The Tokenomics Foundation: Linux Foundation's new push to standardize how AI costs are measured
Almost every FinOps team now manages AI spending, and the Linux Foundation has an answer ready: a standardization organization for token usage. But the launch date is already unclear, and the framework is so far known mostly through secondary sources. Here is what we know — and what we don't.
The news: a standards body for token economics
A new foundation is set to standardize how organizations measure and understand what AI actually costs. The Tokenomics Foundation is a project under the Linux Foundation and works side by side with the FinOps Foundation — the organization that did the same job for cloud spending. Its members include major AI buyers such as banks and retailers, along with cloud providers and model providers, according to InfoWorld (ec7f43b8).
The foundation is described as a standards organization, not a tool vendor. It does not sell a product and does not deliver a cost-tracking tool. Instead, its working groups are to publish open, vendor-neutral definitions, frameworks, cost data standards building on the FOCUS standard for cloud costs, as well as benchmarks and certification (ec7f43b8).
A note on the launch date: the sources conflict. InfoWorld says August 2026, while a Cisco page says June 2026 (3a965f54). None of the available sources is the foundation's own announcement, so the exact date cannot be verified. What the foundation does, however, is consistently described across sources.
Why now: AI spending has become routine financial work
The need shows up in the numbers. According to the State of FinOps 2026 report, 98 percent of respondents say they are handling AI spending — up from 31 percent two years earlier (2613e3c5). In under two years, AI costs went from a niche concern to a core task for finance and operations teams.
The growth in token consumption itself is just as striking. A survey published on arXiv (Chen et al.) shows that weekly token volume on the OpenRouter platform grew from 0.4 trillion tokens in December 2024 to 27.0 trillion in March 2026 — an increase of roughly 68 times over 15 months (d93c7ab6).
Then there are agents. Cisco states, citing Gartner, that agent-based models require 5 to 30 times more tokens per task than a standard generative AI chatbot (3a965f54). This figure should be read with caution: it is cited second-hand, without access to Gartner's underlying report. But the direction is clear — where a chatbot uses one conversation, an agent makes many model calls per task, and the cost multiplies accordingly.
The problem Salesforce ran into
The most concrete illustration of what happens when token usage goes unmanaged comes from Salesforce. Business Insider reported on Tuesday, November 17, that it had seen an internal memo from COO Miguel Milano introducing monthly caps on employees' AI use. According to the reporting, some employees had spent several thousand dollars a month on tokens in the previous quarter. The caps are set at $1,000 per month for technical sales roles and $100 for non-technical sales roles, with rollout the same month (282fd024).
The same reporting says Salesforce expects to spend around $300 million on Anthropic tokens in 2026, a large share of it on coding work. This is the company's own expectation, relayed via Business Insider through Startup Fortune. Neither the memo nor the figures are independently confirmed, and this entire portion of the story rests on secondary sources. But it shows the mechanism: when individuals get direct access to powerful models, costs can grow faster than the controls.
Why direct model APIs are the hardest to manage
The FinOps Foundation's working group identifies direct purchases from model providers' APIs as the most difficult AI channel to manage (2613e3c5). The reason is structural:
- Developers buy directly. The purchase often happens outside IT's procurement routines, and the bill shows up as an abstract API invoice.
- Billing lacks business context. An invoice line tells you how many tokens were used, but not which team, product, or customer caused them.
- Allocation mechanisms are thin. Without tags and keys across calls, it is hard to allocate costs where they belong.
- Pricing varies widely. Model tiers, features, and price per token diverge across providers and models, making comparison and budgeting difficult.
This is the same kind of problem the FinOps movement tackled for cloud costs years ago — and the reason the Tokenomics Foundation is placed alongside the FinOps Foundation under the Linux Foundation's technology value umbrella, where the foundation is to lead work on AI's total cost and how value is measured (2613e3c5).
The frameworks: a cost map and a growth notation
Two of the foundation's proposed tools are worth a closer look — with one important caveat: the descriptions come from an InfoWorld article written in a column style, and the presentation is the author's interpretation, not verified primary documentation from the foundation.
The seven-layer stack is a cost map of everything that shapes the price of an AI response, from electricity at the bottom to the user's request at the top. The point, as InfoWorld presents it, is simple: the map shows where to look before you start cutting (ec7f43b8). It is a counterweight to the most common mistake in cost-cutting: cutting in the wrong place — for example, switching to a cheaper model when the real problem is how many calls an agent makes per task, or the context sent with each call.
Big-T notation is borrowed directly from the programming world. Where Big-O describes how slow code becomes as data volume grows, Big-T describes how quickly token usage grows as more people use the application — from constant to unbounded growth (ec7f43b8). The idea is to force an answer to a question rarely asked at launch: what happens to the bill when the user base grows tenfold?
Both frameworks are, for now, more direction than standard. None of the sources in this story provides access to the foundation's own documents, and a reader who wants to use them today should treat them as a thinking tool rather than an established specification.
What the industry already does: model routers
Some practice is growing independently of the standardization work. In the State of Tokenomics survey from September 2026, 86 percent of respondents say they are evaluating or using a model router, with OpenRouter and LiteLLM the most adopted (2613e3c5).
A model router sends different tasks to different models — cheap models for simple tasks, pricier ones for demanding ones. It is in effect a cost-control mechanism built into the infrastructure, and its prevalence suggests many organizations already find that one model for all tasks is an expensive solution. It is also worth noting that OpenRouter — the most adopted router — is the same platform delivering the most dramatic token-consumption growth figures. Routers can both dampen and concentrate cost growth, depending on how they are configured.
What practical value might look like
For teams buying AI directly from model providers, the available material suggests some concrete moves:
- Map the cost before you cut. The seven-layer stack's logic is to find the right layer — energy, inference infrastructure, model choice, context length, number of calls, agent design, user behavior — before decisions are made.
- Put numbers on growth. Big-T thinking forces an estimate of how token usage scales with usage, as the Salesforce case shows the consequence of not doing.
- Manage the channel problem. Because direct API purchases lack business context in billing, tagging, keys, and internal caps — like those Salesforce is introducing — are in practice the available controls today.
- Consider routers. With 86 percent of respondents evaluating or using them, model routers have become a mainstream part of the cost architecture.
The open questions
Several things remain to be verified. The launch date is disputed by the sources, and no primary announcement is available. The Salesforce figures — the caps, the thousands of dollars in monthly spending, and the roughly $300 million for Anthropic — rest on secondary reporting of an internal memo. The frameworks are known through an interpretive article, not the foundation's own documents. The Gartner figure of 5–30x token usage is cited second-hand.
And the biggest question is not documentary but practical: will the frameworks actually be adopted? The FinOps movement showed that standardizing cloud costs could succeed, but AI costs move faster and diverge more across providers. Whether the Tokenomics Foundation's definitions, FOCUS extensions, and future certification gain the same traction — and what certification would actually mean — remains unanswered. For now, the organization, whatever its launch month, is simply too new.

