# AI agents are blowing up the energy bill of data centers > A research team at the Korea Advanced Institute of Science and Technology, known as KAIST, led by researcher Rhu Min-soo, published results - Author: Maxime Marquette - Publisher: MadMax (https://mad-max.co) - Section: Commentary - Published: 2026-07-06 - Canonical URL: https://mad-max.co/en/article/les-agents-ia-font-exploser-la-facture-energetique-des-centres-de-donnees ## Key takeaways - The researchers used a 70-billion-parameter model and measured actual energy consumption rather than relying on theoretical estimates. - If AI agent requests reached 13.7 billion per day worldwide, data center power demand would reach roughly 198.9 gigawatts, according to the study. - The study was presented at the HPCA 2026 conference. - The author argues that tech companies should be more transparent about the energy footprint of their AI products. - The author links the issue to competition with China and suggests nuclear power and energy efficiency as possible responses. ## Introduction: a Korean study that disrupts the optimistic AI narrative ### A number that changes the conversation about artificial intelligence A research team at the **Korea Advanced Institute of Science and Technology**, known as **KAIST**, led by researcher **Rhu Min-soo**, published results on **July 5, 2026** that should worry anyone seriously tracking the energy trajectory of **artificial intelligence**. Unlike **classic generative AI**, which answers a simple question, **AI agents** capable of autonomously executing complex tasks consume on average **348.41 watt-hours per request**. That figure represents up to **136.5 times more energy** than a standard request sent to a classic language model, a gap so dramatic it forces a complete rethink of energy consumption projections tied to the massive rollout of **agentic AI** across the Western economy. ### Why this technical distinction matters enormously An **AI agent** doesn't just generate text: it **plans**, executes **successive actions**, interacts with **external tools**, checks its own results, and starts again if necessary. This **multi-step architecture** mechanically multiplies the number of computations performed by the underlying **language model** for every task a user requests. The **KAIST** researchers used a **70-billion-parameter** model for their measurements, a representative choice given the systems currently deployed in production at major **Western and Asian tech companies**. *I'll say it plainly: this study should serve as a brutal wake-up call for anyone selling agentic AI as a miracle solution without ever mentioning its real energy bill. You can't celebrate innovation while ignoring its concrete environmental costs.* ## Understanding the difference between generative AI and AI agents ### The shift from a passive tool to an autonomous system Traditional **generative AI**, the kind most users know through consumer chatbots, answers a query with a single response, generally in one computing pass. **AI agents**, by contrast, represent a major architectural shift toward systems capable of **decision-making autonomy** across multiple successive steps. This increased autonomy means every complex task potentially triggers dozens of **internal calls** to the **language model**, each consuming **energy**, unlike a simple query that requires only a single computing pass. ### The real-world uses behind this consumption explosion **AI agents** are increasingly used to automate **complex professional tasks**: in-depth research, multi-step report writing, managing entire **workflows**, or coordinating between several different software tools within the same company. Each of these tasks, however useful for **productivity**, carries an **energy cost** far higher than what users, and even many tech companies, currently seem to realize or publicly disclose. *What strikes me is the industry's relative silence on this reality. We're sold the autonomy of AI agents as pure progress, without ever being told that this autonomy comes with a concrete and considerable energy price tag.* ## Numbers that make your head spin ### A planet-scale projection **KAIST** researchers estimate that if **AI agent** requests reached **13.7 billion per day** worldwide, a scenario considered plausible given the current adoption trajectory, the power required across **data centers** would climb to roughly **198.9 gigawatts**. That figure far exceeds the **energy capacity** currently installed across all **data centers** dedicated to artificial intelligence worldwide, meaning **existing infrastructure** will need to be massively expanded to absorb this projected demand. ### A comparison that puts things in perspective To give a sense of scale, **198.9 gigawatts** represents power comparable to dozens of large **power plants** running simultaneously, solely to meet the computing demand generated by **AI agents** worldwide. This estimate, published as part of a scientific paper presented at the **HPCA 2026** conference, a recognized reference in computer architecture, lends solid academic credibility to figures that might otherwise seem exaggerated or speculative. *Numbers of this magnitude deserve to be repeated calmly, without unnecessary doom-mongering, but also without complacency. We're talking about dozens of additional gigawatts solely for software agents, not for our overall civilian energy needs.* ## The broader context of the AI-driven energy crisis ### A problem the industry already saw coming This **KAIST** study adds to a growing series of warnings from **energy transition** experts who had already flagged that artificial intelligence's exponential growth could hit the physical limits of current Western power grids. Recently published **economic analyses** note that AI-related **energy demand** has already begun creating tension on certain **regional power grids**, particularly in areas where major **data centers** of **American tech giants** are concentrated. ### Tech companies facing their own contradictions Several **major Western tech companies** have made **ambitious climate commitments** while investing massively in **agentic AI** infrastructure whose real energy consumption remains largely under-communicated to the general public, and sometimes even to their own shareholders. This tension between **environmental promises** and the reality of advanced artificial intelligence systems' **power consumption** puts the **tech industry** in an increasingly difficult position to defend publicly. *I remain convinced the West must lead the artificial intelligence race, but not by closing its eyes to the real energy costs. Transparency on these numbers is a necessary condition for building a coherent energy strategy against China and other rival tech powers.