Energy and artificial intelligence: the gap between production and governance

© 2026. All rights reserved. | Juan Antonio Lloret Egea | ORCID: 0000-0002-6634-3351 | Instituto Tecnológico Virtual de la Inteligencia Artificial para el Español™ (ITVIA) | IA eñ™ — La Biblia de la IA™ | ISSN 2695-6411 | CC BY-NC-ND 4.0 | Article 007 · 20 November 2026

Spanish original published by ITVIA on 13 November 2026 · DOI: 10.21428/39829d0b.8a1d7e51

Energy and artificial intelligence: the gap between production and governance

A power plant can generate electricity and a language model can produce a convincing answer. Neither achievement, on its own, tells us who can intervene when something goes wrong or who will bear the cost. Profitability cannot answer those questions. My concern as Director is what happens when the prospect of a return justifies expansion, while some of the risks, costs and obligations fall to citizens and future generations who had no say in assuming them.

In a pressurised water reactor, fission heats water in the primary circuit. That heat produces steam in a separate circuit; the steam drives a turbine coupled to an electrical generator. Operators govern this process through physical systems and verifiable procedures for controlling the reaction, removing heat and containing radioactive material [1]. Power plants also make authorised, monitored discharges, including water containing tritium, and accidental leaks have occurred. A controlled release, contamination and harm to health are distinct findings. Assessing a particular exposure requires keeping them distinct [2].

Waste extends that responsibility beyond the life of the plant. Disposal methods exist, including geological disposal, which combines natural and engineered barriers. Before such a facility can be developed, its long-term safety must be demonstrated [3]. Chernobyl and Fukushima illustrate how serious an accident can be, but the United Nations Scientific Committee on the Effects of Atomic Radiation distinguishes their radiological consequences [4]. For how long, and with what safeguards, does the electricity we consume today bind those who come after us?

Comparing energy sources requires considering the full range of harms. The Intergovernmental Panel on Climate Change identifies nuclear power as a low-carbon source and discusses the costs and challenges of final waste disposal. The life-cycle studies it cites put the health impacts of nuclear power under normal operating conditions below those of fossil fuels and at levels comparable to renewables [5]. Burning fossil fuels releases greenhouse gases and air pollutants with documented effects on health [5, 6]. Renewables, too, require materials, land and infrastructure. Neither today’s price nor a ranking that omits the conditions under which harm occurs is a sufficient basis for choosing among them.

The comparison with artificial intelligence concerns responsibility. The mechanisms are different. Large language models (LLMs) can solve complex tasks and also produce false statements, often called hallucinations. The International AI Safety Report 2026 finds that their internal mechanisms remain poorly understood and that tests before deployment cannot reliably predict every behaviour in practice [7]. Evaluation takes place before and after deployment. Its limits make it necessary to justify each use, particularly where an error could jeopardise someone’s rights or safety.

The link with energy is physical. Training models and running them to generate answers require computing power and cooling. The International Energy Agency estimates that data centres consumed 485 terawatt-hours (TWh) of electricity in 2025 and projects 950 TWh for 2030. These figures cover all data centres; they cannot be attributed to LLMs alone. Electricity use grew faster at centres dedicated to AI. Simple tasks can take less energy even as more complex reasoning and generation, used at greater volume, offset those savings [8]. An efficiency gain on one task does not guarantee a fall in total demand.

Who pays matters. Maryland’s Office of People’s Counsel attributes part of the 2026 increase in electricity supply costs for Baltimore Gas and Electric customers to data centres and their anticipated growth. Distribution costs and extreme temperatures also affect bills [9]. In March, the White House secured pledges from companies to provide for the new generation and grid upgrades they require [10]. Contracts, tariffs and actual bills will show whether those pledges protect households.

The November elections give Texas a means of testing that distinction. In September the governor directed the state environmental regulator to halt data-centre permitting pending an audit by ERCOT, the grid operator [11]. Governor-elect Gina Hinojosa had advocated a moratorium until legal safeguards were in place [12, 13]. Voters can now judge what becomes of that commitment. A change of government creates an opportunity; it does not establish that costs have been fairly allocated. Which obligations will be enforceable, and who will pay if companies’ forecasts prove wrong?

I see no reason to treat the commercial requirements of frontier models as collective needs by default. Before expanding deployment, we should establish what a model is useful for, how reliably it performs that task, how much energy it consumes and who can intervene. Research must have room to pursue less demanding methods, and society must be able to decline uses whose benefits do not justify their costs. Energy planning must account for environmental limits, financial responsibility and public participation. Corporate demand alone is no scientific mandate.

When production expands faster than our ability to assess, limit and correct its effects, the gap between production and governance widens. We can narrow it by setting a pace and scale that the evidence supports, retaining effective means to intervene and assigning costs to those who generate them. That is a way of pursuing technological benefits with lower risks and clear responsibilities, without promising a degree of safety that neither energy systems nor AI can guarantee.

Sources (13 references)
  1. U.S. Nuclear Regulatory Commission. Pressurized Water Reactors. Reactor operation and the primary and secondary circuits.
  2. U.S. Nuclear Regulatory Commission. Backgrounder on Tritium, Radiation Protection Limits, and Drinking Water Standards. See “Tritium from Nuclear Power Plants” and “How do people become exposed to tritium?”
  3. International Atomic Energy Agency. Radioactive waste disposal. Containment, isolation and long-term safety.
  4. United Nations Scientific Committee on the Effects of Atomic Radiation. UNSCEAR 2020/2021 Fukushima Report: Frequently Asked Questions and Answers. Questions 3–4 and 7, including the comparison with Chernobyl.
  5. Intergovernmental Panel on Climate Change (2022). Climate Change 2022: Mitigation of Climate Change. Chapter 6, “Energy systems”, especially § 6.4.2.4 and FAQs 6.1–6.2.
  6. World Health Organization (3 July 2025). Air pollution: tackling a critical driver of the global NCD crisis.
  7. Bengio, Y., et al. (2026). International AI Safety Report 2026. DSIT 2026/001. Executive summary and sections on reliability and risk evaluation.
  8. International Energy Agency (2026). Key Questions on Energy and AI. Executive summary: estimates, projections and differences between applications.
  9. Maryland Office of People’s Counsel (2026). Why is my summer BGE bill so high? Sections on distribution, energy, capacity and transmission.
  10. The White House (4 March 2026). Ratepayer Protection Pledge. Commitments concerning generation, infrastructure and rates.
  11. Office of the Texas Governor (21 September 2026). Governor Abbott Directs TCEQ To Halt Data Center Permits. Official announcement and directive.
  12. Texas Municipal League (6 November 2026). Legislative Update. No. 43, “2026 General Election Results”, pp. 1–2.
  13. Gina Hinojosa for Texas Governor. Data Centers. Campaign policy; cited as a political commitment, not as enacted law or an independent assessment.

Sources consulted on 13 November 2026.

Liked this post? Follow this blog to get more.