AI Data Centers

AI data centers are the physical infrastructure underlying frontier-AI capability: the computing facilities themselves, the energy grids that power them, the water systems that cool them, the land they occupy, and the permitting and siting politics that constrain their construction. Through 2026, data-center buildout emerged as one of the binding constraints on frontier-AI scaling and a focal point of policy and community contestation, with capacity tightly committed, electricity demand rising, and state and local opposition escalating from project-level fights to statewide policy action.

Capacity and market dynamics

By May 2026, terrestrial data-center capacity was nearly fully committed. JLL reported that 97% of global data-center capacity was occupied and more than 75% of new capacity was already committed. The compute shortage reached end users: the Wall Street Journal reported (April 12, 2026) that Anthropic experienced frequent outages and metered users during peak hours, and that OpenAI scrapped its Sora video app to free GPUs for Spud/Codex (Source: wsj.com). Token use in OpenAI's API rose from 6 billion per minute in October 2025 to 15 billion per minute in late March 2026, CoreWeave imposed a price increase of more than 20% with three-year minimum contracts, and Nvidia Blackwell rental reached $4.08/hr, up 48% in two months (Ornn Compute Price Index). J.J. Kardwell of Vultr described "a massive capacity crunch that's unlike anything I've seen in the more than five years I've been running this business" (Source: wsj.com; supporting source folded into this page).

The shortage produced new market structures. xAI began renting compute to Anthropic and Cursor as Grok's commercial traction lagged, monetizing idle data-center capacity (Source: theinformation.com). The startup Amp, which buys excess data-center capacity to resell, raised $1.3 billion in an a16z-led round (Source: nytimes.com). CME Group and Silicon Data launched a futures market for compute capacity priced off daily GPU benchmarks, the first publicly traded GPU-price derivatives venue (Source: cnbc.com).

Anthropic assembled a multi-source compute stack during this period. Its Colossus 1 arrangement with SpaceX (May 6, 2026) provided 220,000-plus Nvidia H100/H200/GB200 chips and 300-plus MW of immediate capacity, with Anthropic buying 100% of the capacity. A $200 billion, five-year agreement with Google (May 5, 2026) represented more than 40% of Google's disclosed cloud "revenue backlog," and a $1.8 billion, seven-year agreement with Akamai (May 8, 2026) added a non-hyperscaler provider. Akamai's Form 8-K later dated that master services agreement to May 5, 2026 and disclosed two further project plans entered on September 18, 2026, under which Anthropic committed approximately $11.6 billion of dedicated cloud capacity over seven-year terms, expandable by up to $9 billion. Akamai estimated related capital expenditure at approximately $5.5 billion, authorized its contract manufacturer Jabil to buy approximately $1.7 billion of memory components, and issued Anthropic a warrant over up to 387,051 preferred shares (7,741,020 common shares as converted) that vests with Anthropic's payments and further commitments (Akamai Technologies Form 8-K: Anthropic Project Plans and Warrant Agreement (September 2026)).

Hardware suppliers and neocloud operators repriced upward on AI demand. Cerebras priced its IPO at $185/share, above its twice-raised range, selling 30 million shares for $5.55 billion at a $56.4 billion fully diluted valuation; it disclosed that Arm and SoftBank had both attempted to acquire it in the preceding weeks, and larger SpaceX, OpenAI, and Anthropic IPOs were expected later in 2026 (Source: cnbc.com). Nebius reported $399 million in Q1 2026 revenue, a 684% increase year-over-year, among the fastest-growing neocloud operators (Source: theinformation.com). ByteDance raised its 2026 capex to roughly $30 billion (more than 200 billion yuan), up at least 25% from its late-2025 plan, directing a larger share to domestic AI chips (Source: scmp.com). Data-center developer Switch was in talks as of June 4, 2026 to raise billions of dollars at a valuation of at least $50 billion, with Brookfield Asset Management and KKR among prospective investors, Goldman Sachs and JPMorgan advising, and an IPO possible as early as the following year (Source: theinformation.com; reuters.com); this is a private-capital counterpart to the private-credit financing of the same buildout.

On July 21, 2026, a BlackRock-backed investor group committed $5 billion to data-center operator Aligned Data Centers (Source: reuters.com). CoreWeave's Q1 2026 results — $2 billion revenue, a $144 million loss, a $100 billion backlog, and 1 GW of active capacity against an 8 GW 2030 target — are a recurring reference point in the bubble debate (see AI Bubble Debate). US power consumption surged more in the prior two years than in the previous 15 combined (Apricitas analysis), and residential electricity prices rose more than 40% since 2020 and continued to climb.

