Power availability has replaced GPU supply as the binding constraint on AI infrastructure expansion, but water access and the regulatory environment are closing in as co-equal obstacles. The average interconnection timeline in PJM, the largest US grid operator, has stretched from under two years in 2008 to over eight years in 2025, while community opposition has blocked or delayed $64B in data center projects since mid-2024. Over 200 state-level bills targeting data center operations were introduced in 2025 alone, with moratorium proposals active in at least eleven states as of early 2026. The companies securing gigawatt-scale capacity are the ones treating site selection as an energy and water procurement problem layered with regulatory risk, rather than a real estate transaction.
"My problem today is not a supply issue of chips; it's actually the fact that I don't have warm shells to plug into. If you can't get the builds done fast enough close to power, you may actually have a bunch of chips sitting in inventory that you can't plug in."Satya Nadella, CEO of Microsoft, November 2025
The scale of planned investment is staggering. Major technology companies are estimated to spend $650B on AI data centers in 2026, with nearly 100 GW of new capacity planned between 2026 and 2030, which would double the global installed base.[1] Colocation vacancy in North American primary markets fell to a record 1.4% at year-end 2025, with 92% of capacity under construction already preleased.[2] Wholesale asking rents hit $195.94/kW-month in H2 2025, a 6.5% year-over-year increase, and are expected to breach $200/kW-month in primary markets during 2026.[2] The question facing developers is no longer whether to build, but where they can actually get power on a timeline that matters.
Speed to power has emerged as the primary criterion in site selection, according to both CBRE and JLL in their 2025-2026 outlook reports, followed by community support, latency, and proximity to customers.[1] In constrained metros, colocation now behaves less like a real estate product and more like a power access product, because the hardest thing to secure is deliverable megawatts on a timeline customers can underwrite.[3]
The grid strain is concentrated in specific regions. PJM Interconnection, which serves 65 million people across 13 states, projects a 6.6 GW shortfall against its reliability requirements for the 2027-2028 delivery year, with capacity prices hitting a record $333.44/MW-day at auction.[4] ERCOT's large-load interconnection queue nearly quadrupled in a single year to 230 GW in 2025, with over 70% of requests coming from data center developers.[5] Up to 11 GW of data center capacity announced for 2026 shows no signs of active construction, raising questions about how much planned capacity will actually materialize.[6]
The response from the largest AI labs has been to bypass the grid entirely. In October 2025, OpenAI and Oracle placed the largest order ever for onsite gas generation: a 2.3 GW plant in Texas for the Stargate project.[7] xAI doubled its gas turbines at the Memphis Colossus facility to 35 units totaling 422 MW, operating beyond its permitted 15-turbine limit while awaiting expanded authorization.[8] Microsoft recommissioned the Three Mile Island nuclear plant to deliver 819 MW, though it will not come online until 2028.[9] Bring-your-own-power (BYOP) has shifted from a niche strategy to a near-requirement for any facility above 100 MW.
Community opposition has emerged as a force that rivals grid constraints in its ability to delay or kill projects. Data Center Watch, an industry tracking organization, reports that opposition campaigns have blocked $18B and delayed $46B in US data center projects since mid-2024, with cancellations accelerating from 2 projects in 2023 to 25 in 2025.[10] By June 2025, the combined figure surged past $98B in blocked or delayed projects across the US and Canada, a 125% increase in three months.[11]
The moratorium movement has jumped from local town boards to state legislatures. Communities in at least 14 states have enacted temporary pauses on data center development, and multiple states introduced moratorium bills in early 2026: Georgia HB 1012 would bar new data center permits until March 2027, New York S9144 proposes a statewide moratorium, and Oklahoma SB 1488 would halt new construction until November 2029.[12] At least 188 organized opposition groups span 40 states, and Senator Bernie Sanders became the first member of Congress to call for a national moratorium in December 2025.[10]
Resident concerns center on energy consumption, diesel generator emissions, rising electricity costs (monthly power prices inflated by as much as 267% between 2020 and 2025 in affected areas), and water usage.[11] In Alberta, a proposed 450-hectare data center technology park was rejected in September 2025 specifically over water concerns, following a series of water-main breaks in Calgary.[13] Data centers in Texas alone consumed an estimated 49 billion gallons of water in 2025, with projections reaching 399 billion gallons by 2030.[14]
Power draws the headlines, but water increasingly determines whether a permitted project can actually operate. US data centers consumed roughly 17 billion gallons in 2023, according to the Department of Energy, and that figure could double or quadruple by 2028 as AI workloads, which use larger and more power-dense chips, drive heavier cooling loads.[18] Over 40% of planned and existing US data centers sit in areas classified as having "high" or "extremely high" water scarcity, creating a structural mismatch between where capacity is being built and where water is available to cool it.[19]
Volume alone understates the problem; the deeper issue is governance. In states like Louisiana, home to Meta's 2,250-acre Hyperion campus, groundwater users face no permitting requirement beyond registering if they pump more than 50,000 gallons per day, and utilities are sometimes barred by nondisclosure agreements from disclosing how much water a data center actually consumes.[20] Google's experience in Iowa illustrates the other end of the spectrum: after Linn County developed one of the state's most comprehensive data center ordinances — including an independent water-balance study, funded by Google at up to $500,000, to evaluate whether the Cedar River could sustain the roughly 12 million gallons per day the six proposed buildings would require — Google pursued annexation of the same 545-acre site into the neighboring City of Palo rather than remain under county jurisdiction. The land never moved; it stayed in Linn County. What changed was the governing authority: municipal oversight replaced the county's ordinance, and the water-balance study that supervisors had negotiated lost its funding mechanism and stalled. Supervisor Sami Scheetz characterized the maneuver bluntly: "We negotiated in good faith. Google's response was to go find a local government that will ask for less."[21] The sequence reveals the adversarial geometry of AI infrastructure siting: a county imposes environmental safeguards, an operator exploits overlapping jurisdictions to sidestep them without moving the project, and the county loses both its regulatory leverage and the funding mechanism it had negotiated — all while the community that would bear the water impact gains no new protections. For site selectors, overlapping municipal and county authority is not a bureaucratic inconvenience; it is a strategic variable.
