AI Agents Crave Control

Glad to have you back to Power Play! Every week, senior writer Molly Taft dives into a topic surrounding this midterm season’s prominent issue: data centers. If you have a question or insight for the column, don’t hesitate to reach out to Molly via email at [email protected] or connect securely on Signal at mollytaft.76.

“What in the world are they constructing all these data centers for?” an incredulous friend asked me not long ago.

They’re not alone in their curiosity: We received numerous similar questions during our recent livestream about data centers. It’s a very valid concern. After all, if AI is achieving all these advancements, why are tech firms incurring massive debts and building some of the largest power plants globally for even more data centers?

The purpose isn’t to assist the average user in searching for recipe ideas or finding vacation spots; basic chatbot interactions are becoming an outdated perspective on AI functionality. Today, AI focuses on agents—though there’s no formal definition, agents can be described as systems based on large language models that are designed to autonomously make decisions in order to accomplish tasks—and this transition is a key factor in the power expansion seen in Silicon Valley.

“Instead of posing a straightforward question to an AI chatbot and receiving an answer, these agents can generate hundreds of smaller prompts derived from a user’s initial query,” explains my colleague Maxwell Zeff, who pens the weekly Model Behavior newsletter. “For instance, if a user asked an AI agent to create a website, it might take hours to generate features, re-prompts itself multiple times to develop various web pages, menus, and datasets that support the project.”

Agents have become central to the work being done in frontier labs focused on AI. They are capable of executing some remarkable—and concerning—feats. Recently, OpenAI revealed that a swarm of over 10,000 agents communicating through 2.7 million messages managed to solve a longstanding mathematical issue. (Mathematicians have challenged the company’s assertions.) While this is an anomaly—AI labs are steadfast in their pursuit of resolving supposedly impossible problems and are prepared to invest unusual resources into this goal—all that communication consumed a substantial amount of processing power. This translates to significant energy usage: likely tens of millions of dollars’ worth, according to Max, though pinning down an exact figure is challenging.

Private AI companies have typically been selective about revealing information related to the environmental impact of their products. Many CEOs often reference the resource consumption of individual queries as a benchmark. During a recent podcast, OpenAI CEO Sam Altman remarked that the water used to grow a single almond equated to 38,000 ChatGPT queries. (This calculation has been contested.)

“Individuals consuming 12 almonds at once generally don’t perceive their actions as having a significant negative impact on water resources,” he noted.

Integrating AI agents, which require considerably more energy than simple queries, complicates these assessments significantly. There’s a substantial lack of data regarding the energy consumption of agents, whose functions can vary from basic tasks to an entire day of independent coding involving a team of parallel “helper” agents. The disparity in power usage among these applications is vast—and could expand indefinitely as tasks grow in complexity.

“In other areas of technological growth, our limitations are based on how many people are driving cars or streaming films,” remarks Boris Gamazaychikov, co-founder and CEO of Sustainable AI, a research and advisory organization. “Currently, this technology seems to operate somewhat independently from user involvement—and if you pay attention to AI leaders, that seems to be their objective. They’re discussing unicorns that require just one employee.”