Organizations spending on AI have told us what's broken, and it isn't the price of a token. In the Tokenomics Foundation's September 2026 State of Tokenomics survey, 43% of 472 respondents named proving value or ROI among their biggest challenges — the most common answer — while only 7% named cost or pricing complexity. When asked what they want from providers, 23% asked for more transparency or granular data and 4% asked for lower prices.
That gap is the story of AI FinOps in late 2026. The bill is arriving and growing, and most organizations aren't confident they can connect it to an outcome their CFO would accept.
Where AI FinOps stands today
AI is no longer a side topic for FinOps teams. By the FinOps Foundation's State of FinOps 2026 count, 98% of respondents now manage AI spend, up from 63% a year earlier and 31% the year before that. AI cost management is now the skill those teams say they need most. One practitioner put the state of the art plainly: "Is your AI providing value? No one can answer that question yet."
The standards work has started. The FinOps Foundation's J.R. Storment framed the problem in May with "Token Economics: The Atomic Unit of AI Value," which defines tokenomics as "the discipline through which AI consumption is metered, attributed, and connected to business outcomes". On August 4, the Linux Foundation launched the Tokenomics Foundation, with Storment as executive director and 29 founding members named in the launch release, from IBM, Oracle, and SAP to JPMorganChase, Flexera, and SHI. Its stated goal is "open, vendor-neutral industry standards, benchmarks and best practices for the economics of AI." The roadmap includes cost to serve "expressed as cost per call rather than cost per token" and value measurement "starting with the share of work completed without human involvement, measured against what the process costs today".
The billing data standard is moving too, more slowly. FOCUS, the open cost and usage spec, added columns in version 1.2 for SaaS virtual currencies such as credits and tokens. Version 1.4, ratified in June, centered on invoice reconciliation and commitment details. Worked examples for token-based AI billing, including cached tokens bought through a cloud marketplace, were approved for the upcoming 1.5 release on October 20. The Tokenomics Foundation's token cost telemetry is slated for FOCUS 1.5 and beyond.
Linux Foundation executive director Jim Zemlin put the gap plainly at the launch: "there is no shared way to connect AI spending to value".
What's broken
1. There is no single unit to attribute
Pricing complexity ranks low as a stated challenge, but it's a big part of why the bill is hard to attribute. The big model providers' bills stack modifiers on a base rate. On one provider's current price list, cached input reads bill at a fraction of the base rate, cache writes bill at a premium, batch jobs run at half price, and a US-only inference option on newer models carries a surcharge — and the modifiers stack. Another provider sells the same model at several service tiers with separate long-context and cached-input prices. On OpenAI's reasoning models, reasoning tokens you never see are still "billed as output tokens". On Azure, provisioned throughput is billed by the hour "regardless of the number of tokens consumed," with reservations to discount it.
Storment's line is the right frame: cost per token "is a property of a configuration," not of a model. Flexera's Jay Litkey, senior vice president of cloud and FinOps, quoted in the launch release, went further: "That's not a maturity problem, it's a measurement problem".
2. Agents break the forecast
In my experience, chat workloads were predictable enough to budget. Agentic workloads aren't. Anthropic reported in 2025 that its agents typically used about 4x the tokens of a chat interaction, and multi-agent systems about 15x. On one benchmark problem, a single run consumed 40.5 million tokens — roughly 38 times the median. In a survey commissioned by AI cost-management vendor Mavvrik, nearly one in four companies missed their AI cost forecasts by more than 50%. Gartner expects more than 40% of agentic AI projects to be canceled by the end of 2027, citing escalating costs, unclear business value, or inadequate risk controls.
Storment's summary: "A token may get cheaper; tokens, in aggregate, are not."
3. The meter is hiding inside your SaaS seats
As Storment notes, a growing share of AI cost is embedded in seat licenses and SaaS tools that buy tokens on your behalf. "Some SaaS providers have begun exposing token-level usage; many have not," he writes. "The seat fee is the floor; the variable component is what drives total spend." The commercial terms are moving under buyers' feet. In April, The Register reported that Anthropic had dropped bundled usage from new Enterprise seat contracts, with legacy seats moving over at renewal; one source it quoted said every token is now billed at API rates on top of the seat.
