Silicon Valley’s technology giants are investing heavily in new data centres, raising questions about the reasons for construction on such a vast scale. The answer lies in the shift from simple chatbot queries to the use of more resource-intensive AI agents capable of performing tasks autonomously.
The evolution of AI: from chatbots to agents
Simple chatbot queries, such as searches for recipes or travel information, are becoming an outdated way of using artificial intelligence. Modern AI systems are evolving into agents — systems based on large language models that can make decisions and perform tasks independently. Unlike chatbots, which answer a single specific question, AI agents can generate hundreds of subqueries to achieve a goal. For example, building a website may require an agent to work for many hours, modifying itself and generating various website components.
AI agents are at the forefront of artificial intelligence research. One recent example involved solving a complex mathematical problem using more than 10 000 agents, which sent 2,7 million messages. Although this is an exceptional case requiring enormous computing resources, it demonstrates both the potential and the energy intensity of such systems. Experts estimate that processing this volume of messages could cost tens of millions of dollars, although the exact figures remain a trade secret.
Energy consumption and data transparency
Companies developing AI have historically been reluctant to disclose environmental metrics for their products. Examples of individual usage are often cited as indicators of energy consumption, such as comparisons between the water used to produce almonds and the number of queries submitted to ChatGPT. However, the emergence of AI agents, which are significantly more energy-intensive than simple queries, complicates such calculations. Information about agents’ energy consumption is extremely limited, and their tasks range from simple activities to complex ones, such as autonomous programming involving numerous supporting agents. This gap in energy consumption could grow without limit as tasks become more complex.
Unlike other areas of technology, where growth is limited by the number of users (for example, the number of drivers or streaming service viewers), the use of AI agents can be ‘decoupled’ from the end user. AI company executives speak of a future in which a single human employee could manage hundreds or thousands of AI agents. This vision of the future explains the current race to build data centres. With reliable data from companies in short supply, enthusiasts are attempting to estimate energy consumption themselves. Climate scientist Zeke Hausfather calculated that his daily use of AI agents could consume more energy than two refrigerators. Although this is not a critical level of consumption for personal use, it represents a new source of emissions amid global warming.
Scaling up and the future of AI infrastructure
Meta recently unveiled Muse, a personal AI agent that will be available to billions of users worldwide. Each user will receive a dedicated cloud computer that will operate even in autonomous mode. Plans call for Muse to be integrated with Meta’s smart glasses, allowing users to interact with agents without even noticing. Although the carbon footprint of personal agent use is still relatively small compared with that of other activities, such as air travel or meat consumption, mass adoption could lead to a substantial increase in energy consumption. This explains the scale of data centre construction, including projects such as Hyperion in Louisiana, which will be powered by 10 gas-fired power stations.
Experts predict that the technologies trained in the data centres now under construction will emerge in three to five years and will differ significantly from today’s chatbots. Small modular nuclear reactors (SMRs) are also being considered as a power source for data centres. They could provide carbon-free energy, but their commercial deployment in the US is still at an early stage. Despite government efforts to accelerate the industry’s development, bringing SMRs to market remains a lengthy process.
Source: WIRED
