For three years, the energy cost of AI was an argument between researchers. In 2026 it became a line on a household power bill, and that changed who is talking about it. When a topic moves from technical debate to pocketbook politics, the volume of critical coverage rises sharply. And critical coverage is what AI assistants retrieve when someone asks about your company's environmental record.
This post is about that second-order effect. Not whether AI uses too much power, which is covered everywhere, but what the argument has done to how machines describe your brand.
What actually happened in 2026
The clearest evidence sits in capacity markets, where scarcity gets priced directly rather than negotiated in a rate case.
| What moved | By how much |
|---|---|
| PJM capacity price, 2024 to 2026 | $28.92 to $329.17 per MW-day, roughly 1,038% |
| Residential rates, Ohio and Pennsylvania | Up 9% and 14% over the trailing year |
| Cost estimated shifted onto public bills | Around $23 billion |
| Data center share of electricity demand growth | Roughly 40% |
| Projects blocked or delayed, Q1 2026 | 75 or more, worth around $130 billion |
| Active local opposition groups | 396 at the end of 2025, 833 across 49 states by March 2026 |
Project cancellations went from 6 in 2024 to 25 in 2025 to 20 in a single quarter of 2026. The most common objection is not carbon. It is water, cited in more than 40% of contested projects, with electricity rates second. We keep the full figures updated in our research on the electricity price impact of AI data centers and data center moratoriums and local opposition.
Data centers are not the only driver of rising prices. Generation retirements, transmission investment, and electrification of heat and transport all contribute, and analysts who push back on the simple story are right that isolating one cause inside a regulated rate is genuinely hard. But that nuance is not what determines your brand's exposure. Volume of coverage is.
Why this reaches companies that do not build data centers
Here is the mechanism, and it is worth understanding precisely because it generalises well beyond energy.
AI assistants are tuned to hedge where their sources disagree. For an uncontested factual claim, an assistant will usually just restate the figure a company publishes. For a contested one it does something different: it presents the claim, attributes it to you, and then presents the counter-analysis. That behaviour is deliberate and it is generally good. It also means that on any topic where credible sources disagree with your marketing, your strongest claim reliably arrives with its criticism attached.
Corporate environmental claims have now crossed into that category. Not because your specific claim is false, but because the topic as a whole is contested enough that the hedging behaviour triggers. Any company publishing an emissions or renewable-energy figure inherits the argument, whether or not it owns a single server.
The uncomfortable version
You do not control whether your sustainability claim is repeated or qualified. The contestedness of the topic decides that. What you control is which version of the correction appears next to your claim, and whether the numbers a model reaches for came from you or from someone rebutting you.
The 100% renewable example
The clearest recurring case is renewable energy matching. A company states it runs on 100% renewable energy. This is almost always annual matching: total clean energy procured across a year equals or exceeds total consumption. That is a real commitment and a real expense, and it is not the same as running on clean power.
A facility can be annually 100% matched while drawing coal-fired electricity at 2am on a still winter night, because the wind generation that balanced the books happened at a different hour. Hourly or 24/7 carbon-free accounting requires matching consumption to clean generation in every hour, which is much harder and much rarer.
That distinction is well covered, clearly written, and easy to retrieve. So when an assistant is asked about a company's renewable claim, the explanation of what annual matching does not mean is right there, usually in cleaner HTML than the claim itself. The companies that publish hourly figures alongside the headline tend to be described more accurately, because the more precise disclosure pre-empts the standard correction instead of inviting it. We go deeper on the regional side of this in carbon intensity of AI by region.
The part nobody expects: it is a format problem
When we look at why independent analysis outranks corporate reporting in AI answers, the intuitive explanation is credibility. A third party is more trustworthy than a self-interested one. That is true, and it is not the main thing happening.
| Corporate sustainability report | Independent analysis | |
|---|---|---|
| Format | Long PDF, often image-heavy | HTML article |
| Where the number lives | Inside a designed infographic | In a sentence, as text |
| Structure | Narrative with figures embedded | Specific claim, specific rebuttal |
| Retrievability | Frequently poor | Good |
A great deal of corporate environmental disclosure is locked in PDFs with the important numbers rendered as graphics. The rebuttal is plain text on a news site. Even a perfectly credible claim loses to a retrievable one, and the most carefully audited figure in your annual report may as well not exist if it only appears inside an image.
This is the same lesson as why ChatGPT cites Reddit more than it cites you, arriving through a different door. Machines reach for what is easiest to parse and quote, not for what took the most effort to produce.
What to actually do
Four things, in order of how much they change the outcome.
Put the numbers in HTML, as text. Keep the designed PDF for the people who want it, and publish a plain page with every headline figure written out in a sentence or a table. This single change does more than the other three combined.
State the standard you are using. The methodology is almost always the point of dispute, so naming it explicitly, annual matching versus hourly, market-based versus location-based, converts an ambiguous claim into a precise one that is harder to mischaracterise.
Pre-empt the known correction. If your metric has a well-understood limitation, say so before someone else does. The correction gets retrieved either way. Being its source is considerably better than being its subject.
Keep the figures current. A stale number invites the observation that it is stale, which is a second criticism you did not need.
The general principle: on contested topics, precision beats favourability. A modest, precisely stated, well sourced figure is represented more accurately than an impressive one that invites qualification. This runs against most communications instinct, which is exactly why so few companies do it.
Why this is the shape of things now
Energy is the current example because it is the argument the AI industry happens to be having in 2026. The mechanism is not about energy. Any topic where credible sources disagree with corporate messaging behaves this way: labour practices, safety records, pricing fairness, data handling. As AI assistants become the default first answer, the gap between what a company says and what a machine says about that company becomes a measurable thing rather than a matter of perception.
Most brands have no idea what that gap looks like for them, because it does not show up in any dashboard they already own. That is the part Presenc measures: how your own claims are represented in AI answers, which counter-sources appear alongside them, and whether changing your disclosure actually moves the result. You can see the underlying analysis in our research on energy and sustainability claims in AI answers.
