An unconstrained agent solving a software engineering problem can cost $5 to $8 in API fees per task. Production sessions routinely run $10 to $100. Multiply by the volume a production deployment actually requires, and you arrive at a number that has very little to do with the pilot that justified the project.

This is the quiet center of Gartner's projection that over 40% of agentic AI projects will be cancelled by 2027. The official causes are escalating costs, unclear business value, and inadequate risk controls. But cost drives the other two. An agent that exhausts its token budget mid-task gets written up as a capability failure. Cost overruns that trigger executive review become governance conversations. The underlying economics never make it into the post-mortem.

The paradox holds together

In Q1 2026, AI captured 80% of global venture funding at $242 billion. Hyperscaler capital expenditure on AI infrastructure is forecast to exceed $450 billion this year. That money is a bet on a category, priced around the probability that agentic AI transforms how work gets done across industries. Category bets have their own economic logic. It was rational for cloud, for mobile, for the railroads.

Individual agent deployments operate on completely different math. A single agent task can trigger planning, tool selection, execution, verification, and response generation, consuming three to ten times more LLM calls than a chatbot interaction. A reasoning loop running ten cycles can burn through fifty times the tokens of a single pass. Context accumulates: conversation history, tool definitions, retrieval documents, telemetry. Each layer adds cost that has nothing to do with the question being asked.

Both sides of this are rational. Four hundred and fifty billion dollars flowing into infrastructure for a category where four in ten deployments won't survive makes complete sense. Two economic logics operating at different altitudes, each internally coherent, with a gap between them where actual projects go to die.

Cost as architecture

The industry's dominant framing treats agent cost the way it treated cloud cost: something you manage after deployment. Build the agent, prove it works, optimize spend later. This sequence made sense for cloud computing because a virtual machine's cost was bounded and predictable before you provisioned it.

Agent economics break this sequence. Every task is an unbounded cost event. One practitioner measuring token usage in production discovered that the prompt itself was a small fraction of total consumption. Most tokens went to the system surrounding it. The cost lives in the architecture that holds the question. The question itself is cheap.

Why pilots don't surface this

Generous vendor credits, small test volumes, flagship models used by default. Pilot environments are specifically designed to prove capability. Unit economics aren't part of the test. The jump to production surfaces costs that were always structural, just hidden by scale.

The money flowing into AI infrastructure will almost certainly build something lasting. Past overbuilds in rail, electrification, and fiber seeded real transformation even as individual projects failed. Though it's worth noting that this particular observation comes from an institutional investor's thesis, which is to say, from the macro altitude where the category logic holds. For teams actually deploying agents, the category was always going to succeed. Whether their deployment's economics could survive contact with production was a separate problem entirely. Right now, the industry surfaces that answer too late.