By [Author Name] | Technology & Energy Correspondent
Ask a chatbot a question and it answers once, then stops. Hand the same job to an AI agent and it plans, clicks, calls tools, checks its own work and keeps going until the task is finished — sometimes for hours. That difference is barely visible on a screen. Inside a data centre, it is enormous.
It explains why Silicon Valley's attention has drifted from the chat window to the agent, and why the industry's next hard limit may not be chips, talent or data. It may simply be electricity.
The Shift: From Answers You Read to Agents That Act
A chatbot query is a single burst of computation. You type, the model predicts, it replies. The machine is done in seconds.
An agent works differently. It breaks a goal into steps, decides which tool to use, runs that step, reads the result, corrects itself if something fails, and repeats. Every one of those steps is a separate model call, and every call draws power.
That is the core of the story: agentic AI does not just change what AI can do. It changes how long the hardware has to stay switched on.
Why an Agent's Electricity Bill Looks Nothing Like a Chatbot's
Data centres were already power-hungry before generative AI arrived. What agents change is the shape of the demand, not just the size of it.
A burst of chatbot traffic creates spikes. Agents create long, sustained loads — a machine that is effectively working a shift rather than answering a question. Sustained load is harder on grids, harder to cool, and more expensive to run.
For operators, that converts a software story into an infrastructure story: land, transformers, substations, cooling systems and water. None of these can be summoned quickly.
How Silicon Valley Got Here: A Buildout Years in the Making
The current expansion did not begin with agents. It began with cloud computing, accelerated through the first generative AI wave, and is now being reinforced by the expectation that agents will run continuously in the background of ordinary software.
Infrastructure moves on a slower clock than software. Grid connections, power purchase agreements, cooling retrofits and new facilities are planned in years, not quarters. Interconnection queues in many markets have historically stretched longer than the construction itself.
That mismatch — fast software ambitions, slow physical delivery — is where most of the tension sits.
Who Pays for the Compute Boom — and Who Feels It First
The people who notice first are often those living near data centre clusters. They see changed land use, higher water demand for cooling, construction traffic, noise, and questions about what it does to local power costs.
Ratepayers in some markets have already been part of public debates over who should fund grid upgrades built for very large industrial customers. Those debates are local, but they are spreading.
Then there are the workers. Demand for power engineers, thermal specialists, electricians and grid planners is rising faster than the pipeline of people trained for those roles — a skills gap that quietly shapes how fast any of this can be built.
What the Industry and Grid Operators Are Saying
To be clear about sourcing: the material behind this story does not include on-record statements from companies, utilities or regulators. No quotes have been attributed, and none have been invented.
What can be said is structural rather than quoted. Utilities typically plan capacity years ahead and must justify large new loads to regulators. Data centre operators, meanwhile, tend to secure long-term power arrangements before breaking ground, because an unpowered facility is a stranded asset.
Regulators in several markets have begun asking harder questions about incentives, grid cost allocation and water use — a signal that AI infrastructure is now a public policy topic, not only a technology one.
The Real Cost Isn't the Chip — It's the Watt
There is a comforting assumption that efficiency will solve this. Newer models often do produce the same output with less energy per token, and that progress is real.
But agents multiply tokens. A task that once needed one answer now needs dozens of internal steps, many of them invisible to the user. Efficiency can improve and total consumption can still rise — a rebound effect familiar from every previous wave of computing.
The practical consequence: measuring energy per query is becoming an outdated metric. The more useful unit is energy per completed task.
Confirmed Facts vs What Still Isn't Clear
Reasonably well established: the industry's focus is moving toward agentic AI; agents execute multi-step, longer-running workloads; those workloads raise compute demand per task; data centre capacity is being expanded to meet projected demand.
Genuinely unclear: how much energy a typical agent task consumes across different systems; how much of that growth is offset by efficiency gains; how much of the buildout will be used versus left idle; and how the cost burden will be split between operators, shareholders and ordinary electricity customers.