Google's annexation maneuver routed around the regulatory body entirely, but in East Texas, a developer tried the opposite approach — securing permits through the existing authority — and the community fought back in court. Conservation Equity Management, a firm controlled by hedge fund operator Kyle Bass, obtained preliminary approval for 40 high-capacity wells above the Carrizo Wilcox Aquifer. After Wayne-Sanderson Farms and local residents challenged the permits, a district court voided the approvals and ordered three independent impact studies before drilling could proceed. Bass's firm then sued the Neches & Trinity Valleys Groundwater Conservation District, claiming board members "unconstitutionally" delayed the permit process.[22][23] The two cases bracket the range of tactics available to both sides: developers can shop for favorable jurisdictions or push permits through sympathetic boards, while communities can impose stringent ordinances or litigate to void approvals already granted. Neither side lacks for legal tools, and the resulting uncertainty adds months or years to project timelines that GPU depreciation schedules cannot absorb.
State legislatures are responding with a wave of regulation that has no recent precedent in the data center industry. Over 200 bills targeting data center operations were introduced across all 50 states in 2025, with more than 40 enacted into law, and the pace is accelerating in 2026.[24] The legislation falls into three categories: transparency mandates, cooling technology requirements, and outright construction moratoria.
On transparency, California's AB 1577 now requires monthly energy and water consumption reports to the State Energy Commission, Iowa's HF 2447 mandates quarterly water usage filings, and Georgia's SB 421 would prohibit local governments from signing nondisclosure agreements that conceal data center resource consumption.[25] On cooling technology, South Carolina's HB 4583 would require closed-loop systems achieving "zero net water withdrawal," prohibiting both groundwater extraction and municipal water use for cooling entirely. Kansas's SB 400 imposes similar closed-loop requirements and grants municipalities and the state attorney general standing to seek injunctions against violators.[25]
The moratorium bills represent the sharpest regulatory tool. At least eleven states had active moratorium proposals in session as of early 2026: South Carolina's H. 5286 would bar all data center permits and incentives until January 2028, Georgia's HB 1012 would freeze permits until March 2027, and Virginia's bill would halt new site applications until July 2028.[26] At the municipal level, Denver announced plans for its own moratorium in February 2026, and Maryland's Frederick County imposed a six-month pause followed by a zoning map that constrains eligible sites, while Baltimore County froze development until 2027.[27] For developers evaluating a three-to-five-year buildout timeline, the regulatory environment in a given jurisdiction may shift materially before construction is complete.