Stijn Van Nieuwerburgh's Brookings Papers on Economic Activity paper (Fall 2026) sizes the U.S. pipeline from Cleanview project-level data accessed July 23, 2026: about 57 GW of operating capacity and a planned pipeline of about 509 GW, which the author states is not a completion forecast. Of 2,933 non-canceled projects holding 566.0 GW, 662 planned projects with 353.6 GW report no completion date. Assigning realization shares that fall with the reported horizon, his central scenario has 182.7 GW completed in 2025–2032, 117.2 GW after 2032 and 226.9 GW never completed. He costs a representative 200 MW training campus at about $8.216 billion: $2.2 billion for the facility at $11 million per MW, $0.4 billion for incremental power infrastructure, and $5.616 billion for about 1,170 NVIDIA GB300 NVL72 racks at about 142 kW each, $4.0 million per rack before networking, storage and integration. That is about $41 billion per GW, 68% of it IT equipment. At a power usage effectiveness of 1.2, the campus has about 166.7 MW of critical IT load, and a continuous 200 MW load uses about as much electricity a year as 170,000 average U.S. households. The paper also reports that the 36-month market beta of data-center REITs rose from 0.53 in 2021 to 1.00 in 2026, and their beta with the S&P 500 excluding AI stocks from 0.36 to 0.76. The author reads this as data centers behaving less like defensive infrastructure and more like part of the same systematic risk as AI stocks (Financing the AI Buildout (Van Nieuwerburgh, BPEA Fall 2026)).

CoreWeave's ratio of 1 GW active versus its large backlog functions as a marker in the AI Bubble Debate: bulls read a closing gap as validation of demand, bears read an inverting gap as overbuild, and whether the gap closes, inverts, or remains structural is unresolved.

Individual hyperscaler campuses continued to grow in committed scale. On July 13, 2026, Meta committed an additional $40 billion to its Richland Parish, Louisiana data-center campus, known as Hyperion and its largest, bringing the announced expansion to $50 billion for at least 5 GW of capacity and total expected investment at the site beyond $250 billion, with 2 GW targeted by 2030 and completion around 2032 (Source: bloomberg.com) (Source: cnbc.com).

Snapshot

Newest on top. Dated quantitative and deal facts.

Energy demand and the physical footprint

Inside the Dirty, Dystopian World of AI Data Centers — Matteo Wong (The Atlantic, April 2026) (March 2026) provides primary numerical anchors for terrestrial data-center scale. The xAI Colossus facility in Memphis consumes roughly the annual electricity of 200,000 American homes, runs up to 35 natural-gas turbines on-site, and used 11 million gallons of water in September alone; the combined Colossus and two nearby xAI facilities draw about 2 GW, roughly twice the consumption of Seattle. OpenAI has set a target capacity of more than 30 GW, described as more than the largest recorded demand for all of New England. Big Tech AI capex since ChatGPT — Amazon, Microsoft, Meta, and Google combined — exceeds $600 billion, which inflation-adjusted is more than the federal interstate-highway system. The IEA projects that by 2030 US data centers will consume more electricity than all heavy industry combined, spanning cement, steel, chemical, auto, and other industrial sectors. The Department of Energy's Lawrence Berkeley National Laboratory, in analysis reported June 24, 2026, projected that data-center electricity demand could exceed 15 percent of total US consumption by the end of the decade, a figure the department paired with a call for efficiency gains (Source: insideaipolicy.com). The same day, the Federal Energy Regulatory Commission pressed grid operators to revise their market rules for large power users so that data-center costs are not shifted onto other customers (Source: insideaipolicy.com).

Regional concentration is most acute in Northern Virginia. Loudoun County, Virginia hosts 199 operating data centers with roughly 30 in the pipeline, representing 13% of global data-center capacity within 520 square miles, and data centers fund nearly 40% of the county budget. Dominion Energy forecasts 5.5% annual electricity-demand growth in Virginia, doubling by 2039, which it has called the largest growth in power demand since the years following World War II.

Utility ownership began to consolidate on AI-driven demand. The Financial Times reported on May 15, 2026 that NextEra and Dominion Energy had entered talks to combine into an estimated $400 billion US utility, explicitly driven by electricity demand for AI data centers (Source: ft.com). The companies announced the deal on May 18, 2026 as an all-stock merger valued at approximately $67 billion, materially smaller than the FT's $400 billion combined-entity figure, that would create the world's largest regulated electric utility by market capitalization, serving roughly 10 million customer accounts across Florida, Virginia, and the Carolinas (Source: pbs.org). Under the terms, Dominion shareholders receive 0.8138 NextEra shares per Dominion share plus a $360 million one-time cash payment at closing; NextEra holders own 74.5% of the combined entity and Dominion shareholders 25.5%; NextEra CEO John Ketchum serves as chairman and CEO; and closing is expected in 12 to 18 months (mid-2027 to late 2027), pending Nuclear Regulatory Commission and other approvals. In morning trading on May 18, Dominion stock rose 9.6% and NextEra fell about 5%. Both companies framed the combination as a response to AI-driven electricity demand, and Dominion powers hundreds of Northern Virginia data centers. NRC approval is the binding regulatory step.

