AI agents are completely changing the game when it comes to energy consumption in the tech sector.
If you have ever wondered why big tech companies are spending billions to build ever-larger data centers, the answer goes way beyond the chatbot you use to ask a quick question or look up a cake recipe.
The era of the simple chatbot is behind us, and what is coming next is a lot heavier, literally. 🔋
The new frontier of artificial intelligence is driven by systems that make decisions on their own, execute complex tasks, and keep running for hours without needing a human at the controls. That comes at a cost, and that cost goes far beyond money.
Carbon emissions, water consumption, and the demand for computing power are all growing at a pace that has a lot of people worried, including climate scientists and independent researchers trying to understand the true scale of this impact. The problem is that companies are not exactly transparent about these numbers, which makes the debate even more complicated.
In this article, we are going to explore:
- What AI agents are and why they consume far more energy than a regular search
- How OpenAI experiments with thousands of agents reveal the scale of that consumption
- Why projects like Meta’s Hyperion include 10 natural gas power plants to feed a single data center
- What all of this means for global sustainability goals
- And whether nuclear energy has any real shot at being part of this equation
Let’s dig into what is really going on behind the scenes of this tech race. 🚀
What are AI agents and why are they so different
When we talk about a traditional chatbot, the process is relatively simple: you type a question, the model processes it, and gives you an answer. It is a one-off interaction, almost like flipping a light switch on and off. AI agents, on the other hand, work in a completely different way. They do not wait for a single command to act. They plan, execute steps, consult external tools, review their own results, and keep operating in loops until they complete a complex task, all without a human needing to step in at every turn.
As journalist Maxwell Zeff from WIRED explained, instead of asking a simple question and getting a straightforward answer, these agents can generate hundreds of small prompts from a single user request. If someone asks an agent to build a website, for example, it can run for hours assembling features, reprogramming itself dozens of times in the process to create pages, menus, and the databases that hold everything together.
Receive the best innovation content in your email.
All the news, tips, trends, and resources you're looking for, delivered to your inbox.
By subscribing to the newsletter, you agree to receive communications from Método Viral. We are committed to always protecting and respecting your privacy.
Think of it this way: while a chatbot answers a question, an AI agent can browse the internet, read reports, compile data, write code, test it, and even send an email with the results, all autonomously. Each one of those steps requires calls to the language model, which means the demand for computing power multiplies exponentially compared to a simple conversation. And when you have thousands of these agents running at the same time, the impact on infrastructure becomes very clear.
When 10,000 agents solve a math problem
One of the most impressive, and at the same time somewhat alarming, examples came from OpenAI itself. The company announced that a swarm of more than 10,000 agents sending 2.7 million messages had solved a longstanding math problem. It is worth noting that some mathematicians disputed those claims, so take the news with a grain of salt.
Even so, the episode is revealing. It shows just how far cutting-edge labs are willing to go to try to solve problems once considered impossible, throwing absurd amounts of resources at the process. All those messages burned through a mountain of computing power, which translates to a whole lot of energy spent. The cost is estimated to have been in the tens of millions of dollars, although the exact figure is virtually impossible to pin down.
This kind of operation is an extreme case, sure. But it illustrates the central point of the discussion: the model of prolonged, intensive use by agents is the main factor pushing major tech companies to rethink, and more importantly to aggressively expand, their physical infrastructure.
The transparency problem
Private AI companies have historically cherry-picked what they share when it comes to environmental metrics for their products. Many executives tend to point to individual queries as a way to measure resource usage, which ends up downplaying the perception of the real impact.
An interesting example came from Sam Altman, CEO of OpenAI. In a recent podcast interview, he claimed that the water needed to grow a single almond would be equivalent to 38,000 ChatGPT queries. The math was widely challenged by experts. According to Altman, people who eat 12 almonds in one sitting do not feel like they are doing something terrible from a water consumption standpoint, most of the time.
The problem is that adding AI agents into that equation changes everything. They are far more energy-intensive than simple queries, and the tasks they perform can range from small jobs to an entire day of autonomous coding involving a team of helper agents running in parallel. The gap between those two consumption scenarios is enormous, and the expansion is potentially limitless as tasks grow more complex.
Boris Gamazaychikov, co-founder and CEO of the research and consulting group Sustainable AI, sums it up well. According to him, in other areas of tech growth there was a natural ceiling, like the number of people driving a car or watching Netflix. Now, that consumption is somewhat decoupled from users. And if you pay attention to what AI leaders are saying, that is exactly what they want, openly talking about billion-dollar companies with a single human employee.
But a single human employee, in that imagined world, would come with hundreds or thousands of AI agents working behind the scenes. That is the future these companies are racing toward, and it helps explain the urgency to build data centers everywhere.
Doing the math: the real consumption of agents
Since so little reliable information comes out of the companies themselves, some enthusiasts decided to try running the numbers on their own. Climate scientist Zeke Hausfather published a post analyzing how much energy his own AI usage, which relies heavily on agents, actually consumes. Using various sources, he concluded that his average daily session on Claude could use more energy than what it takes to keep two refrigerators running.
