Data center backlash: what it means for AI power and capacity

For most of the AI boom, the limits on new compute were chips, power equipment and grid connections. The US elections on 3 November 2026 added another one: local consent. According to Ballotpedia, there were 41 local ballot measures about data centers in 10 states, and in 40 of them the result went against new development. Ohio alone had 21 of the measures.

The reaction came quickly. On 5 November the Nasdaq Composite fell about 5.5% and Nvidia about 9.3%, with several reports linking part of the sell-off to worries about the data center buildout. The same day OpenAI published a "National Community Compact" for towns that host its sites, promising to pay the full cost of the electricity and grid upgrades it needs.

This article explains why data centers meet local resistance, how their power use compares with a town's, and what it means for anyone who plans AI compute. It reports positions and figures and takes no side on parties, candidates or ballot questions. For context, see the MLOps and AI infrastructure topic page.

What happened

Voters back limits in 40 of 41 local data center measures (3 November 2026)

Ballotpedia lists 41 data center-related local measures on the 3 November ballot, across California, Florida, Illinois, Kansas, Maine, North Dakota, New Jersey, Nevada, Ohio and Wisconsin. In 27 cases voters approved a measure that limited development or expressed opposition. In 13 they defeated a measure that would have supported development. The one exception was Harwood, North Dakota, where voters rejected an extra review requirement for industrial projects, data centers included. Other outlets' totals differ slightly.

The measures were not all the same kind of rule:

Type of measure	Example	What it does	Result
Outright ban	Alhambra, California (Measure DB)	Prohibits data centers in the city	Approved, 83% yes
Size cap	Conneaut, Ohio (Issue 10)	Prohibits data centers above 25 MW	Approved, 65% yes
Moratorium	Wells, Maine (Article 5)	180-day pause on data center development	Approved, 82% yes
Conditions on tax deals	Perry, Ohio (Issue 16)	Requires financial guarantees from hyperscale developers seeking certain tax incentives	Approved, 87% yes
Rezoning	Wilmington, Ohio	Would have rezoned about 119 acres from rural residential to light industrial	Defeated, 20% yes

The differences matter. A ban stops a project, a moratorium pauses one, a size cap allows small sites but blocks AI-scale ones, and a rule on tax deals changes a project's price rather than forbidding it.

Exit poll puts AI and data centers among voters' concerns (3 November 2026)

The NBC News exit poll, with data from SSRS, found that 56% of voters called AI and data centers a crucial or major factor in their vote for Congress. Other issues scored higher on the same question: government corruption at 90% and the cost of living at 89%. AI and data centers mattered to many voters, but usually alongside other concerns.

Earlier polling points the same way. Gallup, with data from March 2026, found that 71% of Americans opposed an AI data center in their area, 48% strongly, against 53% who objected to a nearby nuclear plant. Strong opposition came from 56% of Democrats and 39% of Republicans, so the concern crosses party lines. Opponents cited water and energy use, pollution and utility costs. Supporters cited jobs and tax revenue.

AI stocks fall (5 November 2026)

On Thursday 5 November the Nasdaq Composite fell 5.52% and Nvidia 9.26%, according to market data compiled by HeyGoTrade. Other chip stocks also fell sharply. Commentators gave several reasons at once: post-election uncertainty, possible zoning, environmental and power delays for data centers, and cautious forecasts from Qualcomm and Arm. One day of trading says little about long-term demand.

OpenAI publishes a National Community Compact (5 November 2026)

On 5 November OpenAI released a National Community Compact and a Compute Campus Policy Blueprint. As reported by SiliconANGLE and The New York Times (republished by GV Wire), the compact covers future sites OpenAI owns and operates, which it calls "compute campuses". The company says it will:

pay the full cost of the electricity it uses and the grid upgrades needed to serve it;
post financial guarantees so local households and businesses do not pay for upgrades if a project is cancelled or delayed;
reduce its power draw during periods of high grid demand;
publish water and power use for each site, plus construction plans and noise projections;
stop requiring public officials to sign non-disclosure agreements, apart from narrow security and trade-secret protections;
publish expected jobs, a local agreement for each site and an annual report.

The blueprint asks lawmakers to apply similar rules across the industry, and asks regulators to make large users pay the grid costs they add. The NYT report notes two limits: the pledges cover only sites OpenAI owns, while partners such as Oracle build and own much of its capacity, and they extend commitments first made in January. That January plan said: "we commit to paying our own way on energy, so that our operations don't increase your electricity prices." SiliconANGLE reported that Google made a similar electricity pledge the same day for a planned $17 billion data center in Wyoming, and that Meta and Anthropic had made comparable commitments earlier in the year.

Analysts estimate how much capacity is at stake (4 to 6 November 2026)

GovTech, among others, reported a Goldman Sachs estimate that around 45% of planned US data center expansion could be affected. The New York Times reported that winning candidates for governor in Texas, Georgia and Ohio had pledged to pause new data center permits and tighten construction rules. The full Goldman note is not public, so read the 45% as a share of planned capacity exposed to possible restrictions, not a forecast of cancellations.

