Are AI data centers straining the power grid?

Short answer

Yes, the grid strain is real and growing, but most public alarms mix a measurable engineering trend with a separate, faster-moving political sentiment. Several independent surveys converge on a clear public-opinion signal — roughly six in ten U.S. adults are concerned about the electricity AI uses, and about seven in ten oppose a data center in their own community. The harder engineering question is how quickly demand is rising versus how fast transmission, generation, and cooling infrastructure can catch up; the answer there is genuinely unsettled.

Why people are concerned

The concern is anchored in concrete stories: data-center siting fights in Loudoun County, Virginia; the explosive water-use controversies around cooling in Arizona and Texas; and recurring utility filings about multi-year waits to interconnect new AI campuses with the grid. Those local conflicts travel upward into national surveys.

In a Reuters/Ipsos poll from August 2025, about 61% of U.S. adults said they were concerned about the electricity needed to power AI (the Reuters URL itself returns 401 outside its network, so the figure is cited here via the aggregation in Pew Research’s follow-up coverage). Stanford HAI’s 2026 AI Index Public Opinion chapter elevates the issue further: it reports that about 70% of Americans now oppose an AI data center in their own community — a Gallup March 2026 figure the Index restates — and lists the environmental footprint of AI infrastructure among the rising-concern clusters year over year.

It is not just an American worry. KPMG’s 2025 global study found that environmental and sustainability concerns cluster with privacy as a top-tier global AI risk, and that 78% of respondents globally want governments to act on AI’s environmental impact. Pew’s 25-country survey from October 2025 shows that the median trust in national governments to regulate AI is uneven — 37% in the United States, 53% in the European Union, and 27% in China — which makes “who decides what to do about data centers” as politically loaded as the energy question itself.

What is true

The engineering signal that data-center electricity demand is rising fast is well established. The Stanford HAI 2026 AI Index Technical chapter reports that U.S. data center electricity consumption is projected to reach roughly 6–12% of total U.S. electricity by 2028, depending on scenario assumptions about hyperscaler build-out, training-versus-inference mix, and grid additions. That range is wide because the underlying IEA and analyst projections disagree, but the central estimate is materially higher than the roughly 4% share data centers held in 2023.

Several other established facts round out the picture. KPMG’s longitudinal data show that worry about AI has risen from 49% (2022) to 62% (2024) across 17 countries it tracks, and trust fell in 13 of those 17. Pew’s April 2025 study found that 55% of the U.S. public and 57% of AI experts want more personal control over how AI is used, with environmental externalities part of what they mean by “control.”

What is also true, and often lost in the alarm: AI is a significant driver of data-center electricity growth, but not the only one. Cryptocurrency, general cloud growth, on-shoring of manufacturing, and electrification of buildings and transport are all pulling on the same wires in the same regions.

What is exaggerated, misleading, or unsupported

A few claims have outrun the evidence.

First, “AI is collapsing the grid.” Unsupported. The U.S. has not experienced a single AI-attributed blackout, and utilities publish granular load forecasts that account for AI campuses. Strain is real; collapse is not established.

Second, the framing that the public simply “rejects” AI because of its energy footprint is too broad. Stanford HAI’s 2026 synthesis reports that 59% of people globally still say AI offers more benefits than drawbacks, and KPMG’s 2025 survey finds 58% of respondents still view AI systems as trustworthy. Concern about energy is rising; wholesale rejection is not.

Third, the political spin that the issue is partisan-coded on direction is misleading. Pew’s 2025 data and Gallup’s 2026 Gen Z work both show concern crosses age, gender, and party lines. What is partisan is the proposed solution (carbon tax versus nuclear versus renewables versus a moratorium), not the underlying anxiety.

Fourth, the word “dominant” should not be attached to AI’s share of data-center load without a primary engineering source. Stanford HAI’s Technical chapter reports data-center electricity projections but does not single-cause-attribute all of them to AI.

What remains uncertain

The forecast band itself is the central uncertainty. A swing between 6% and 12% of U.S. electricity by 2028 is a doubling; that range reflects genuine disagreement among IEA, EPRI, Goldman Sachs, and utility-integrated resource plans about how fast new generation and transmission come online versus how fast hyperscalers place load.

Three other open questions matter:

• Per-facility water use. Total volumes are partly reported, but cooling-technology mix (air-cooled versus water-cooled versus closed-loop) varies sharply by site and operator, and disclosure is inconsistent.
• Embodied carbon in hardware. Chip fabrication, server replacement cycles, and rare-earth supply chains are real but rarely quantified in the same studies that report operational energy.
• Behavioral response. Whether consumer and enterprise users shift workloads, pay premium for greener regions, or simply accept higher utility bills as AI becomes normal is not yet measurable.

Where we are likely headed

The following is informed editorial judgment, not research output, and covers roughly the next two to five years.

Expect the public-opinion signal to keep rising before it plateaus. As more data centers break ground near residential voters, siting fights will produce more local news, and Stanford HAI’s “rising-concern cluster” framing will harden into a recognized category. Expect hyperscalers to lean harder into power-purchase agreements, behind-the-meter generation, and small modular reactor announcements as reputational cover — whether or not those deals land on the timeline promised. Expect the regulatory answer to fragment: more state-level disclosure rules, more interconnection-queue reform talk, and a slow attempt to weave environmental impact into the next round of EU AI Act guidance. The risk that the grid simply cannot absorb the load is low; the risk that local opposition and interconnection delays slow AI build-out more than any technical limit is high.

What this means for people and small businesses

For most households, the practical effect in the next two to three years is a modestly higher electricity bill in regions with new data centers and a longer wait for utility rebates or new connections if you are planning rooftop solar, an EV charger, or a home battery. None of that requires panic; it requires attention to local utility filings.

For small businesses, three things are worth doing now. First, watch the disclosure rules — if you sign a cloud or AI-vendor contract, ask whether the provider will disclose regional energy mix and water intensity on request; that is becoming a procurement question. Second, treat data-center availability as a real continuity risk in any region where utility interconnection is already constrained, and price redundancy into contracts. Third, do not assume energy cost is purely a household problem — small AI-heavy SaaS bills will be exposed to whatever pass-through rates emerge.

None of this means you should stop using AI tools, hold off on adoption, or treat the technology as uniquely toxic. It means the energy footprint is now a measurable, surveyed, and increasingly regulated dimension of the product, and ignoring it is no longer cost-free.

Bottom line

AI data centers are genuinely pulling on the power grid, and the public knows it. The most-cited U.S. polling puts concern about AI’s electricity use around six in ten adults, opposition to a local data center around seven in ten, and projected U.S. data-center electricity demand in a wide but clearly rising 6–12% band by 2028. The honest unknowns — water use, embodied carbon, the speed of grid build-out — are real. The exaggerated versions — imminent grid collapse, mass rejection of AI, AI as the sole cause of data-center growth — are not supported by the surveys or the engineering forecasts. Take the strain seriously; treat the panic as overblown; and watch the local utility docket, not the headlines.