There are 4,011 data centers in the United States as of 2026 — nearly double four years ago. Before you can shape how they're built, governed, or held accountable, you need to understand what they actually are and what they demand from your community.
A data center is a building — sometimes a campus of buildings — filled with servers, cooling systems, and power infrastructure. Every AI query, cloud file, and streamed video runs through one. There are 4,011 of them in the U.S. as of 2026. Before you can shape what gets built in your community, you need to know what you're dealing with.
The term "cloud" is deliberately abstract. There is no cloud. There are buildings — in your state, in specific zip codes, drawing real water from real reservoirs, generating real noise in real neighborhoods. The abstraction serves the industry. Specificity serves you.
What's changed with AI is scale and intensity. A traditional server rack draws 5–15 kilowatts of power. An AI-optimized rack running NVIDIA's latest GPUs draws 120–140 kilowatts — the equivalent of about 30 residential furnaces in a single rack. NVIDIA's roadmap targets 600 kW per rack by late 2026. This isn't a marginal increase. It's a fundamental shift in what these buildings demand from the grid and the water system.
"There is no cloud. There are buildings — in specific places, drawing specific amounts of power and water, in someone's community."
When a data center developer presents to your community, they will have polished projections. Apply the same standard to those projections that you'd apply to any other claim: What's the source? Who benefits from me believing this? What's missing from this picture? What does the contract actually say versus what's being promised verbally?
Much of the public confusion around data centers is deliberate. "Hyperscale" sounds impressive but just means very large. "Carbon neutral" and "net zero" sound equivalent but mean very different things. "Water positive" sounds like the company is adding water — it means they're buying credits in water restoration projects elsewhere, which may be geographically and temporally disconnected from where they're actually consuming water.
The distinction between water withdrawal (taken from source), water consumption (lost, primarily through evaporation), and water discharge (returned) matters enormously. Companies typically report the metric that makes them look best. Always ask which definition is being used.
Every major infrastructure buildout in American history has produced similar patterns: a private interest with capital, a public resource at stake, speed pressure that advantages the builder over the community, and a gap between what was promised and what was delivered.
The transcontinental railroad brought economic connection — and land seizure, labor exploitation, and the displacement of communities that had no seat at the negotiating table. Interstate highway construction in the 1950s and 60s deliberately routed through Black neighborhoods in dozens of American cities, destroying communities that took generations to rebuild. Both were presented as national necessities. Both were built faster than communities could respond.
This is not an argument against infrastructure. It is an argument for going in with eyes open. The communities that fared best in every major infrastructure era were the ones that understood what was being built, organized before the contracts were signed, and negotiated from knowledge rather than desperation.
"The communities that fared best were the ones that organized before the contracts were signed."
The data center buildout is happening faster than most infrastructure eras. A data center can go from proposal to construction in two years. A transmission line takes ten. A community's institutional knowledge about what to demand takes time to build — which is exactly why the speed is a feature, not a bug, for developers who would prefer communities not have time to organize.
The European Union required data centers above 500 kW to monitor and report energy performance from September 2024. Germany mandates waste heat reuse with financial penalties for non-compliance. Ireland ended its moratorium in December 2025 but with strict conditions: 80% renewable energy, on-site backup, grid operator veto authority. Singapore awards data center capacity competitively — only 380 MW total in 2023–2024 — requiring PUE of 1.3 or better.
The U.S. has no equivalent federal framework. That gap is a choice — and choices can be changed.
From roughly 2005 to 2020, U.S. electricity consumption grew at just 0.1% annually — driven by LED adoption, efficiency standards, and the 2008 recession's lasting effects. This 15-year plateau bred complacency. Utilities systematically under-forecasted demand and underbuilt transmission. When AI-driven demand materialized suddenly around 2023, it hit infrastructure designed for steady-state load. The national five-year load forecast in 2024 was five times higher than 2022 predictions. That is how fast the collision happened.
The data center industry is undergoing a genuine engineering revolution — driven not by altruism but by physics. As AI racks approach 120–600 kW per rack, air cooling physically cannot remove heat fast enough. The industry is being forced to innovate.