An earlier analysis gives a counterweight. In September, SemiAnalysis estimated that only about 2.3 GW of US capacity was actually delayed by local and state moratoriums, against a projected 38 GW of new US IT capacity in 2027, because it counted a moratorium only if it blocked an approval a specific project still needed. The figure predates the election and may now be higher.

Context: Microsoft and Chevron's gas deal (22 June 2026)

Some roundups listed this deal with this week's news, but it is older. On 22 June 2026, as reported by The Next Web, Chevron announced a 20-year deal to supply up to 2.67 GW from a new gas-fired plant near Pecos in West Texas, serving a Microsoft data center without drawing from the Texas grid. First power is planned for 2028.

Why data centers meet local opposition

A large data center is an unusual neighbour: a big construction project, then a quiet building with a very large, constant appetite for electricity. Objections tend to fall into six groups.

Electricity prices. Utilities recover the cost of new plants and lines from all customers through regulated rates. When a large user needs new supply, residents worry they will pay part of the bill. This is the concern the OpenAI and Google pledges address.

Grid upgrades. A data center may need new substations, transmission lines and transformers, which take years to build. If a project is cancelled halfway, those costs can be left behind. Hence the financial guarantees in OpenAI's compact and in Perry, Ohio's measure.

Water. Some cooling designs evaporate large amounts of water. In dry regions or towns with stressed supplies, this competes directly with homes and farms.

Noise and land. Cooling fans, backup generators and new power lines change the feel of a place. Many measures were really zoning questions: whether farmland or residential land becomes industrial.

Tax deals and secrecy. Incentives are often negotiated before the public knows who the developer is, sometimes under non-disclosure agreements. OpenAI's decision to stop requiring NDAs from officials responds to exactly this complaint.

Jobs. Construction employs many people for a few years. A finished data center is highly automated and usually employs far fewer. Supporters point to tax revenue and long-term jobs; opponents to the gap between the size of the site and the size of the payroll.

None of these concerns is about AI models as such. They are about land, water, wires and money.

How much power a data center uses, compared with a town

Power is measured in two ways, and mixing them up causes confusion.

Capacity (megawatts, gigawatts) is how much power something draws at once. It is like the width of a pipe.
Energy (megawatt-hours, terawatt-hours) is how much it uses over time. It is like the total water that flows through the pipe in a year.

Grid operators worry most about capacity, because wires, transformers and plants must be sized for the highest draw, not the yearly total.

The International Energy Agency's Energy and AI report put it plainly: a typical AI-focused data center consumes as much electricity as 100,000 households, and the largest ones under construction will consume 20 times as much. The IEA projected that global data center electricity use would more than double by 2030, to about 945 TWh.

Two features make a data center harder for a grid than a town of similar average use. A town's demand rises and falls through the day, while a data center runs close to full power around the clock. And a town grows a few hundred homes at a time, while a data center can request hundreds of megawatts in one application, faster than new transmission lines can be built.

This explains the 25 MW cap in Conneaut, Ohio. The 2.67 GW West Texas plan is about a hundred times that size. A cap at 25 MW leaves room for smaller facilities and blocks AI-scale campuses.

How companies respond

The industry's answers fall into five groups. Each has costs and limits.

Paying for the grid upgrades

The simplest answer is to make the newcomer pay: special tariffs for very large users, long minimum contracts and upfront guarantees, so that if a project shrinks other customers are not left with the bill. The hard part is separating costs caused by one customer from upgrades that benefit everyone is a long-running argument in utility rate cases.

Behind-the-meter generation

"Behind the meter" means the power plant sits on the customer's side of the utility meter. The data center makes its own electricity on site or next door, rather than buying it through the public grid. Think of a house with its own generator and no mains connection, scaled up to a power station. The Chevron and Microsoft project is the example: up to 2.67 GW supplied directly, without using the Texas grid.

The benefits are speed and separation. The project skips the queue for a grid connection, and residents' bills are not tied to its fuel costs. The drawbacks are real too. On-site gas plants have their own emissions and air permits, which bring their own local debates. An off-grid site needs its own backup, and the owner carries fuel price risk for decades. Behind-the-meter supply can ease the price worry, but not the land, water or noise worries.

Demand response and flexible data centers

Demand response means a large user agrees to cut consumption when the grid is under stress, usually on the hottest or coldest days, in return for payment or a faster connection. It is like a family agreeing to run the dishwasher after peak hours when the energy company asks.

A grid is built for its few busiest hours of the year, so most of the time it has spare room. Researchers at Duke University's Nicholas Institute estimated in February 2025 that the largest US grid regions could take about 76 GW of new load if those loads accepted curtailment of 0.25% of their annual energy, about 85 hours a year in their model. At 1% curtailment the figure rose to about 126 GW. These are modelled estimates, not guarantees, but they explain why "flexible" data centers have become a policy goal, and why OpenAI's compact includes cutting power draw at peak times.

Training jobs can checkpoint and wait, and batch inference can run later or elsewhere; live chat traffic cannot. A flexible data center is like a delivery company that can move some deliveries to night, but not the urgent ones.

