AI agents and rising inference demand

Inference is the compute used to run a finished AI model. A simple chatbot reply is one inference call: a question goes in, an answer comes out. AI agents change that pattern entirely. An agent breaks a goal into steps, calls a model repeatedly to plan, act, check its work, and try again until the task is done.

That shift matters for compute. Where a single answer might take one call, an agent completing a task might take dozens. As agents spread into software development, research, customer support, and operations, the total number of inference calls multiplies far faster than the number of users would suggest.

The key change is structural. With chatbots, demand scaled roughly with how many people were typing. With agents, demand scales with how many tasks are running and how many steps each one takes, a number that can grow even while the human headcount stays flat.

Agents are not yet reliable. The Stanford AI Index notes that AI agents still fail about 1 in 3 attempts. Failure is not free. When an agent gets a step wrong, it often retries, takes a different path, or runs extra checks, and every one of those is another inference call that consumes compute.

So the very thing that makes agents promising, their ability to keep working at a problem until it is solved, also makes them compute hungry. As they improve, more tasks become worth automating, which raises total demand even if each individual task gets cheaper to run.

This is the counterintuitive part. Better, more reliable agents do not necessarily reduce compute demand. By making automation worthwhile for a wider range of tasks, they tend to expand the total amount of work being handed to models.

This is not a forecast for some distant year. The IEA reports that AI-focused data centre electricity surged about 50 percent in 2025, while overall data centre demand grew about 17 percent. Major model providers reported a roughly threefold rise in active users over the past year, according to the IEA.

More users and more agents both point the same direction. Inference, the everyday running of models, is becoming a larger and larger share of why GPU compute stays in short supply. The grid figures are an early, measurable signal of a workload that is still in its early growth.

What makes this shift important is its durability. A single viral product can spike demand and then fade, but agents embedded in software, support, and operations create a baseline of usage that does not switch off. As that baseline rises, it changes how operators plan, because they must provision for steady around-the-clock inference rather than occasional peaks.

Agents are moving out of demos and into real workflows. In software development, they read code, propose changes, run tests, and revise their work across many steps. In research and analysis, they gather sources, draft, and check findings. In customer operations, they handle multi-step requests that once needed a person. Each of these patterns multiplies model calls per task.

What ties them together is autonomy. Instead of a single prompt and reply, an agent runs a loop of plan, act, and check until a goal is met. The IEA reports that AI-focused data centre electricity surged about 50 percent in 2025, and the spread of these multi-step workloads is one of the forces behind that jump.

As more software is built to call models automatically, the line between a user and a workload blurs. A handful of people can set off thousands of agent runs, which is why inference demand is starting to scale with tasks and automation rather than with headcount alone.