* ## The methodology behind these results ### A rigorous measurement protocol **Rhu Min-soo's** team measured the actual energy consumption of **70-billion-parameter** **language models** executing tasks representative of **AI agent** uses currently deployed in professional settings, rather than relying on abstract theoretical estimates. This **empirical approach**, based on direct measurements rather than extrapolations, considerably strengthens the **scientific credibility** of the conclusions published by the **South Korean researchers** in this study presented at the **HPCA 2026** conference. ### Methodological limits to keep in mind Like any study of this kind, these results rest on specific assumptions about the types of tasks executed and the hardware architecture used, meaning real-world figures could vary depending on the technological configurations deployed by different companies. Nonetheless, even accounting for a **reasonable margin of error**, the magnitude of the consumption difference between **classic generative AI** and autonomous **AI agents** remains large enough to warrant serious attention from Western **policymakers** and industry leaders. *This Korean team's methodological rigor deserves recognition. Too many AI studies settle for vague estimates, whereas here we have concrete empirical measurements on real hardware, running real tasks.* ## The strategic stakes for the West against China ### A tech race that also carries an energy cost The competition between the **West** and **China** for dominance in **artificial intelligence** isn't limited to raw computing power or algorithm quality: it now includes a crucial energy dimension that few Western analysts had anticipated at this scale. If **AI agents** become as indispensable as expected in the global economy of coming years, a country's ability to produce enough **reliable, affordable electricity** to power its **data centers** will become a decisive factor in its overall **technological competitiveness**. ### The risk of the West falling behind better-prepared rivals **China**, which is investing massively in **power generation capacity**, including **nuclear** and **coal**, could paradoxically find itself better positioned to fuel its **agentic AI** ambitions than certain Western countries mired in political debates over the energy transition. This reality forces the **West** to treat the AI energy question not as a secondary environmental issue, but as a genuine matter of **strategic security** and technological sovereignty against rivals who don't face the same internal political constraints. *This is a point I find underestimated in Western public debate: if we want to win the AI race against China, we also need to win the race for the energy that powers it. Ignoring this dimension is condemning ourselves to a major strategic setback.* ## The technological solutions the industry is considering ### Toward more energy-efficient architectures Faced with these findings, several research labs are actively working on more efficient **language model** architectures capable of accomplishing the same complex tasks with significantly reduced energy consumption compared with current systems. These efforts include **model compression** techniques, **optimization of redundant computations**, and the development of **specialized chips** designed specifically to maximize the energy efficiency of inference tasks tied to **AI agents**. ### The role of next-generation semiconductors Western and Asian chipmakers are investing massively in developing **semiconductors** specialized for artificial intelligence, with the explicit goal of reducing energy consumption per unit of computation while maintaining, or even increasing, available processing power. These investments in **next-generation semiconductors** could partially offset the surge in demand tied to **AI agents**, but **experts** agree these **efficiency gains** will likely be insufficient to fully neutralize the projected growth in total consumption. *I trust the Western industry's ability to innovate on energy efficiency, it's historically one of our strengths. But let's not kid ourselves: technological innovation alone probably won't be enough to offset such staggering demand growth.* ## Implications for regional power grids ### Pressure already visible in some regions In several regions where major American **data centers** are concentrated, **power grid** operators are already reporting signs of strain tied to growing demand from artificial intelligence infrastructure, a situation that could worsen considerably as **AI agents** become widespread. This pressure translates concretely into **longer wait times** for connecting new industrial projects, along with growing fears about **power grid** stability during simultaneous **demand peaks** between residential and industrial use. ### The debate over who should pay AI's energy bill A political and economic debate is gradually emerging over whether the tech giants operating these energy-hungry **data centers** should shoulder a larger share of the electrical infrastructure costs needed to support their growth. Some **Western lawmakers** are beginning to propose **specific pricing mechanisms** for large industrial consumers tied to AI, an approach that could significantly redefine the tech industry's **economic model** in the years ahead. *It would be unfair for ordinary residential consumers to bear higher electricity bills to finance the energy-hungry expansion of tech giants. Those who benefit most from this growth must also bear a fair share of the costs.