| Factor | Data Point | Source | Conf. |
|---|---|---|---|
| Grid interconnection | PJM avg. timeline: <2 yrs (2008) to 8+ yrs (2025) | RMI, PJM data | High |
| Grid capacity gap | PJM projects 6.6 GW shortfall for 2027-2028 delivery year | PJM capacity auction | High |
| ERCOT queue | 230 GW large-load requests in 2025, 4x YoY; 70%+ from data centers | ERCOT, Latitude Media | High |
| Vacancy rate | Primary market colocation vacancy: 1.4% (year-end 2025) | CBRE H2 2025 | High |
| Construction cost | $10-12M/MW avg., up to $20M+/MW for AI-optimized facilities | JLL, Turner & Townsend | High |
| Land cost | $5.40/sqft avg. for parcels 50+ acres, up 23% from 2023 | Cushman & Wakefield | Med |
| Colocation pricing | $195.94/kW-month wholesale avg. (H2 2025), expected to breach $200 | CBRE H2 2025 | High |
| Community opposition | $64B+ blocked/delayed; 25 cancellations in 2025 (vs. 2 in 2023) | Data Center Watch | Med |
| Moratoriums | 14+ states with enacted or proposed development pauses | Good Jobs First | High |
| Water consumption | Texas: 49B gal (2025), projected 399B gal (2030); 40%+ of US DCs in high-scarcity zones | HARC, Univ. of Houston, Circle of Blue | Med |
| Water regulation | 200+ state bills introduced in 2025, 40+ enacted; SC, KS mandate closed-loop cooling | WilmerHale, MultiState | High |
| Jurisdiction shopping | Google shifted $1B+ Iowa project from county (strict water rules) to city (fewer requirements) | CBS2 Iowa | High |
| Federal land program | DOE selected 4 sites (INL, Oak Ridge, Paducah, Savannah River) | US Dept. of Energy | High |
Power and land constraints in established hubs are driving development into emerging markets. Northern Virginia led all primary markets with 1,102 MW of net absorption in 2025, but its colocation vacancy rate sat at 0.72% in H1 2025, leaving almost no room for new entrants.[2] Dallas-Fort Worth absorbed 470.8 MW, up 424 MW year-over-year, with 89% of under-construction space preleased. Phoenix has 4.2 GW of planned capacity and 1.3 GW under active development.[2]
The next wave of development is moving to markets where land is cheap, grid capacity is available, and community opposition has not yet organized: parts of Texas outside the Dallas-Austin corridor, Wisconsin, Michigan, Ohio, and the Southeast.[15] Markets with low power costs, such as Atlanta, Charlotte-Raleigh, and San Antonio, are positioned for accelerated supply growth according to JLL's 2026 outlook.[1]
The DOE's federal lands program represents a parallel track. In response to Executive Order 14179 ("Removing Barriers to American Leadership in Artificial Intelligence"), DOE identified 16 potential sites and selected four, including Idaho National Laboratory, Oak Ridge Reservation, the Paducah Gaseous Diffusion Plant, and the Savannah River Site, for private-sector AI data center development with co-located energy infrastructure.[16] These sites offer existing energy infrastructure, streamlined federal permitting, and the ability to fast-track new generation including nuclear. Developer selection at the Paducah site was open for proposals through January 2026.[17]
| Dimension | Grid-Connected (Utility) | BYOP (Onsite Generation) |
|---|---|---|
| Time to power | 24-96+ months (PJM avg. 8+ years) | 6-18 months for gas turbines |
| Scale ceiling | Limited by substation headroom and feeder capacity | Theoretically unbounded; xAI at 422 MW, Stargate at 2.3 GW |
| CapEx per MW | $10-12M (facility only; grid upgrade costs borne by utility) | $10-12M + $0.8-1.5M/MW for generation equipment |
| Energy cost | $0.04-0.08/kWh (varies by market) | $0.06-0.10/kWh (natural gas at current prices) |
| Permitting risk | Utility manages; developer waits | Developer manages air quality, emissions permits directly |
| Community risk | Moderate (grid impact diluted) | High (visible turbines, emissions, noise complaints) |
| Water exposure | Municipal supply; subject to local disclosure rules and rate structures | Often well-based; groundwater permits increasingly contested (see East Texas) |
| Reliability | Grid-backed with utility SLAs | Self-managed; fuel supply chain dependency |
| Carbon profile | Grid mix (varies; renewable PPAs available) | Natural gas baseline; carbon offsets or future nuclear required |
| Regulatory trend | FERC co-location reforms improving (Dec 2025 order) | EPA scrutiny increasing; xAI permit violations under investigation |
| Trade-off | Certainty and cost efficiency at the expense of speed | Speed and scale independence at the expense of regulatory and carbon risk |
$650B in planned spending for 2026, with Stargate alone committing $400B+ over three years, demonstrates that capital is not the bottleneck. The DOE federal lands program and FERC's December 2025 co-location reforms signal that federal policy is actively working to accelerate buildout.
PJM's 8-year interconnection timelines, 6.6 GW capacity shortfall, and 11 GW of announced-but-unstarted 2026 capacity suggest that grid infrastructure cannot keep pace. ERCOT's 230 GW queue, the vast majority of which will never be built, reflects speculative demand that obscures the actual supply pipeline.
Over 40% of planned US data centers sit in high or extremely high water-scarcity zones. State legislatures introduced 200+ data center regulation bills in 2025, with 40+ enacted, and moratorium proposals are active in at least eleven states. Developers face a regulatory environment that can shift materially within a single buildout cycle, and jurisdiction shopping (as Google demonstrated in Iowa) only accelerates the legislative response.
Community opposition is a wildcard that pure infrastructure analysis consistently underweights. 188 organized opposition groups across 40 states and moratorium bills in 14+ state legislatures represent a structural risk with no engineering solution, only a political and communications one. Projects that invest in community engagement early can succeed, but the default trajectory is toward more resistance as power prices and water consumption rise in data-center-adjacent communities.