Congressional attention to grid-cost allocation advanced on July 21, 2026, when the House Energy and Commerce Committee approved a data-center energy package including the Ratepayer Protection Act (H.R. 9340, passed 52-0), which would have states consider making data centers with 100 MW of demand pay full grid-upgrade costs (Source: broadbandbreakfast.com; energycommerce.house.gov). The bill text defines the covered customers by function and size, as non-residential consumers whose facilities take electricity primarily for information-technology infrastructure with 100 MW or more of aggregate peak demand at a single site or campus, and does not use the words "artificial intelligence" or "data center." It works through PURPA's consider-and-determine mechanism: state regulators must begin considering the standard within one year and reach a determination within two, may decline to adopt it, and are exempt where the state has already implemented, held a proceeding on, or legislatively voted on a comparable standard (Ratepayer Protection Act, H.R. 9340 — text as reported in the House (September 10, 2026)). The House passed the bill 417–3 on September 16, 2026 (Source: clerk.house.gov); the Senate has not acted on it, and a companion bill introduced by Sen. Jon Husted, S. 5028, is in the Energy and Natural Resources Committee (Source: congress.gov). President Trump meanwhile expanded his voluntary ratepayer-protection pledge to 23 governors and 187 companies (Source: apnews.com; thehill.com). Reporting published July 24, 2026 described the committee as having cleared a set of data-center and energy-infrastructure bills on bipartisan votes ahead of the summer recess, and said OSTP Director Michael Kratsios expressed interest in working with Rep. Josh Riley (D-NY) to codify the ratepayer-protection pledge in statute (Source: insideaipolicy.com; insideaipolicy.com). On the environmental-permitting side, an EPA interpretation that became public by July 24, 2026 exempts "islanded" data-center power generation — on-site generation not connected to the grid — from Clean Air Act acid-rain provisions (Source: insideaipolicy.com). In the Senate, Mark Warner introduced legislation, made public July 23, 2026, conditioning a tax exemption for AI data center developers on efficiency standards plus energy- and water-use reporting (Source: insideaipolicy.com).

Researcher estimates published July 16, 2026 put hyperscaler power growth in longer perspective: Google's electricity consumption rose more than 140% between 2021 and 2025, and Google, Microsoft, Amazon, and Meta added combined new power demand roughly twice New York City's annual consumption between 2022 and 2025. In the same coverage, MIT economist Daron Acemoglu argued the AI boom could slow on economic grounds if investment keeps outpacing demand (Source: axios.com). See AI Bubble Debate.

Hyperscaler procurement has also turned to dedicated renewable capacity. During the week of July 13, 2026, Google announced its largest solar-and-battery purchase to date: the Steel River Energy Center in Arkansas, adding 1 gigawatt of solar and 1.9 gigawatt-hours of battery storage in its first two phases, backed by $3.5 billion in secured financing, and scaling to about 1.8 gigawatts by 2029 — which would make it the largest solar facility in the United States. TechCrunch noted the site sits 40 miles north of xAI's unpermitted gas-turbine power plant serving Colossus 2 (Source: techcrunch.com).

A range of energy suppliers raised capital specifically to serve data-center load. Fervo Energy ran a roughly $2 billion IPO at a valuation above $10 billion to supply geothermal power for AI data centers, the first major IPO of an AI-data-center power supplier separate from a utility (Source: theinformation.com). Fusion startup Helion Energy said on June 4, 2026 that it had raised $465 million in a Thrive Capital-led round, nearly tripling its valuation to $15.5 billion on demand for fusion power for data centers, a datapoint for the fusion and advanced-nuclear strand of the AI-power story given that hyperscaler demand is now large enough to underwrite frontier energy R&D ahead of commercial deployment (Source: theinformation.com; see AI Power Concentration).

Demand flexibility emerged as an industry answer to interconnection delay. On September 16, 2026 Emerald AI, Google and Nvidia launched the AI Energy Management Alliance, with 18 launch partners including Anthropic, National Grid, AES, Constellation, NRG and RWE, to advocate faster grid connections for data centers that can reduce their power draw during peak grid stress through colocated generation, storage or computational flexibility (Source: businesswire.com). The group is a relaunch of the Advanced Energy Management Alliance, founded in 2014 to advocate for demand-response companies. Google's head of advanced energy market innovation, Tyler Norris, said the company has committed 1 gigawatt of reducible power demand through utility agreements nationwide, and Emerald AI chief executive Varun Sivaram said facilities should qualify for faster or larger connections only if their ability to cut demand is "verifiable and enforceable" (Source: axios.com). Emerald AI had raised a $150 million Series A at a $1.05 billion valuation the previous month (Source: businesswire.com).

Manufacturing economics

Per-transistor cost stopped falling in 2011, and the EUV cost-down curve has since reversed, which Azhar and Petrovic (May 11, 2026) describe as the breaking of the economic bargain underlying Moore's Law; they assess that TSMC will not deploy High-NA EUV through 2029 (Source: exponentialview.co).

Hardware retirement and the secondary market

Datacenter accelerators are typically phased out on roughly three-year cycles aligned with warranty periods, so each generation transition pushes large volumes of still-functional hardware into secondary markets. A 2026 Cambridge, Oxford and Quettaflop AI preprint puts NVIDIA's 2023 datacenter GPU shipments at over 3 million chips and reports second-hand V100 16GB prices falling to about $60 by March 2026, against a $10,000 launch price in May 2018, with A100 40GB units at about $1,900 (DumpsterCluster: From Dumpster Diving to Serving LLaMA-70B on $60 GPUs). The authors report building a 128-GPU pod entirely from retired components — GPUs, CPUs, memory, motherboards and NICs — for about $32,000 in 2024, falling to about $22,000 at March 2026 prices, and running it in production for one year at a measured hardware failure rate of roughly 6.25% of devices per year, with 86.2% of the 58 incidents logged recoverable in software rather than requiring replacement.