* ## What this means for the businesses using this technology ### An operational cost that will need to factor into profitability calculations Companies deploying **AI agents** to automate internal processes will now need to more seriously factor the real energy cost of these tools into their **return on investment** calculations, a factor often overlooked in the initial enthusiasm for these new technologies. This awareness could slightly slow the adoption pace of the most resource-hungry **AI agents**, in favor of **targeted solutions** better optimized for specific, **high-value-added** tasks. ### The emergence of a new discipline: AI energy optimization We're already seeing the emergence, within major tech companies, of teams dedicated specifically to optimizing the energy consumption of **artificial intelligence** deployments, a discipline that barely existed two or three years ago. This **professionalization** of AI energy management reflects a growing awareness that the issue is no longer marginal, but is becoming a **central concern** in the **technology strategy** of the most advanced Western companies. *This is a logical and necessary evolution. We can no longer treat AI's energy consumption as a secondary technical detail when it's becoming a determining factor in the economic and environmental viability of these technologies.* ## Reactions from the international scientific community ### A study that confirms fears already voiced elsewhere This **KAIST** study aligns with the conclusions of other recent work by Western researchers who had already warned about the unsustainable trajectory of AI-related energy consumption without significant technological or political intervention. The convergence of these different studies, conducted independently in several countries, considerably strengthens the **collective credibility** of these warnings and makes it harder for **industry players** with an interest in downplaying this **issue** to dismiss them. ### The importance of South Korean research in this field **South Korea**, through institutions like **KAIST**, is gradually establishing itself as a leading scientific player in studying the technical and energy implications of **artificial intelligence**, contributing to a global debate often dominated solely by American and Chinese voices. This South Korean scientific contribution illustrates the importance of maintaining a diversity of academic perspectives in such a strategic field, rather than depending solely on narratives promoted by the major tech companies themselves. *I welcome this Korean contribution to the global AI debate. Too often, we only hear American and Chinese voices on these issues, while countries like South Korea bring valuable scientific rigor and a welcome independent perspective.* ## Possible scenarios for the coming years ### A scenario of gradual technological adaptation In the most optimistic scenario, the tech industry manages to develop more energy-efficient **AI** architectures quickly enough to absorb a significant share of projected demand growth, thereby limiting strain on Western **power grids**. This scenario assumes massive, coordinated investment in research and development, along with clear political will to prioritize energy efficiency in future generations of **semiconductors** dedicated to artificial intelligence. ### A scenario of prolonged energy tension In a more pessimistic scenario, demand growth tied to **AI agents** far outpaces the **West's** technological and infrastructural capacity to adapt, leading to prolonged strain on electricity prices and potentially forced slowdowns in the rollout of certain artificial intelligence applications. This scenario could paradoxically favor countries with surplus power generation capacity, regardless of their democratic values or geopolitical alignment with the **West**, a prospect that should worry Western strategists. *Unfortunately, I lean toward an intermediate scenario, where the West moves forward in fits and starts, without the strategic coordination needed to truly anticipate this energy demand explosion. There's still time to correct course, but the window is closing fast.* ## The role of Western public policy ### Governments still behind on this issue Despite the severity of the figures revealed by this **KAIST** study, many Western governments have not yet fully integrated the energy dimension of **agentic AI** into their national industrial and energy policies, a worrying gap given how quickly these technologies are being deployed. This political slowness contrasts with the urgency of the investments needed in **electrical infrastructure**, which generally require years of planning and construction before they can meet such rapidly growing demand. ### Possible actions to speed up Western preparedness Experts particularly recommend accelerated investment in **nuclear energy** production capacity, considered by many the most realistic solution to provide the stable, massive electricity needed for **AI agent** **data centers** without compromising Western climate goals. Others propose targeted tax incentives to encourage tech companies to invest directly in their own power generation capacity, rather than simply drawing from public **power grids** already under strain. *Nuclear power should be at the center of the Western energy strategy for AI, without ideological hesitation. It's the only currently available technology capable of providing the stable, massive electricity our data centers will need in the coming years.* *It's time for Western governments to stop treating nuclear as just one controversial option among others. Given the scale of needs revealed by this study, it has become an unavoidable strategic necessity.* ## What consumers should understand from this study ### An invisible but very real energy bill For the general public, this **KAIST** study makes visible a reality that was largely invisible until now: every time an **AI agent** executes a complex task on your behalf, that action consumes far more energy than most users imagine. This awareness could, over time, influence the usage habits of consumers and businesses, favoring more thoughtful, targeted use of **AI agents** rather than blanket deployment for every possible task. ### The importance of transparency from tech companies This study strengthens the case for greater **transparency** from tech companies about the real energy footprint of their **artificial intelligence** products, transparency that today remains largely insufficient given the scale of the issues this research reveals. Consumers and business clients have a legitimate right to know the real energy cost of the tools they use, particularly at a time when climate concerns occupy a central place in Western public debate. *Transparency isn't optional, it's a democratic necessity. Western citizens need to know exactly what it costs, in real energy terms, every time they interact with these systems that are becoming omnipresent in our professional and personal lives.