Closed-loop liquid systems circulate coolant directly to chips — eliminating evaporative water loss entirely. Microsoft committed to zero-water cooling for all new builds from August 2024. Meta's newest facilities use closed-loop systems consuming effectively zero operational water. These systems can deliver coolant at 55–60°C — high enough to supply building heating systems directly.
The critical insight: the innovation pressure AI creates is generating engineering solutions that could benefit everyone — but only if those solutions are mandated, not left voluntary. When a company adopts zero-water cooling because it's profitable, that's good. When it's required as a condition of building, that's transformative.
Every efficiency gain in computing history has been consumed by more computation, not less total energy. When data centers become 30% more energy-efficient, the industry builds 30% more data centers. This is not speculation — it's the documented pattern since 1965. Efficiency gains are necessary but not sufficient. They must be paired with demand accountability and cost equity to produce public benefit.
AI infrastructure is splitting into two types. Training — building the model — requires massive, centralized GPU clusters running for weeks or months. Inference — using the model — can run on smaller hardware, closer to users, with more flexibility. By 2026, inference represents about two-thirds of all AI compute. This shift matters for communities because inference facilities are smaller, more distributed, and can be sited closer to population centers — which means more communities will face these decisions, not just the handful of major data center markets.
The most powerful moment in a data center's lifecycle — from a community's perspective — is before the zoning approval. Once a facility is permitted, the negotiating leverage shifts dramatically toward the developer. The strategy is to engage early, with specific knowledge, in the specific forums where decisions are made.
Data center developers negotiate these agreements professionally, repeatedly, with legal teams and lobbyists. Most local officials and community members do it once. The way to close that gap: connect with communities that have already done it. The NAACP's Stop Dirty Data Centers initiative has community benefit agreement templates. The Coalition for Responsible Data Center Development has a free Resistance 101 toolkit. Brookings Institution's data center research provides independent analysis you can cite. You don't have to build the knowledge from scratch.
Data centers use public resources — power grids built with ratepayer investment, water from public reservoirs and aquifers, roads and land in communities that had no vote on whether to become data center hosts. The ethical question is not whether private companies can use public resources. It's what they owe in return — and what happens when they don't pay it.
The current system in most U.S. jurisdictions socializes costs and privatizes benefits. Data centers pay rates as low as 5.5¢/kWh in Virginia while residents pay 9–14¢. Ninety-five percent of data center interconnection costs in PJM have been rolled into general transmission charges paid by all ratepayers — including the 21 million households that were already behind on utility bills before the AI buildout accelerated prices further.
This is not an accident. It is the result of specific policy choices — NDAs with local officials, lobbying against transparency legislation, regulatory structures that haven't been updated to account for a new category of industrial load. Naming it as a policy choice is important, because policy choices can be changed.
The last question is genuinely hard. Opposition that protects one community by relocating harm to a less-resourced community is not the same as solving the problem. The ethical answer is not just "not in my backyard" — it's advocating for the standards that should apply everywhere, not just where you live.
"Opposition that protects one community by relocating harm to another is not the same as solving the problem."
A Virginia circuit court ruled in November 2025 that water usage data is public information — not a trade secret. Judge Ciaffone wrote: "There are few resources more precious than water." That ruling should be the national standard. The argument that disclosing how much public water a facility uses would reveal competitive secrets does not hold up to scrutiny — and courts are beginning to agree.
81% of Virginia localities hosting data centers have signed NDAs with operators. Those agreements prohibit sharing "business plans" and "non-public information," instruct officials to "share as little as legally possible" if FOIA'd, and give companies advance notice to intervene legally. In Pima County, Arizona, a county supervisor learned he was bound by an NDA regarding Amazon's $3.6 billion proposal only when a developer's spokesperson accused him of violating it by speaking to a newspaper. This is not a narrow protection for proprietary technology. It is a systematic suppression of public information about how public resources are being used.
This section applies the AI Thinking Model™ — a framework for critical thinking, wisdom, innovation, strategy, and ethics developed by Liz B. Baker, Global Institute for AI & Humanity. Learn more →
