Tokens are only a small share of what AI really costs

Forrester has written up the Linux Foundation's analyst summit of 8 October, which put tokenomics, FinOps, IT asset management and the FOCUS billing specification in one room around a single question: how does an enterprise connect what AI costs to the business value it creates?

The term Forrester singles out is TCA, total cost of AI. The claim behind it is that tokens make up only 10% to 25% of total AI spend once the rest of the stack is counted: infrastructure, energy, capital, software and licensing, and the labour needed to produce and consume AI. Total cost of ownership, the long-standing measure, treats AI as an asset to own; TCA treats it as a production-and-consumption system.

The figures come from the State of Tokenomics 2026 study, which covers 472 respondents across 11 industries with $4.6 trillion in combined revenue. Among its findings:

• 96% use frontier models, 87% use cloud providers and 64% use AI embedded in existing products.
• Half operate across at least four of six AI channels.
• 86% are evaluating or already using model routers to match workloads to models instead of sending everything to the most capable option.

The conversation is also moving to finance. Organisers pointed to growing CFO involvement in the Tokenomics Foundation and announced a CFO forum on 8 December under the Chatham House Rule. The practical problem raised was mapping direct and indirect AI costs to the general ledger and chart of accounts, so that month-end allocation stops being manual.

For software buyers, the implication is that per-token pricing is the most visible line and the smallest one. Budgets built on token prices alone understate the bill by a wide margin, and the vendors that can report cost in the units finance already uses will be easier to justify.