AINEWS
FinOps for AI needs token-native KPIs, not cloud-era cost levers
98% of FinOps teams now manage AI spend, but the meters built for reserved instances and rightsized VMs cannot read token billing.
98% of FinOps teams now manage AI spend, but the meters built for reserved instances and rightsized VMs cannot read token billing.
Ninety-eight percent of FinOps practitioners now manage AI spending in their organizations, up from 31% two years ago, according to the FinOps Foundation's State of FinOps 2026 Report, as reported by SiliconANGLE. The adoption curve is steep. The measurement layer underneath it is not keeping pace.
The cost levers that defined cloud-era FinOps, including tagging, rightsizing, and reserved capacity commitments, were built for a billing model that priced compute in hours and storage in gigabytes. Token-based AI services price in units that have no clean mapping to either. A single LLM call can burn thousands of tokens before a request returns, and the cost driver is the model's behavior on a prompt, not the duration of a virtual machine. Trying to govern token spend with VM-era tooling is, in practice, governing the wrong line item.
Victoria Levy, a senior staff FinOps analyst at SailPoint Technologies Inc., framed the gap this way to SiliconANGLE and theCUBE at FinOps X 2026: "The KPIs are going to be way different. We have tokens, so people are going to come up with your cost per token, maybe tokens per — whatever other business driver there is out there." The candidates she named, including cost per token and tokens per business driver, are not extensions of the old dashboard. They measure the cost of producing an outcome, not the cost of running an instance. With a single named practitioner in the reporting, the framing is a clear industry signal rather than a settled consensus.
That shift has operational consequences. Recommendations, the advice-style output that FinOps platforms have historically produced, decay quickly in a token environment because the cost surface changes with every model update, every prompt revision, and every change in provider pricing. Without automated enforcement, the recommendation is a memo. With it, the recommendation is a guardrail. The choice between memo and guardrail is the difference between a cost program that drifts and one that holds.
"If you don't have automation and you tell people to do some of the best practices and rightsize, that's only a one-time thing," Levy noted. "You need to implement that enforcement to make sure that the work you've already done stays there and then you can build on it and go do different things."
The cross-functional piece matters too. Token costs land in product, engineering, and finance simultaneously, and the unit economics only make sense if those teams are reading the same number. A cost-per-token metric that finance tracks but engineering does not see is a report, not a control. The work FinOps teams are now being asked to do is less about reducing a cloud bill and more about building a shared instrument that multiple functions can act on, where the success metric is a business outcome priced in tokens, not a workload priced in hours.
Two things to watch next. First, whether the FinOps Foundation's reported KPI taxonomy moves from cost-per-token toward outcome-based metrics, since the latter is harder to standardize but closer to the value the business actually buys. Second, whether the major cloud and AI platforms expose the billing granularity that any of these new KPIs will require, because the best FinOps framework in the world cannot enforce a metric it cannot measure. As of this week, the gap between AI adoption and AI accountability is real, the adoption number is large, and the instruments are still being built.