Speculation — not fact: claims that agents will inevitably drive a permanent step-change in national power demand, or that efficiency will inevitably neutralise it. Both are projections. Neither is settled.
Why the Compute Layer Is Hard to Dislodge
If this were only about model quality, the advantage would be fragile. It isn't only about that.
The companies leading this buildout hold something harder to copy: secured land, queued grid connections, long-term power contracts, custom silicon, cooling expertise, and developer ecosystems that make their platforms the default place to build.
In simple terms, the moat is not the model. It is everything the model needs to run — and how difficult that is to assemble from scratch on a deadline.
The Case Against the Boom Narrative
There is a credible bear case, and it deserves space.
If agents deliver less value than promised, or if customers resist paying for long-running autonomous tasks, demand could soften and some capacity could sit underused. Historically, infrastructure booms have produced overbuilds as often as shortages.
Add to that rising community opposition, water stress in dry regions, tighter regulation, and the possibility that model efficiency improves faster than expected. Any one of these could change the arithmetic that today's construction plans are based on.
None of this means the buildout is wrong. It means it is a bet, not a certainty.
From Search Boxes to Software That Never Sleeps
Step back and the wider pattern is clear: computing is shifting from humans asking, to machines acting.
For thirty years, software waited for a click. Agentic systems initiate work on their own schedule, across services, in the background. That is a different kind of internet — one whose electricity demand is set by machines rather than by how many people are awake.
Energy planning, data centre siting and even urban policy will have to adjust to that assumption.
What This Means If You're a Student, Investor, or Job Seeker
Students: the scarce skills here are unfashionable ones — power systems, power electronics, thermal engineering, grid planning — paired with machine learning infrastructure. That combination is unusually valuable right now.
Investors: the useful question is no longer only which model performs best, but who controls power, land and interconnection rights. Those determine who can actually deploy at scale.
Job seekers: the buildout creates demand well beyond AI research — electrical contracting, cooling maintenance, facility operations and compliance roles among them.
Residents near new sites: tariff proceedings and local consultations are where cost allocation is decided. They are public, and they are where individual input has the most leverage.
What to Watch Next
The most informative signals won't be product launches. They will be grid connection approvals, utility capital spending plans, cooling technology choices, and whether the industry moves toward reporting energy use per task rather than per prompt.
Watch, too, how agentic AI is priced. If autonomous tasks are billed by duration or by completed outcome, that will tell you how the industry itself expects the economics to work.
Our Take
The story of agentic AI is usually told as a software story — smarter models, better tools, more capable assistants. But the constraint that will decide how fast any of it reaches ordinary people is physical: watts, water and wires.
The honest position is that both sides of the argument are early. The demand growth is real and visible in construction. The offsetting efficiency is also real and improving. What the industry lacks is transparent measurement — and without it, everyone from regulators to ratepayers is arguing over projections rather than evidence.
That is the part worth watching. Whoever first measures energy per completed task credibly will do more for this debate than another benchmark score.
Frequently Asked Questions
What is agentic AI, in simple terms?
Agentic AI refers to systems that don't just answer a question but work toward a goal on their own — planning steps, using tools, checking results and continuing until the task is done. A chatbot replies; an agent acts.
Why do AI agents use more power than chatbots?
Because a single chatbot reply is one computation, while an agent may run dozens of computations in sequence for one task, often over minutes or hours. More computation, sustained for longer, means more electricity and more cooling.
Could this raise electricity bills for ordinary households?
It depends on the market. Where large data centre loads connect to the same grid as homes, the cost of new transmission and generation becomes part of a public regulatory process. Outcomes differ by region, and there is no single global answer.
Will energy demand keep rising even as AI gets more efficient?
Possibly, and it's debated. Efficiency reduces energy per computation, but agentic workloads perform many more computations. Whether the second effect outpaces the first is an open empirical question, not a settled one.
