AI's Energy Bill Comes Due: Why Footprint Disclosure Is the Next Credibility Test for AI Vendors

Short answer: AI has become a physical industry. US electricity use is set to hit new records in 2026 and 2027, global data centre demand is projected to more than double by 2030, and grid operators such as PJM say data centres are outrunning new supply. That turns energy from a PR topic into a buying and political question. AI vendors that publish measured, clearly scoped footprint numbers will shape how buyers, regulators and AI assistants describe them. Vendors that publish nothing will be described with someone else's numbers.

AI is now an infrastructure story

The constraint is no longer only chips. It is power. The US Energy Information Administration projects electricity use rising from 4,195 billion kWh in 2025 to 4,268 billion kWh in 2026 and 4,391 billion kWh in 2027, new records driven largely by data centres and electrification (World Economic Forum). Globally, the International Energy Agency estimates data centres used about 415 TWh in 2024, around 1.5% of the world's electricity, and expects that to reach about 945 TWh by 2030, slightly more than Japan uses today (IEA).

The strain is most visible on the largest US grid. PJM projects peak demand will grow 32 gigawatts from 2024 to 2030, all but 2 gigawatts of it from data centres. Its December 2025 capacity auction hit a record $16.4 billion and still fell short of its reserve target, and some customers face bill increases of 1.5-5% from June 2026 (Canary Media). "Data centre electricity consumption has grown by around 12% per year since 2017." - IEA (Energy and AI report) - statistic.

Two true numbers that tell opposite stories

The public debate swings between two kinds of figures, and both are accurate. Per query, AI looks tiny. Google reported in August 2025 that a median Gemini text prompt uses 0.24 watt-hours, emits 0.03 grams of CO2 and uses 0.26 milliliters of water, and that energy per median prompt fell 33-fold in a year (MIT Technology Review). Sam Altman put the average ChatGPT query at about 0.34 watt-hours.

In aggregate, AI looks enormous, because billions of queries, ever larger models, image and video generation, and training add up to gigawatts. Efficiency per query is improving fast; total demand is still climbing. A vendor that cites only the first number sounds evasive to someone who has read about the second. "The average query uses about 0.34 watt-hours." - Sam Altman ("The Gentle Singularity," 2025) - company statement.

The per-prompt figures also come with caveats. Google's estimate covers text only, not images or video, and uses market-based emissions that reflect its clean energy contracts rather than local grid averages. Researchers welcomed the disclosure but warned it is not a basis for comparing companies. "This is not a replacement or proxy for standardized comparisons." - MIT Technology Review (Sasha Luccioni, Hugging Face, 2025) - expert view.

From PR to buying criterion

Three pressures are moving energy from the sustainability report into the sales cycle. First, regulation: under the EU Energy Efficiency Directive, data centres with at least 500 kW of installed IT power must report 24 indicators, including energy, water, renewable share and waste heat, every year by May 15, feeding an EU-wide database (CMS). Second, politics: when data centres show up in household bills, local opposition and state debates over who pays follow. Third, procurement: enterprise buyers with their own emissions targets increasingly need supplier data to report against them.

Example: a procurement team evaluating two AI platforms asks each for energy and emissions data per workload. One supplies measured figures with a methodology note. The other sends a press release about renewable energy commitments. The first is easier to approve, because the buyer can put its numbers into their own reporting.

How the story gets retold

Buyers increasingly research vendors through AI assistants, and those assistants answer "how sustainable is [vendor]?" from whatever is published and cited. When a vendor has published nothing specific, the most likely answer draws on industry-wide figures, headlines about grid strain or a competitor's framing. When a vendor has published clear, scoped numbers with a methodology, there is at least a credible source to cite.

That makes footprint disclosure part of a vendor's public fact base, alongside pricing, security and product capabilities. The same discipline applies: specific numbers, clear scope, consistent wording across the website, reports and third-party coverage, and regular updates as efficiency changes. Some GEO/AEO teams and providers, such as Hordus AI, already work on keeping brand facts consistent across these surfaces (Hordus). The substance, though, has to come from the vendor's own measurement.

A practical disclosure checklist for AI vendors

1. Measure per unit of work. Publish energy, emissions and, where relevant, water per request, per task or per customer workload, not only company totals.
2. State the scope. Say what is included (chips, servers, cooling, idle capacity) and what is not (training, images, video).
3. Show both accounting methods. Report location-based and market-based emissions, so readers can see grid impact and clean energy purchasing separately.
4. Report the trend. Efficiency gains are real and persuasive when they are dated and repeatable.
5. Don't hide the aggregate. Acknowledge total demand growth and how you plan to meet it, including power sourcing.
6. Avoid unsupported green claims. "Sustainable AI" without numbers invites skepticism and, in some markets, regulatory scrutiny.
7. Publish it where it can be found. One clear, dated, answer-first page that sales, investors and AI assistants can all point to.

What to watch next

Watch PJM's proposals for closing its supply gap and how costs are allocated to large loads, the European Commission's next steps on data centre performance standards using the EED data, and whether more model providers follow Google with detailed per-prompt methodologies. Also watch the IEA's next update on AI demand, which will show whether efficiency gains are bending the curve.

The bottom line

AI's energy footprint is no longer an abstract debate. It shows up in grid auctions, utility bills and procurement questionnaires. Both the per-query and the aggregate numbers are true, and credibility goes to vendors who publish their own, scoped clearly and updated often. Silence is also a disclosure: it lets others decide what your footprint is.