The paper frames the secondary market as a supply-chain channel as well as a price one: it is geographically distributed, less exposed to export controls on frontier technology, and available without the six-to-twelve-month lead times reported during recent GPU shortages, which matters for buyers facing vendor prioritization of hyperscale customers (Export Controls (AI)). It also notes an adoption paradox — the prices that make the approach attractive reflect weak demand for older parts, and systematic uptake would raise them. The environmental accounting is treated at AI Environmental Impact and the cost-per-token comparison at Inference Economics and Token Pricing.

The short economic life of accelerators is also a financing issue. Van Nieuwerburgh's BPEA paper notes that GPUs "may become economically obsolete within three to six years," while the buildings and power connections financed alongside them last decades, and that capacity commissioned in 2025–26 may need substantial GPU replacement by around 2030. He identifies the mismatch between long financing maturities and shorter asset lives as one of the buildout's main vulnerabilities. In his revenue model, shortening the assumed IT life from six years to three raises the capacity's required mature annual revenue from about $3.7 trillion to about $6.0 trillion (Financing the AI Buildout (Van Nieuwerburgh, BPEA Fall 2026)).

International and sovereign-compute buildout

National AI-infrastructure commitments expanded across Europe and Asia-Pacific. SoftBank announced on May 30, 2026 that it would invest up to €75bn ($87bn) to build and operate 5 gigawatts of AI data-center capacity in France, described as Europe's largest AI-infrastructure project. An initial €45bn ($52bn) phase delivers 3.1 GW across the Hauts-de-France region, with sites at Dunkirk, Bosquel, and Bouchain, by 2031, and SoftBank partnered with Schneider Electric on a Dunkirk data-center-component manufacturing and robotics hub. Masayoshi Son framed France, chosen for its low-carbon nuclear grid and fast-tracked AI-facility approvals, as "uniquely positioned to become a leading AI infrastructure hub in Europe." The commitment is SoftBank's largest AI bet outside the US, following its 10 GW Ohio project and its $60bn+ OpenAI stake; the 5 GW target is roughly New York City's peak electricity demand and will require substantial additional financing from unnamed partners. The announcement preceded an investment conference hosted by President Emmanuel Macron, who has pressed for nuclear-powered domestic data centers to narrow the AI gap with the US and China (Source: ft.com; wsj.com). In a June 1, 2026 briefing with Macron, Son confirmed the full €75bn (~$87bn) / 5 GW plan, following the first €45bn phase (3.1 GW across Dunkirk, Bosquel, and Bouchain by 2031) announced May 31. Separately, SoftBank overtook Toyota to become Japan's most valuable company (Source: cnbc.com; techzine.eu).

Australian data-center operator AirTrunk announced on June 5, 2026 a $30 billion investment to build 5 gigawatts of data-center capacity in India by 2030, with CEO Robin Khuda meeting Prime Minister Narendra Modi the same day, extending the Asia-Pacific leg of the buildout at a scale (5 GW) comparable to the SoftBank–France project (Source: reuters.com; see AI Power Concentration). India's RMZ said on June 19, 2026 that it planned a $35 billion push to expand data-center capacity, a further sign of concentrated buildout in the Indian market (Source: reuters.com). Tata Consultancy Services said on September 5, 2026 that its subsidiary HyperVault and partners would invest up to 700 billion rupees ($7.41 billion) in a 1-gigawatt campus in Telangana, having secured 264 acres in Hyderabad for phased development aimed at AI companies and hyperscalers; TCS chief executive K Krithivasan said Hyderabad offered the scale, talent and ecosystem needed to serve global customers, and the company said the campus is expected at full buildout to rank among India's largest AI infrastructure facilities (Source: reuters.com). Also on June 19, 2026, Abu Dhabi's MGX was reported to be weighing a multi-billion-dollar acquisition of Singapore-based data-center operator DayOne, which had been pursuing a US IPO at a roughly $20 billion valuation (Source: reuters.com). A Microsoft-G42 $1 billion geothermal data center in Kenya stalled around May 10, 2026 after the Kenyan government declined to guarantee annual capacity payments (Source: reuters.com).

The Asia-Pacific buildout extended to frontier-lab leases in Australia. Anthropic's first Australian data-center lease, at Zerra DC's proposed Western Downs Digital Park in Queensland, became public on September 16, 2026; two people familiar with the deal told Reuters the campus has planned capacity totalling 2.16 gigawatts about 250 km from Brisbane, is planned to begin coming online in 2027, would run inference rather than training, and is subject to Foreign Investment Review Board approval. Reuters described Australia as pitched to AI firms as a low-cost, renewable energy-rich alternative to congested markets in the U.S. and Asia, noted OpenAI's offtake agreement for a Sydney data center the previous year, and reported that the Australian government has promised to set restrictions on data-center resource use from 2027 while economists estimate the investment boom may total A$150 billion ($107 billion) by 2030 (Source: reuters.com). The campus cost has been reported at about A$30 billion (Source: afr.com).

