Local Opposition to Data Centers: What It Means for AI

One week, two signals: a $17 billion campus and a wall of "no" votes

On November 5, Google announced a $17 billion investment in its first Wyoming data center, a campus in Laramie County built with four energy and infrastructure partners. The same week, US voters rejected 40 of 41 data center-related ballot measures, and candidates who campaigned against expansion won key races in Texas and Georgia. Capital is pouring in at record speed while local opposition to data centers hardens into something that shows up on ballots.

If your company runs on AI tools, or plans to, you are standing downwind of this collision. Somewhere between the checks being written and the permits being refused sits the price, speed and reliability of every AI service you buy. A smaller company's own summer crisis shows exactly how that works, and the numbers are more instructive than any forecast.

Most executives read headlines about data centers as an energy-sector story or a tech-giant story. It's neither. It is a supply story, and supply decides what you pay for AI next year.

Every model your team calls, every agent that drafts a contract or routes a purchase request, ends up as a workload on a GPU in a building somewhere. Buildings need land, electricity, water and, increasingly, the consent of the people who live next door. That last input is the one nobody budgeted for.

Where the money is going: Google's campus and fal's compute

Google: $17 billion and a long list of promises

The Wyoming project is the clearest picture of how hyperscalers now behave. According to reports, Google's partners are Tallgrass Energy, Black Hills Energy, American Electric Power and the developer Related Digital. The campus is described as a 702-megawatt facility, though one analysis notes that figure refers to the fuel-cell power plant rather than a disclosed measure of computing capacity. Google also says the design is air-cooled to limit pressure on local water, and it pledged to fund local water replenishment programs.

The commitments go further. Google said it would pay for all the power it uses and cover infrastructure needs created by its growth. Earlier this autumn it also outlined community programs tied to data centers, full property taxes on new campuses, earlier public disclosure of its role in projects, and an end to nondisclosure agreements with state and local governments (with narrow exceptions).

Read that list again. It is the shopping list of a company that has learned the hard way that a signed contract with a utility is not the same as permission from a town.

fal: a smaller company buying the same scarce thing

On November 3, fal, a San Francisco inference startup, announced a $1.3 billion Series F led by Bessemer Venture Partners, with Nvidia among the participants. According to Bloomberg, the round values the company at $16 billion, double its April valuation. Fal hosts image, video and audio models behind a single API, so businesses can add generative features without running their own servers.

Fal isn't building a Wyoming campus. It is buying and leasing capacity, and the reason it raised so much is telling. Its CEO said that demand grew so fast this summer that the company could not serve all its customers and had to throttle them. If you want to understand how money moves between chipmakers and the companies that rent out their silicon, our explainer on Nvidia's circular investments covers the mechanics.

Google is spending tens of billions to build capacity. A company like fal is spending hundreds of millions to rent it. Both are racing the same clock.

How local opposition to data centers is reshaping the AI supply chain

The scale: billions in projects stuck at the permit stage

Data Center Watch, a research project of the firm 10a Labs, has been tracking this since 2025. Its numbers have moved fast:

• May 2024 – March 2025: in its first report, the group estimated that local opposition had blocked or delayed $64 billion in data center projects.
• Q1 2026: at least 75 projects worth about $130 billion blocked or delayed, per the group's data.
• Q2 2026: at least 45 projects worth about $68 billion blocked or delayed, according to a September Bloomberg report. The group counted 843 opposition groups across 49 states.

Two caveats matter. The researchers themselves note that organized resistance can't explain every delay, because permitting backlogs and regulatory changes play a role too. And the Q2 2026 figure is lower than Q1, so the pace of new stoppages isn't rising in a straight line. What has changed is the shape: the resistance is spreading into town councils, state legislatures and referendums. About 30 statehouses have introduced or adopted rules on siting, power use or water consumption, and some towns now pass moratoriums before a developer has even filed.

Public polling points the same way, with recent coverage citing majorities of Americans who oppose a data center being built in their own area.

What residents actually object to

Strip away the politics and the complaints are practical: electricity bills, water use, noise, and a suspicion that costs get shared while profits don't. In Pine Island, Minnesota, a local survey in August found about three in four respondents against Google's proposed 780-megawatt project, even though the company planned to back 1.4 GW of wind, 200 MW of solar and 300 MW of battery storage alongside it.

In Wyoming, a University of Wyoming poll put support for expanding AI infrastructure at only 30% of residents, and Google had already gathered input from more than 200 people in Laramie County. Residents there want straight answers on jobs, tax contributions, school funding and what several developments around Cheyenne add up to.

An analysis of 46 projects across 20 states, worth about $170 billion and publicly delayed or cancelled after community pushback between 2024 and May 2026, found that secrecy was the most common trigger. Projects wrapped in NDAs, with unclear ownership or undisclosed end users, drew faster and sharper organizing than those that named their tenants from day one.

That finding explains Google's new disclosure policy better than any press release does.

The case: what a compute shortage looked like for fal's customers

Abstract risk is easy to shrug off. Here is the concrete version, using only what has been reported.