Efficiency

Every watt saved in the chip, the cooling or the software shrinks the grid request. Each accelerator generation, such as the NVIDIA B200, aims to do more work per watt, though total site power usually rises because operators install more chips. In software, techniques such as knowledge distillation, quantisation, the KV cache and prompt caching cut the compute needed per request.

Siting

Finally, companies choose where to build. Spare grid capacity, a cool climate, non-potable cooling water, existing industrial zoning and a willing local government all reduce conflict. The 2026 results suggest that local politics now belongs on the siting checklist next to fibre routes and land prices. The earlier article on AI's infrastructure bottlenecks covers the chip and optics side of the same supply problem.

What it means in practice

Capacity becomes less certain. A project announced is not a project powered. Permits, local votes, grid studies and possible state pauses make delivery dates harder to predict. Estimates of the effect range from SemiAnalysis's pre-election 2.3 GW to Goldman's 45% of planned expansion exposed.

Costs move from the public to the project. Paying full grid costs, posting guarantees and building private power make each site more expensive. Some of that may reach cloud and AI service prices over time, by an unknown amount.

Transparency becomes a condition. NDAs, unpublished water use and late disclosure of the developer were common triggers. Expect more disclosure up front, voluntary or required.

Efficiency becomes a business issue. When power is scarce, a team that needs fewer GPU hours can grow while others wait.

For most users the effect is indirect: slower capacity growth in some regions, tighter rate limits at busy times or higher prices for the largest models.

What to do now

Treat capacity as a risk, not a given. If your plans depend on a provider adding GPUs in one region next year, ask what that depends on and keep a fallback.
Measure compute per task. Track GPU hours and cost per thousand requests or per training run, and set cost and latency budgets so growth is a decision, not a surprise.
Use the smallest model that meets the bar. Many extraction, classification and routing tasks run well on small language models. Send only hard requests to large models with LLM routing.
Cut repeated work. Cache prompts and responses, batch offline jobs and remove duplicate calls.
Separate urgent and flexible workloads. Mark which jobs can pause or move, such as training, evaluation and batch scoring, so you can use off-peak or flexible capacity.
Ask providers about power and water. For large contracts, ask where capacity sits, whether it is grid or behind-the-meter, how they handle demand response and what they publish on water.
If you build or host, start with the community. Disclose early, publish water and noise plans, and agree who pays for grid upgrades before any vote.
Find idle capacity. Reserved GPUs that sit idle waste power and money. Model monitoring and utilisation dashboards help find them.

Risks and open questions

The size of the effect is unclear. Many measures affected small towns rather than the largest planned campuses, and the most-cited estimate, Goldman's 45%, has not been published in full.

Pledges are voluntary and partial. OpenAI's compact covers only sites it owns, and many AI data centers are built by partners and colocation firms. How regulators turn "pay the full cost" into tariffs will decide whether bills actually stay flat.

Behind-the-meter power moves the debate. On-site gas avoids grid queues but raises questions about emissions, air permits and fuel costs. Some communities may object to the power plant as much as to the data center.

Flexibility has limits. The Duke estimates assume loads really curtail when asked. Live AI services cannot pause easily, and regulators will want proof.

Watch next: whether newly elected state officials pause permits and for how long, how utility regulators write large-load tariffs, whether other developers adopt compacts like OpenAI's, how many projects move behind the meter, and whether cloud providers change GPU prices or regional availability.

Frequently asked questions

What did voters actually decide on data centers?

According to Ballotpedia, there were 41 local measures in 10 states, and in 40 the outcome went against new development. They included bans, temporary moratoriums, size caps, conditions on tax incentives and rejected rezonings. Ohio had 21 of them.

Do data centers raise electricity bills?

They can, depending on how a utility and its regulator share the cost of new supply and grid upgrades. If those costs are spread across all customers, residents pay part of them. Special tariffs, upfront guarantees and on-site power are ways to prevent that.

What does behind-the-meter mean?

The power plant sits on the customer's side of the meter, so the data center generates its own electricity instead of buying it through the public grid. It avoids grid queues but brings its own emissions, permits and fuel risks.

What is demand response for a data center?

An agreement to cut power use when the grid is under stress, usually for a limited number of hours a year. It suits work that can pause or move, such as training and batch jobs, better than live services.

Will this make AI services more expensive or scarce?

Possibly, but the size and timing are uncertain. Higher site costs and slower approvals may reach cloud prices over time. Smaller models, caching and off-peak scheduling reduce exposure.

Summary

The 3 November 2026 elections made local consent a real constraint on AI infrastructure. Voters backed limits in 40 of 41 local data center measures, polls show cross-party unease about nearby sites, and OpenAI responded with a compact promising to pay full electricity and grid costs and publish water and power use. Estimates of the capacity at stake vary widely. The concerns are practical, about bills, grids, water, noise, tax deals and jobs, and so are the responses: paying for upgrades, behind-the-meter power, demand response, efficiency and better siting. For anyone planning AI compute, treat capacity as a risk, measure cost per task and use the smallest model that does the job.

Voters back data center limits in 40 of 41 local measures: what it means for AI capacity