China's state-directed buildout has drawn parallel scrutiny. A ChinaTalk analysis of CAICT data published July 10, 2026 concluded that China's "Eastern Data, Western Compute" initiative has largely failed to move compute to the country's west: Hebei, Guangdong, and Jiangsu lead chip deployments, state media report one western intelligent-compute center running below 50% rack-up and under 30% utilization, and nearly 400 additional centers were under construction or planned as of late 2024 — a buildout the analysis argues risks leaving poorer western provinces in debt (Source: chinatalk.media).

Beijing set a headline capacity target two months later. A five-year industry plan released by the Ministry of Industry and Information Technology on September 7, 2026 sets a goal of 9,800 exaflops of intelligent computing capacity by 2030 and calls for 3.8 trillion yuan ($532 billion) in cumulative information-infrastructure investment over 2026–2030, including the orderly deployment of clusters of 100,000 or more accelerator cards. Capacity stood at 2,185 exaflops at the end of June 2026 across 52 facilities each holding more than 10,000 accelerator cards, up 177 per cent year on year, and about 2,450 exaflops by the end of July per the National Data Administration (Source: scmp.com). The target and the utilisation findings above bear on each other: the ChinaTalk analysis describes existing western centres running below half rack-up, which the capacity figures do not address.

Orbital data centers

Terrestrial siting pressures, energy constraints, and community opposition have pushed a share of frontier capex into orbital data-center plans. SpaceX has reframed corporate strategy around orbital AI data centers, and Musk displayed an "AI Sat Mini" rendering at a March 2026 event with multi-kilometer solar arrays. Google and SpaceX entered talks on May 12, 2026 to launch orbital data centers in partnership, with Planet Labs as prototype partner under "Project Suncatcher" and prototype satellites planned for 2027. Blue Origin formally entered the orbital-data-center race, Nvidia posted an "orbital data-center system architect" job, and Rocket Lab CEO Peter Beck stated that "solar arrays of multikilometers in scale are what's needed." Data Centers in Space: A Pipe Dream, or AI's Next Big Thing? — Micah Maidenberg, Merrill Sherman (WSJ, May 12 2026) covers the engineering tradeoffs, including heat dissipation as the binding constraint, solar arrays larger than the Starship rocket, and "savage" launch economics per industry skeptics. Startup entrants have followed the incumbents: Orbital Compute Inc., a five-month-old startup with $5 million in pre-seed funding, filed with the FCC — per a July 7, 2026 report — for up to 100,000 orbital data-center satellites cumulatively capable of 10 GW of compute, using 100 kW satellites at 500–850 km altitude, with a 2027 pathfinder demonstrator carrying an Nvidia Blackwell chip (Source: datacenterdynamics.com). Google's first Project Suncatcher test satellite, a refrigerator-sized spacecraft called MVP built by Planet Labs and carrying four TPU accelerators, is scheduled to launch on October 1, 2026 on SpaceX's Transporter-18 rideshare aboard a Falcon 9; the launch date became public on September 24. Its solar panels supply about one kilowatt, and its cooling system can run the chips on Gemini workloads only in spurts of about 15 minutes. The satellite is to operate for a few months, and the two custom satellites Google planned for 2027 remain on its agenda (Source: arstechnica.com).

Rural siting and economic effects

Per Data centers are coming for rural America — Abigail Bassett (The Verge, May 13 2026) (May 13, 2026), Pew Research (April 2026) found that 67% of planned US data centers are headed to rural areas and 39% to counties currently without any. More than 35 states offer incentives or tax breaks to attract data centers (NCSL). A causal analysis of 254 Texas counties by Michael Hicks (Ball State) found net long-term data-center job creation of approximately zero: whatever long-term jobs existed at data centers were offset by losses elsewhere in the same sector. The source defines a "neocloud" as a specialized GPU-compute facility exceeding 100 kW per rack with direct-to-chip or immersion cooling, distinct from hyperscale general-compute facilities.

NVIDIA, with SPAN and PulteGroup (XFRA), began placing industrial server nodes — 16 Blackwell GPUs, 4 EPYC CPUs, and 3 TB of memory per node — in residential backyards to bypass a 2,600 GW utility-interconnect backlog, described as the first public residential-datacenter deployment by a major frontier-compute vendor (Source: sebastianbarros.substack.com).

Community opposition and political backlash

Opposition to data centers escalated through 2026 from local objections to organized action across multiple jurisdictions and into incidents of violence. AI-Driven Political Violence and Beyond Misuse: Artificial Intelligence, Grievance, and the Future Landscape of Political Violence — Yannick Veilleux-Lepage (Combating Terrorism Center at West Point, April 2026) provide the analytical framework. In Indianapolis on April 6, 2026, City Councilman Ron Gibson, who had supported a data-center project, had 13 rounds fired into his home overnight with a "No Data Centers" note left under his doormat; the council subsequently passed a measure allowing officials to keep home addresses private. Soufan Center reporting found that the most common online threats now involve "physical sabotage of proposed or operational data centers" (cited in The AI Backlash Could Get Very Ugly — Lila Shroff (The Atlantic, May 13 2026)). In Q1 2026, a record number of proposed data-center projects were canceled following local pushback (Heatmap News).