The situation. Fal reached about $800 million in annualized revenue, a figure Bloomberg attributed to a person familiar with the matter. Its customers reportedly include Adobe, Shopify and Canva. Over the summer, request volume outran the GPU capacity fal had lined up, and the company had to rate-limit customers.

The cost of the squeeze. Throttling is a quiet failure. Nothing crashes. Image generation just slows, queues form, and a product feature that worked in spring becomes unreliable in July. For a retailer using generated product imagery, or a marketing team producing video at volume, that means missed launch windows rather than a clear outage to escalate.

The response. Fal raised $1.3 billion, its second large round in about seven months, after $350 million in April at an $8 billion valuation. The money funds more compute and growth of the team to 185 people. Even so, TechCrunch reported, citing people familiar with inference providers' finances, that gross margins across the sector are thin because compute is so expensive to buy or lease. Fireworks AI reportedly raised $1.5 billion in July at a $17.5 billion valuation, and Baseten and Modal are raising too.

The lesson in numbers. A provider with about $800 million in annualized revenue, a $16 billion valuation and Nvidia on its cap table still ran short of capacity for part of a summer. Thin margins mean that when capacity tightens, there is little cushion to absorb it. Either prices rise, or limits appear. As local opposition to data centers slows new supply, that cushion gets thinner still.

Capacity is the new lead time. Nobody notices it until a shipment doesn't arrive.

What business leaders should do about AI capacity risk

You can't change a county zoning vote in a state where you don't operate. You can decide how exposed your operations are to the outcome. Five moves are worth making this quarter.

Map which processes depend on which provider

List every workflow that calls an external model: invoice matching, contract review, customer replies, content production. Next to each, write the provider, the region it runs in and what happens if throughput halves for two weeks. Most companies discover that two or three critical processes sit on a single API with no fallback.

Write capacity terms into contracts

Ask vendors directly: what are your rate limits at peak, how do you notify us of throttling, and what do we get when you can't deliver? Fixed-price multi-year deals look attractive until a provider with thin margins hits a shortage. Our checklist for choosing an AI agent vendor includes the questions that separate a reliable supplier from a lucky one.

Build in a second route

Architect critical flows so the model behind them can be swapped. Route routine, high-volume work to smaller, cheaper models and reserve the expensive frontier ones for decisions that justify them. This is as much about spend as about resilience: a lighter model for 80% of requests cuts exposure to both price hikes and queues.

Put governance around agents before you scale them

More agents mean more inference calls and more places where a throttled service can stall an approval chain. Central oversight, usage budgets and clear ownership keep that from turning into chaos. We described the operating model in managing AI agents in the enterprise.

Track the siting news that touches your vendors

Ask your main providers where their capacity is being built and whether those projects face contested approvals. It's a ten-minute conversation. A vendor with a Wyoming-style campus backed by pre-signed power and a public community plan is a different risk from one whose growth plan depends on a rezoning that three hundred neighbors are fighting.

Do these five things and something shifts in how the job feels. Instead of waking up to a vendor outage and improvising, you know which process depends on what, what the fallback is, and who owns it. That's what control looks like in practice: calm, because the surprises have already been priced in.

It also changes how the room sees you. A board or an investor who has watched competitors get caught by throttled APIs and sudden price changes will read a documented capacity plan as evidence of a leader who sees supply chains coming, not one who reacts to them.

FAQ

Is local opposition to data centers actually slowing AI growth? It is delaying specific projects, not stopping investment overall. Data Center Watch counted about $68 billion in projects blocked or delayed in Q2 2026 alone, while companies like Google keep committing tens of billions elsewhere. The practical effect is that capacity arrives later and in different places than planned.

Why would a company like fal need to raise $1.3 billion? Because inference, the work of running trained models for every user request, is capacity-hungry and expensive. Fal's CEO said demand outran its computing power over the summer, forcing it to limit customers. The raise is meant to fund more compute and its team.

Will AI services get more expensive because of this? It's a real risk rather than a certainty. Inference providers already report thin gross margins because of compute costs, so shortages are more likely to show up as higher prices or usage caps. Locking in capacity terms and keeping a second provider ready are the cheapest hedges.

What is Google doing to win over local communities? Google says it will pay for the power it uses, cover related infrastructure, pay full property taxes on new campuses, disclose its involvement earlier, and stop using NDAs with local governments. In Wyoming it also chose an air-cooled design and promised water replenishment programs.

Should a small or mid-sized business care about data center politics? Yes, indirectly. You don't build data centers, but you rent their output through every AI tool you use. Knowing which of your processes would stall if a vendor throttled service is basic operational hygiene, whatever your size.

Google's Wyoming campus and fal's emergency rounds are two answers to the same question: where will the compute come from, and who will let it be built? Neither answer is settled. Opposition is organized, local and increasingly effective, and the capital behind AI is not backing down.

Take the fal story as a mirror. Which of your processes would slow down if your main AI provider throttled you for a month? Is there a named owner, a fallback route, and a clause in the contract that covers it? If any of those answers is "I'm not sure," that is your starting point. Write the answers down, compare them with the case above, and decide how much of your operation you're willing to leave on a single supplier's good summer. If you'd like a second pair of eyes on that map, start a conversation in the AI chat.