Token Economics: A Capacity Planning Framework for Enterprise GenAI Cost Governance

Abstract

Enterprise generative AI cost is hard to predict because token consumption does not behave like ordinary infrastructure demand. A token is the unit a model provider bills on, a chunk of text roughly three quarters of a word long. Token use rises with how complicated a prompt is, which model was picked, and how people end up using the feature. It does not rise neatly with headcount or transaction volume. So finance ends up explaining invoices after they arrive instead of forecasting spend before it happens. This paper treats tokens as a capacity resource rather than a line-item expense. It borrows three practices that are decades old in compute and storage capacity planning and almost entirely missing from token cost management: modeling demand by type of work rather than by team, measuring what normal consumption looks like for each type and raising an alert when it moves, and treating the number of steps a model takes as something to forecast rather than something to accept. The framework sets out a six-step cycle that runs from sorting work into classes through to reconciling the forecast against the invoice, an instrumentation layer defining what every call record has to carry, and a control model that separates three kinds of limit organizations routinely confuse: a safety cap that blocks, a soft threshold that alerts, and a spending ceiling that escalates. No threshold numbers are published, because every threshold has to be set against a baseline measured inside the organization deploying it. The framework has not been deployed, instrumented or measured. This is a design contribution, not an empirical one.