A series of mid-May 2026 actions extended opposition beyond local-NIMBY contests. Cleveland City Council rejected a data-center permit application on May 14, 2026, and Hill County, Texas voted the same week for a one-year data-center moratorium. Senator Elizabeth Warren opened an investigation into data-center companies. The March 2026 AOC–Sanders data-center-moratorium bill circulated again after Y Combinator's Garry Tan posted criticism of it. An Atlantic "Galaxy Brain" podcast episode (May 29, 2026) with Charlie Warzel and Heatmap reporter Jael Holzman called AI data centers "the most polarizing buildings in America" and described a left-right coalition coalescing around environmental, economic, and grid-capacity objections that, in the episode's argument, unites environmental-left and cost-of-living-populist-right constituencies (Source: theatlantic.com). TechCrunch reported (May 13, 2026) that xAI's Mississippi data center was running nearly 50 gas turbines unchecked (Source: tbpn.substack.com; techcrunch.com).

A wholesale capacity-reallocation case emerged around Lake Tahoe (Fortune, May 12, 2026; via Futurism May 14). Nevada-based NV Energy notified residents of Lake Tahoe — roughly 49,000 residents on both the California and Nevada sides — that it would stop supplying power after May 2027, leaving California-based Liberty Utilities, which gets 75% of its power from NV Energy, to seek a new supplier. The disconnection is explicitly driven by AI-data-center demand on Nevada's grid: an estimated 22% of Nevada's total electricity generation capacity went to data centers in 2024 (Desert Research Institute), a share that could rise to 35% by 2030. Reconnecting Liberty Utilities to a new California-side energy provider would cost "hundreds of millions of dollars," per Liberty president Eric Schwarzrock. Danielle Hughes, a California Energy Commission Efficiency Division supervisor quoted as a resident, said, "It's like we don't exist" (Source: futurism.com; original Fortune reporting: fortune.com).

Industry groups have begun measuring whether that opposition transfers to AI generally. The Information Technology and Innovation Foundation released polling associating public opposition to data centers with negative sentiment toward artificial intelligence, in an account published August 27, 2026. Opening an August 26, 2026 ITIF event on the results, president Daniel Castro said there is a "risk that opposition to data centers becomes, in effect, opposition to AI itself." The account states that the polling suggests the inverse is true; the article body is paywalled and the finding itself was not retrievable, so the direction and magnitude of the measured association are not established here (Source: insideaipolicy.com). The underlying poll was not located.

A national survey by the AP-NORC Center for Public Affairs Research and the Energy Policy Institute at the University of Chicago, conducted July 6–24, 2026 among 3,424 adults and released September 16, measured concern about local impacts directly. Asked about data centers in the communities where they are built, 63% were extremely or very concerned about electricity prices, 57% about water supply and 53% about power outages; 73% supported requiring companies that build data centers to pay for the grid upgrades needed to support them, 65% supported requiring clean energy sources, and 62% supported limiting the number of new data centers. Fewer rated data centers as extremely or very beneficial for technological advancement (24%) or job creation (20%). Counting those somewhat concerned, the release put concern about electricity prices at 84%, and the survey's factsheets put support for the grid-upgrade requirement at 79% of Democrats and 76% of Republicans (AP-NORC/EPIC 2026 Energy Survey: AI, Data Centers, Climate and Energy; EPIC's release: epic.uchicago.edu).

A Pew Research Center survey of 10,548 U.S. adults conducted July 20 to August 9, 2026 and published September 22 tracked change since January 2026: the share saying data centers are mostly bad for the environment rose from 39% to 54%, for home energy costs from 38% to 50%, for people's quality of life nearby from 30% to 49%, for local jobs from 15% to 24% and for local tax revenue from 12% to 21%, while 60% said they would be not too or not at all comfortable with a new data center operating in their area (Americans' Views of Data Centers Have Turned More Negative (Pew Research Center, September 2026)).

State-level policy responses

State governments moved from project-by-project siting fights toward statewide policy in mid-2026. Maine's proposed moratorium (April 2026) would have been the country's first statewide moratorium, an 18-month pause on facilities exceeding 20 MW for study; Gov. Janet Mills vetoed it on April 24, citing 125 to 150 jobs at the proposed Jay (Androscoggin Mill) facility (Data centers are coming for rural America — Abigail Bassett (The Verge, May 13 2026)).

The New York state legislature passed a bill establishing a one-year moratorium on new data centers on June 4, 2026, alongside the Bores-sponsored Bioterrorism Prevention Act on gene-synthesis screening (see AI Biosecurity) (Source: insideclimatenews.org). As of the June 5–6 update, the bill awaited Gov. Kathy Hochul, who has until December to sign or veto; it covers centers with peak demand of 20 MW or more, mandates developer-funded public hearings, and was narrowed from an earlier three-year version. The New York Independent System Operator (NYISO) is reviewing 24 data-center proposals totaling over 9,000 MW (Source: theverge.com; see Data Center Siting / AI Power Politics).