4. Nobody owns it
In State of Tokenomics, 12% of organizations said no one owns AI economics, and not one of them could connect AI spend to an outcome their CFO would accept. Finance owns it in only 5%. AI spend is spread across the P&L, and "no single budget line captures it".
5. Value has no definition yet
This is the root problem. Without an agreed unit of value, cost per token is a number with nothing to divide by. Nearly three in four State of Tokenomics respondents were either not confident or only slightly confident they could link AI spend to an outcome their CFO would accept. The Tokenomics Foundation's first proposed value measure — the share of work completed without human involvement, against what the process costs today — is a start, not an answer.
One caution before anyone celebrates rising usage. As The Decoder argued about Google's token milestone, headline token counts mostly measure compute effort, not value. More tokens can mean more work done, or more retries, longer prompts, and bigger context windows.
What I learned metering my own agents
I've been running Claude Code agents all year to build and operate my own projects, so I had to solve a small version of this problem for myself. Every job I hand to an agent carries a budget cap. There's a daily spending cap across all of them, and every job writes an entry to an append-only ledger with its reported cost and whether I approved the result.
Three lessons from that ledger line up with the industry findings above.
The meter and the bill don't match. My agents report an API-equivalent cost per job, but I pay a subscription. Both numbers are true, and they answer different questions. If a one-person operation needs two cost views, an enterprise buying through seats, cloud platforms, and commitments needs more than that.
The ledger only sees what's routed through it. Jobs I dispatch are metered to the cent. Interactive sessions where I work with an agent directly don't show up in the ledger at all. That's shadow AI at the scale of one person — the same kind of blind spot Storment describes.
Cost per job means nothing without a decision on value. The useful number isn't cost per job. It's cost per job I accepted. Recording the approval next to the cost is what turns a spend log into something that resembles unit economics.
None of this needed new tooling. It took deciding, before the first job ran, what would be measured and who signs off.
Where I see the opening
I've watched this cycle before. Over 30 years in IT, I've seen on-prem licensing turn into subscription SaaS, and subscription turn into consumption cloud. And now we see the explosion of AI Token Costs! As I've seen it, each shift broke the existing cost model first, and the tooling caught up later: a confusing bill, then tagging and allocation, then a standard billing format, then reconciliation.
AI spend is following that pattern through more buying channels at once. Tokens are bought from model providers, through cloud platforms such as Bedrock and Vertex, and inside SaaS tools. In State of Tokenomics, half of AI consumers buy through four or more of six channels. The FOCUS working group's new examples already include cached tokens bought through a cloud marketplace. Each hop adds a pricing layer and removes a little visibility.
That's where I think the real work sits for the next two years:
- Attribution across channels. Connecting a token bought through a cloud platform, a marketplace, or a SaaS tool back to the team, product, and outcome that consumed it.
- Cost per outcome, not cost per token. The Tokenomics Foundation is right to push toward cost per call and value metrics. Every organization will still have to define its own outcome units.
- Ownership between finance and engineering. In the survey, none of the organizations without an owner could show a CFO outcome.
The tooling pieces exist in silos today. AWS supports cost-allocation tags on Bedrock model usage. Open-source gateways such as LiteLLM track spend and budgets per key and team. FinOps platforms such as Vantage ingest model provider bills. What's missing is the layer that ties them to value.
What to do this quarter
- Name an owner for AI economics. Make it someone who answers to both finance and engineering.
- Inventory where tokens come from. That means direct API, cloud platforms and marketplaces, and SaaS seats. Ask each SaaS vendor whether they expose token-level usage.
- Put a budget cap on every agent workload before it runs, and a daily cap above that.
- Log cost next to acceptance. A job's cost only means something next to whether its output was used.
- Map your billing data to FOCUS now, so the AI work planned for FOCUS 1.5 and later lands on a foundation you already have.
The standards bodies have named the problem correctly. It's a measurement problem, and the organizations that start measuring now will have the baseline when the standards arrive.
If you're working on AI cost attribution, I want to compare notes: what's your unit of value? Follow me on LinkedIn for the next piece in this series.
— Todd Carter, CISSP
Sources: FinOps Foundation, Token Economics: The Atomic Unit of AI Value; Linux Foundation, Tokenomics Foundation launch; Tokenomics Foundation, State of Tokenomics; FinOps Foundation, State of FinOps 2026. Other sources are linked where cited.
Comments
Post a Comment