There is a gap opening up between how companies budget for AI and what AI actually costs when it leaves the pilot stage and runs for real. Google's 2026 State of AI Infrastructure report, drawn from 1,400 senior IT leaders globally, puts numbers on the problem. Eight in ten organizations say they need to upgrade their existing systems to support AI agents at production scale.
The reason is structural. When companies ran AI chatbots and assistants, each user request triggered a single processing call. AI agents work differently. They plan, retrieve information, check their own outputs, call external tools, and sometimes start over when the first attempt falls short. A single task can trigger 10 to 20 processing calls. Gartner's analysis puts the token consumption of agentic AI at 5 to 30 times higher per task than a standard AI tool. Per-token prices are dropping, and dropping fast, but the volume of tokens consumed is rising faster.
This creates what analysts are calling the inference paradox: the cost per unit of AI processing keeps falling, yet enterprise bills keep rising. Inference refers to the act of running an AI model to produce an answer or take an action. It now accounts for nearly half of all AI computing workloads, up from 28% that went to training new models. One analysis puts inference at 85% of the total enterprise AI budget in 2026. The companies that built those budgets a year ago were not modeling this.
Uber is the clearest example on record. The company exhausted its entire 2026 AI budget by April, after its AI coding tool spread rapidly across 5,000 engineers. Monthly costs for heavy users ran between $500 and $2,000 per person. The CTO confirmed the overrun and said the company was back to the drawing board on its planning assumptions. The tools delivered genuine productivity. That was not the problem. The problem was that usage-based pricing, which charges by consumption rather than by seat or licence, does not behave like any software budget that finance teams know how to model. By early June, Uber had introduced per-employee monthly caps.
That gap, between pilot economics and production economics, is now a documented pattern across industries. Budget overruns of 300 to 400% in the first year of production are appearing in multiple reports. Companies whose AI agents run in the background continuously, monitoring, scanning, and processing, face a further complication: those workloads cannot be turned off without losing the business value they provide.
Ninety-six percent of senior IT leaders in the Google report said cost efficiency is now the primary factor guiding AI infrastructure decisions. And 91% said they factor power consumption into hardware choices. Energy has moved from a technical concern to a boardroom one, because running agents at scale requires a lot of it.
For organizations outside the United States, a second pressure layer is building. Google's report projects that 75% of non-US enterprises will have a formal digital sovereignty strategy by 2030. Sovereignty here means deciding which AI workloads can run on major US-owned cloud platforms and which must stay within national or regional boundaries. It is not just a technical choice: storing data with a US-headquartered cloud provider, even in a European data centre, still exposes that data to US federal access laws regardless of where the servers physically sit. The EU AI Act, which reached full enforcement for high-risk systems in August 2026, adds documentation and audit requirements that make sovereignty planning harder to defer.
The strategic reading here is not that AI agents are too expensive to use. The Uber story is actually about genuine productivity gains that outpaced the ability to govern costs. The question every leadership team should be asking is whether their finance and procurement functions have the tools to manage AI as a variable operating cost rather than a predictable licence fee. Most do not yet. Setting per-user caps, routing simpler tasks to cheaper models, and tracking consumption in real time are operational disciplines that most organizations have not built. The companies that build them first will get the productivity benefits without the budget shocks.