Gamazaychikov, whose group is expected to release research soon with more precise calculations on the environmental footprint of agents running on closed models, praised Hausfather’s effort but noted that his math was based on somewhat outdated data. Not exactly surprising, given how little academic work exists on the topic and how opaque tech companies are when it comes to sharing emissions metrics.
Hausfather concludes that, in the grand scheme of his personal life, having AI usage equivalent to a few extra refrigerators is not a number that is going to end the world. But he raises an important point: this usage represents a net new source of emissions at a time when global temperatures are skyrocketing and carbon reduction targets are increasingly off track. And it is far larger than the fraction-of-an-almond figures that Altman has been using as a benchmark.
Bigger data centers, heavier energy bills
Hausfather admits he uses AI and agent-based tools more than most people, but that could change fast. Meta recently launched a personal AI agent that, according to the company, was built to work for billions of people around the world. Called Muse, it promises to maintain a dedicated cloud computer for each user, running even when the person is offline. The company plans to integrate Muse into its smart glasses later this year.
It is entirely possible that in the near future, Meta users wearing the glasses or scrolling through Facebook will be delegating tasks to agents without even realizing what they are doing. And if Meta envisions a future where everyone uses an agent, that explains the jaw-dropping scale of some of the data centers being built.
The Hyperion project in Louisiana is the most telling example of this race. It is a massive data center that will be powered by 10 natural gas power plants. That is right: a single data facility with its own fossil fuel-based power generation campus, which raises serious questions about where we are heading when it comes to sustainability.
As Gamazaychikov puts it, the technology that will be trained by the data centers being proposed and built right now is still three to five years away. And it is going to look and feel very different from the simple chatbot window we know today.
Sustainability on the line: what the numbers say
The debate around sustainability in the context of artificial intelligence has gained an extra layer of complexity because the companies themselves control which data they make public. There is no global transparency standard that forces big tech to report the specific energy consumption of their AI systems. This means the numbers circulating in the media tend to be estimates made by external researchers, not verified official data. This lack of clarity makes it much harder to create effective public policy.
Water consumption is also part of this equation and tends to fly under the radar in public discussion. Data centers use cooling systems that depend on large volumes of water to keep servers at safe operating temperatures. As the demand for computing power grows, water usage scales proportionally, creating additional pressure in regions already facing water stress. In other words, the environmental impact goes beyond carbon emissions and touches resources that are essential to entire communities surrounding these facilities.
Nuclear energy: real solution or a bet on the future
A question that comes up a lot is whether small nuclear plants could work to power data centers. The short answer is: yes, they could be a great carbon-free energy option. Several startups and data center developers dream of a future where these facilities happily operate alongside small modular reactors, or SMRs, connected to the power grid.
The problem, as always with nuclear, is how long it could take. No small modular reactor is commercially operating in the United States, and only one design has been licensed for sale, despite decades of development. The U.S. government has been trying to help the sector mature faster, including a pilot project at the Department of Energy for 11 startups to hit a key milestone. At least a few of them got there. Now comes the long journey of bringing those products to market.
But many data center developers do not want to wait years for these companies to prove their worth. So they are installing gas turbines right now. In other words, they are not waiting for utopia to arrive. Meanwhile, giants like Microsoft, Google, and Amazon have already signed deals or expressed interest in nuclear energy. Microsoft, for instance, closed a deal to restart the Three Mile Island plant in Pennsylvania, specifically to power its operations.
The most realistic scenario for the coming years is that no single energy source is going to solve the AI agents consumption equation on its own. The trend points to a combination of renewable sources, nuclear, energy efficiency in the chips themselves, and changes in language model architecture to reduce the computational cost per task. It is a race on two fronts: increasing the supply of clean energy and reducing the demand for raw power.
What to expect going forward
The expansion of AI agents is not a passing trend. On the contrary, as models become more capable and use cases multiply, the demand for data centers is likely to grow even more. And the effects are already being felt in practice, including politically. In some U.S. states that offered billions in tax incentives to attract data centers, governments are already walking back those deals. In places like Texas, the issue has even become an election topic, pushing some conservative voters away from their traditional parties.
Governments around the world are starting to pay attention, whether because of the impact on local power grids or because of climate targets set in international agreements. On the corporate side, investor pressure for consistent environmental metrics is also forcing a more serious conversation about how to balance growth with responsibility. It is not that big tech does not care about sustainability, it is that the financial incentives to expand AI capacity are, for now, much greater than the incentives to slow that growth for environmental reasons.
What becomes clear in all of this is that the development of AI agents is not just a technology question. It is a question of infrastructure, energy policy, sustainability, and collective choices about what kind of future we want to build. Technology moves fast, and the impacts come right along with it. Understanding this context is the first step toward any serious debate about where artificial intelligence is heading, and about who is going to foot the bill in the long run. 🌍