* ## The cryptocurrency precedent and lessons to draw from it ### An energy crisis we've already seen with crypto mining The Western tech industry has already lived through a comparable episode with the surge in energy consumption tied to **cryptocurrency mining**, which forced several regions to urgently rethink their electricity supply policies in the face of unforeseen, massive industrial demand. The lessons drawn from that earlier crisis, particularly on the need to anticipate demand rather than simply endure it, should directly inform how Western governments approach today's foreseeable surge in **AI agent** consumption. ### Avoiding a repeat of the same planning mistakes Unlike the cryptocurrency episode, whose scale caught many governments by surprise, **KAIST** researchers now offer Western decision-makers precise, scientifically grounded **numerical projections** on which to build proactive rather than reactive energy planning. Ignoring these warnings would amount to deliberately repeating past mistakes, especially now that the data needed to act intelligently and in time is publicly available thanks to work like this. *We already had a warning with cryptocurrencies and we didn't really listen. This time, with numbers as precise as KAIST's, the West will no longer have the excuse of saying it wasn't warned in time.* ## Conclusion: an energy bill the West can no longer ignore ### An alarm signal that deserves a coordinated response The study published by **KAIST** on **July 5, 2026** is a rigorous scientific alarm signal that Western policymakers and industry leaders can no longer afford to ignore. The difference in energy consumption between **classic generative AI** and autonomous **AI agents** is too significant to be treated as a minor technical detail. Facing a projection that could reach **198.9 gigawatts** of additional demand, the West must quickly develop a coherent energy strategy combining technological innovation, investment in electrical infrastructure, and greater transparency from tech industry players. ### An issue that goes beyond environmental concerns alone Beyond legitimate climate concerns, this energy question directly touches the West's ability to maintain its technological leadership against rivals like **China**, which could exploit any Western delay in preparing for the energy demands of the **AI agent** era. The response to this challenge will largely determine who truly dominates the next phase of the **artificial intelligence** revolution, and the West cannot afford to enter this decisive race a step behind on its fundamental energy infrastructure. **By Maxime Marquette, columnist** ## Columnist's transparency note ### Editorial stance I am not a journalist but a **columnist and analyst**. My expertise lies in observing and analysing the geopolitical, economic and strategic dynamics that shape our world. My work is to take political strategies apart, to understand global economic movements, to put the decisions of international actors in context and to offer analytical perspectives on the changes that are redefining our societies. I do not claim the cold objectivity of traditional journalism, which stops at reporting facts. **I claim analytical clarity**, rigorous interpretation and a deep understanding of the complex issues that concern us all. My role is to give meaning to the facts, to place them in their historical and strategic context, and to offer a critical reading of events. ### Method and sources This text keeps the basic distinction between **verified facts and interpretive analysis**. The rule never changes: a factual claim is published only when a verifiable source supports it, and the sources this article actually used are listed under **Sources**, never here. **Primary source families** used by the publication, when the story has them: official statements from governments and international institutions, public statements by political leaders, reports from intergovernmental organisations, dispatches from recognised international news agencies. **Secondary source families**: specialised publications, internationally recognised news media, analyses from established research institutions, reports from industry organisations. When an article cites statistical, economic or geopolitical data, it comes from the institutions that produce it (intergovernmental organisations, central banks, national statistics offices), and the exact institution is named under **Sources**. ### Nature of the analysis The analyses, interpretations and perspectives in the analytical sections of this article are a **critical, contextual synthesis** based on the information available, the trends observed and the expert comments cited in the sources consulted. My role is to interpret these facts, to set them in the context of today's geopolitical and economic dynamics, and to give them a coherent meaning within the larger story of the changes shaping our time. These analyses reflect expertise built through continuous observation of international affairs and an understanding of the strategic mechanisms that drive global actors. This article describes a documented state of affairs on its publication date, not a prediction: later developments may shift its outlook. No update is promised in advance; when an article is corrected or completed, the change is dated in the text. ## Sources ### Primary sources - Chosun Biz, KAIST study on the energy consumption of AI agents — July 5, 2026 ### Secondary sources - Forbes, the energy transition hits the wall of AI compute — June 30, 2026 - Chosun, markets and economic implications of AI energy demand — July 2, 2026 - Chosun, Korean industry facing the energy challenges of artificial intelligence — July 1, 2026 - CNBC, technology news — 2026 - Reuters, business and industry news — 2026 --- Source: Maxime Marquette, MadMax. When citing, name the author and link https://mad-max.co/en/article/les-agents-ia-font-exploser-la-facture-energetique-des-centres-de-donnees