On July 14, 2026, New York became the first US state to impose a moratorium on large new data centers: Hochul ordered a one-year halt on facilities drawing 50 MW or more, directed preparation of a Generic Environmental Impact Statement, and said she would pursue repeal of the state's data-center sales tax exemption (Source: reuters.com). See New York Data-Center Moratorium (2026).

Illinois Gov. JB Pritzker announced on June 5, 2026 that the state will pause processing of data-center tax-incentive applications beginning July 1, after lawmakers declined to act on a proposed two-year suspension; Pritzker called on the legislature and stakeholders to build a framework addressing affordability, water, and grid impacts in the fall veto session (Source: nbcnews.com; see Data Center Siting / AI Power Politics, State-Level AI Regulation).

Sales-tax exemptions

By August 2026 the incentives themselves had become the contested instrument, in states that had previously competed to offer them. Reporting on August 2, 2026 described governors and legislatures moving to repeal data-center sales-tax breaks, with The Information calculating that repeal could raise the price of every gigawatt of AI computing by several billion dollars and add 7% or more to equipment costs (Source: theinformation.com).

Texas supplies the fullest public accounting. At a Senate Finance Committee interim hearing on July 27, 2026, chief revenue estimator Brad Reynolds testified that 69 data centers were approved for the state exemption between 2021 and 2025, that 59 more were certified in 2026 alone with roughly ten further approvals expected per month, and that of the 138 projects then qualifying, 20 had been audited and six found out of compliance. The comptroller's office estimated in 2025 that the state will forgo about $3.2 billion in sales-tax revenue over two years. Sen. Joan Huffman (R-Houston) said she has committed to filing legislation on the exemption in the session beginning January 2027 (Source: communityimpact.com).

Constraints on buildout

Multiple physical and political axes constrain data-center construction:

• GPUs (Nvidia/AMD compute): Binding for some labs (OpenAI Stargate stalls, Nvidia $100B "on ice"). Less binding for Anthropic, which assembled a multi-source SpaceX, Google, Amazon, and Akamai stack.
• Power generation: Binding. NEPA permitting reforms appear in America's AI Action Plan Pillar II; nuclear, geothermal, and fusion bets are in play.
• Power transmission: Increasingly binding. Interconnection queues stretch years and are a major bottleneck.
• Water (cooling): Binding in arid regions, with siting fights in Texas, Arizona, and Nevada.
• Land: A lower constraint; siting politics is often the operative proxy.
• Permitting / siting politics: Increasingly binding, with state and local opposition to large-load data centers.

Data centers as a cybersecurity target

In a prepared statement dated September 2, 2026 to the House Permanent Select Committee on Intelligence, Hudson Institute senior fellow David Feith argued that data centers create a grid-security problem new in kind rather than degree, because a campus drawing 300 MW to a gigawatt or more is densely computerized and can change its electricity consumption almost instantaneously, unlike traditional industrial load (A New Counterintelligence Map for the AI Age (Feith, HPSCI, 2026-09-02)). Trade coverage of the hearing, held September 3, reported the testimony as a call to treat data centers as a major cybersecurity target for adversaries including China (Source: insideaipolicy.com); data centers are one of five ecosystems the statement addresses.

The statement's anchoring incident is a routine transmission fault in Northern Virginia on July 22, 2026, which Feith says caused roughly 3,800 MW of data-center load to transfer rapidly from the grid to backup generation, producing voltage and frequency swings across the eastern United States from Virginia to Chicago. He states that operators stabilized the system, there was no blackout, and it was not a cyberattack, but that the event demonstrated the physics an attacker would study: several gigawatts of digitally controlled demand can move within seconds across a wide geography. The inferred threat is an adversary with synchronized access to enough inverters, batteries, uninterruptible power systems or other energy controls to manipulate large blocks of demand.

That framing makes hardware provenance inside the facility a security question rather than only a procurement one. Feith presents cryptocurrency mines as an existing instance — he states that almost all US crypto mines rely on Chinese-made compute and equipment, and cites the 2024 CFIUS-ordered divestment of a PRC-owned mining operation within a mile of Wyoming's Warren Air Force Base, along with subsequent Senate Intelligence Committee warnings about remotely controllable Chinese-made mining machines. Two supporting claims are attributed rather than asserted: that the Chinese inverter manufacturer Deye demonstrated in 2024 it could remotely disable installed equipment abroad during a commercial dispute, and that the research firm Strider has identified more than 2,700 Chinese research papers studying grid vulnerabilities, including hundreds it assesses as highly relevant to potential US grid attacks. He uses Foxconn — Taiwanese, central to AI-server and rack manufacturing, with a longstanding Greater China footprint — to argue that a company's flag alone cannot answer the supply-chain-security question (A New Counterintelligence Map for the AI Age (Feith, HPSCI, 2026-09-02)).

The statement is explicit that this is not an argument against building data centers, which Feith says America needs to build faster and at extraordinary scale; the asked-for measures are visibility into the installed base, trusted-supplier and firmware standards, elimination of adversary remote access, network segmentation, and grid rules designed for gigawatt-scale digital loads, with large data centers treated as critical grid infrastructure. He credits America's AI Action Plan with the principle that the AI stack and its supporting energy and telecommunications infrastructure should be free of foreign-adversary technology, and names the FCC's action on foreign power inverters and an executive order restricting foreign-adversary hardware in the bulk-power system as starts.

