AI Agents Could Use 100x More Energy Than

An AI agent is a shopping list, a backup store when the first one's out of stock, a verification pass at the register, and sometimes a full return-and-rebuy cycle when the order comes home wrong.

That's the gap a new report, "AI Agents Are Thirsty for Power," tries to quantify, and it's why the conversation about AI's energy footprint is shifting. According to Wired, the report connects the rise of agentic workloads, systems that plan, call tools and run multistep tasks, to continued data-center construction. The infrastructure challenge is no longer just training large models once; it's serving repeated inference loops at scale, every hour, for everyone using an agent.

Why One Prompt Becomes Fifty

If you've watched an agent work (Cursor grinding through a refactor, a research assistant running search after search), you've seen the mechanism. One user request fans out into:

a planning step, where the model decides what to do
several tool calls: web searches, code execution, database lookups
calls to other models, maybe a big one for reasoning and a small one for formatting
verification passes, where the agent checks its own work
retries when something fails, because agents fail a lot

Each step is its own inference request. Digital Trends reports that research on this found agentic AI can be over a hundred times worse on energy than a plain chatbot query. Gizmodo's framing was blunter: when it comes to energy use, agents could make chatbots look like pocket calculators.

The hundredfold figure is a multiplier ceiling, not a prediction. It assumes a task where the agent runs many rounds of reasoning and tool use. Ask an agent for a one-line answer and the multiplier collapses. Ask it to debug a codebase for twenty minutes and it balloons. The number describes the shape of the risk, not a utility bill you can expect next month.

What One Scientist Found When He Counted

The aggregate forecasts are one way to see this. The other is to just... log it. Fast Company profiled a scientist who tracked the energy behind every prompt he sent while using agentic tools, treating his own AI usage like a electricity meter reading. The punchline of that experiment is structural: the individual prompts felt trivial, but the volume of calls underneath a single "task" is what adds up. You don't notice fifty inferences when they're wrapped in one progress bar.

That's the measurement problem in miniature. We've spent three years talking about AI energy use in tokens and queries, because that's what we could see. Agents hide the unit of consumption. Your meter is now the task, and the task is opaque.

Where the Forecasts Go Wrong (in Both Directions)

Here's where I'd pump the brakes on anyone selling you a firm number, in either direction.

Forecasts built on peak capacity can overstate near-term electricity use. A data center wired for 500 megawatts doesn't draw 500 megawatts on day one; chips arrive over years, and utilization (how hard the hardware actually runs) starts low. Utilities and grid planners have seen this movie with crypto mining and with early cloud buildouts: announced capacity is a ceiling, not a load.

At the same time, ignoring persistent inference growth understates the real constraints. Training is a project with an end date. Inference is a subscription. If agent adoption compounds, the load doesn't just grow, it changes character: it's continuous, geographically concentrated wherever the data centers are, and it comes with cooling and water demands that transmission upgrades alone don't solve. The brief here is blunt on this point and I agree with it: the actual effect depends on adoption rates, model efficiency, hardware utilization, and how much work gets routed to smaller models. None of those variables has a locked-in value.

And the efficiency story cuts both ways. Model efficiency has improved enormously; the cost per token of a given capability keeps dropping. But Jevons' paradox (the old economics observation that cheaper consumption often increases total consumption) is the ghost at this table. If agents make AI-driven tasks cheap enough, the number of tasks explodes. Efficiency per unit, eaten by volume.

The Questions the Next Debate Will Fight Over

The report, per Wired, points toward a debate that connects model design to physical infrastructure, and the battlegrounds are already visible:

Power procurement. Where does the electricity come from? Tech companies have been signing nuclear deals, renewable PPAs and (controversially) gas deals to feed data centers. Agents raise the stakes on every one of those contracts.

Water. Cooling is thirsty, and data centers increasingly compete with agriculture and cities for it. This fight has already played out in Arizona, Chile and the Netherlands over existing facilities; agent-scale inference adds pressure to every siting decision.

Transmission and siting. The best renewable resources are rarely next to the best fiber. Interconnection queues for new grid capacity run years in the US, and that's now a direct constraint on how fast AI capacity can come online.

Where the metrics live. Gizmodo's calculator comparison is memorable because it reframes the unit. The reporting suggests efficiency gains will need to be measured against total task completion, not tokens alone. That's a hard engineering problem: how do you benchmark energy per completed, correct task when tasks vary wildly in difficulty? Nobody has a standard. Whoever sets one shapes what "efficient AI" means for the rest of the decade.

What I Can't Tell You Yet

I try to be honest about the holes, and this story has big ones. The hundredfold figure comes from specific research setups; I don't have independent replication across different agent frameworks, and the sources here don't quantify real-world adoption curves. The Wired report connects agentic AI to data-center construction, but attribution is tricky: data centers serve many workloads, and separating the agent-driven slice from everything else isn't yet possible from public data.

So treat the range, not the point estimate, as the finding. The floor is that agents multiply inference per user action by a factor of at least several, often. The ceiling is a hundred or more for hard tasks. Where between those the world lands depends on choices that haven't been made yet: how aggressively agents get deployed, how well small models handle subtasks, how much effort goes into energy-aware orchestration (routing easy steps to cheap models, caching aggressively, killing retry loops).

That last category is where I'm watching for the next story. The companies building agents have both the incentive and the telemetry to optimize energy per task; it's on their electricity bill. If orchestration efficiency becomes a competitive metric the way cost per token is now, the scary multiplier shrinks. If it doesn't, the grid planners inherit the problem.

Either way, the era of evaluating AI systems by what they say is ending. The next evaluation is what they do, step by step, in kilowatt-hours, and whether the physical grid can tell the difference between a chatbot and an agent that just looks like one.

By Yuki Okonkwo