalysis (Nov. 1, 2026) models 33.8 GW US net additions in 2027 (roughly three-quarters already under construction, largely locked in) versus 62.2 GW in 2028 (most not yet started). Roughly 30 GW of the 2028 pipeline sits in states with pause-favoring candidates (vs. ~5.7 GW under conditions-only scenarios); Texas alone represents ~30% (~18.7 GW) and faces a TCEQ permitting freeze plus grid audit. Bernstein pegs only ~35% of the overall announced US pipeline as credible due to grid timelines, transformers, and cooling bottlenecks. Seaport’s Jay Goldberg estimates just 30 GW of 55 GW announced for 2027 will materialize (45% shortfall) from electrician/turbine/grid shortages. Morgan Stanley projects a 32 GW US power gap by 2028 (~34% of needs), potentially widening to 72 GW by 2029.[14][15]

For 2027–2028 planning, the credible pipeline is far smaller than headline announcements (“bragawatts”); those with secured sites, power, and construction underway hold a decisive edge, while pure announcements face high attrition.

Hyperscalers and select tech names rallied after the Nov. 2026 midterms while power/infrastructure and chip stocks sold off, reflecting divergent bets on spending pace versus buildout volume.[16][17]

On Nov. 4, 2026 (post-Democratic congressional sweep and key state wins), Amazon rose ~6–8%, Microsoft ~4–7%, Alphabet/Google ~3–6%, and Meta ~2–4% (combined market value gains cited up to ~$750B–$1T), as investors anticipated slower data-center spending could preserve cash flow for hyperscalers. Nvidia fell ~1–3%, with broader chip/memory names (Broadcom, Micron, Intel, SK Hynix) also lower. Power and infrastructure stocks dropped sharply: Vistra –10–14%, Constellation Energy –6–8%, GE Vernova –7–10%, Vertiv –9–13%; smaller names like Modine (–24%), Sterling Infrastructure (–19%), and Comfort Systems (–16%) saw steeper declines. CoreWeave and Nebius (neoclouds) fell ~7–10%.[18][19]

The moves occurred against a broader market rally (S&P +~2%, Nasdaq +~2.6%) and followed pre-existing pressure on merchant generators. Analysts described the power-stock reaction as potentially knee-jerk or “incrementally negative,” noting strong recent order backlogs (e.g., GE Vernova, Vertiv) and that many 2027 projects are already advanced. Hyperscaler gains were framed as relief from capex intensity rather than reduced AI demand overall.[20]

These reactions highlight a split in the AI trade: slower infrastructure rollout may benefit hyperscalers’ margins short-term but pressure upstream suppliers and power names more immediately. Other drivers (earnings, backlog updates, pre-election positioning) were already influencing prices; the election amplified sentiment around regulatory risk, particularly in high-growth states like Texas.