The framing places data-center security alongside the physical-siting and grid questions treated above, and connects to the compute-security concerns in Compute Governance and to the agent-security legislation introduced the same day (Stop Rogue AI Act (2026)).

Policy levers and tensions

The principal policy levers include NEPA streamlining (America's AI Action Plan Pillar II) on the federal-permitting side; state and local siting authority as a counterweight to federal streamlining; the CHIPS Act and Stargate as a capex-subsidization layer; hyperscaler procurement decisions, where data centers get built tracking where compute commitments are made; and state-level data-center incentives, with Texas, Arizona, Virginia, Iowa, and Oregon competing on tax abatement and power-access fast-tracking.

Three tensions recur. Datacenter-vs-residential power: residential prices rose more than 40% since 2020 with data-center load a contributing factor, and state public utility commissions increasingly face the question; the Lake Tahoe case above represents this as a wholesale capacity reallocation rather than only a price-pressure dynamic. Datacenter-vs-environment: water usage in arid regions, transmission-line politics, and cooling-tower thermal pollution. Datacenter-as-bubble-marker: the CoreWeave active-versus-backlog ratio is a reference datapoint in AI Bubble Debate.

On water specifically, an ITIF report by Robin Gaster argues the problem is one of governance rather than engineering or scarcity: "technology exists, and policy instruments are available… What's missing is institutional coordination, regulatory specificity, and a set of standardized mechanisms and metrics" (The Data Center Water Problem Is Soluble (Robin Gaster, ITIF, July 2026)). Two quantitative claims underpin it: indirect consumption through electricity generation exceeds direct cooling use by more than tenfold, while data centers overall "account for a very small fraction of total U.S. consumption"; and near-zero-water cooling is now available — "a bit more expensive, but close-to-zero water consumption is possible," with some hyperscalers already adopting zero-water designs. The implication is that public attention concentrates on the smaller share, which also has an available technical fix, while the larger share is a function of the electricity mix.

Relationships

• supports: Compute Governance, AI Bubble Debate, Procurement-Driven AI Governance.
• supports: AI-Driven Political Violence — physical infrastructure as target class
• related: A New Counterintelligence Map for the AI Age (Feith, HPSCI, 2026-09-02) — data-center hardware provenance and load-shift as grid-security risks
• supports: Data Center Siting / AI Power Politics — siting-specific dynamics
• supports: AI Environmental Impact
• related: America's AI Action Plan (Pillar II — infrastructure).
• related: Ratepayer Protection Act (H.R. 9340), Ratepayer Protection Act, H.R. 9340 — text as reported in the House (September 10, 2026) — federal large-load cost-allocation standard passed by the House September 16, 2026
• related: AI Energy Management Alliance (AEMA), Emerald AI — demand-flexibility advocacy for faster interconnection
• related: Americans' Views of Data Centers Have Turned More Negative (Pew Research Center, September 2026) — Pew's January–August 2026 tracking of perceived data-center effects.
• related: AP-NORC/EPIC 2026 Energy Survey: AI, Data Centers, Climate and Energy — national survey of concern about local data-center impacts and support for developer-pays grid requirements
• related: Zerra DC — developer of Anthropic's first Australian leased campus
• related: Akamai Technologies, Akamai Technologies Form 8-K: Anthropic Project Plans and Warrant Agreement (September 2026) — non-hyperscaler cloud capacity contracted by Anthropic; the 8-K discloses commitment, capex and memory pre-purchase figures
• related: Financing the AI Buildout (Van Nieuwerburgh, BPEA Fall 2026) — BPEA campus-cost model, Cleanview capacity scenario, financing risks; Stijn Van Nieuwerburgh
• related: CoreWeave, Anthropic (compute Snapshot block), Google DeepMind, Amazon, Microsoft.
• depends-on: Inside the Dirty, Dystopian World of AI Data Centers — Matteo Wong (The Atlantic, April 2026) — primary numerical anchor (energy, water, emissions, transmission)
• depends-on: Data centers are coming for rural America — Abigail Bassett (The Verge, May 13 2026) — rural-targeting + Maine case + Hicks net-zero-jobs finding
• depends-on: Data Centers in Space: A Pipe Dream, or AI's Next Big Thing? — Micah Maidenberg, Merrill Sherman (WSJ, May 12 2026) — orbital-data-center engineering and players
• depends-on: WSJ, "AI Is Using So Much Energy That Computing Firepower Is Running Out" (folded supporting source, Source: wsj.com) — compute crunch + token-pricing dynamics

Sources

Stub created 2026-05-11; substantially expanded 2026-05-14 with Inside the Dirty, Dystopian World of AI Data Centers — Matteo Wong (The Atlantic, April 2026), Data centers are coming for rural America — Abigail Bassett (The Verge, May 13 2026), Data Centers in Space: A Pipe Dream, or AI's Next Big Thing? — Micah Maidenberg, Merrill Sherman (WSJ, May 12 2026) foundational source ingest.