---
title: Tokenomics: don't count tokens, count what they bought
description: Tokenomics is AI economics, the discipline of turning energy and capital into AI and then into measurable business value. Notes on J.R. Storment's five-minute AGNTCon keynote for the new Tokenomics Foundation: why a token count is the wrong metric, what the bill really includes, and the gap between seeing the spend and showing the value to the CFO.
date: 2026-10-24
tags: AI economics, tokenomics, LLM costs, FinOps, AGNTCon
---

Tokenomics is AI economics: the discipline of turning energy and capital into AI, then using that AI well enough to show business value. It's not the crypto kind. Its first rule sounds odd for something named after tokens: **don't count tokens**. Count what they bought.

That's the pitch from J.R. Storment, executive director of the new [Tokenomics Foundation](https://www.tokeneconomics.com) at the Linux Foundation, in a [five-minute keynote at AGNTCon + MCPCon](https://youtu.be/iNV5-YkPHW4?t=2381) on October 22. He opened with two questions for the room. How many of you know what your agents are costing you? A few hands. How many can tie that to value and ROI? One, at the back.

## Energy in, value out

His slide gives the definition: take energy and capital, put them through chips to create intelligence, consume that intelligence efficiently, and drive measurable business value at the other end. His shorter version: energy creating intelligence driving value.

The definition is still a draft. The Linux Foundation launched the Tokenomics Foundation on August 4 to write open, vendor-neutral standards, benchmarks and best practices for the economics of AI, and it works closely with the FinOps Foundation.

<figure class="fd-fig tk-chain" aria-labelledby="tk1-cap" data-desc="The Tokenomics Foundation's definition as three steps, with a marker that walks through them. Energy and capital go in. They pass through chips and come out as intelligence. That intelligence is consumed efficiently, and what comes out of that is measurable business value. His one-line version, shown as the big label: energy creating intelligence driving value. A strip under it says the definition is still a draft, version 0.5.2, dated September 1, 2026.">
<p class="tk-big">Energy creating intelligence driving value</p>
<ol class="tk-row">
<li class="tk-node tk-node-1"><b>Energy and capital</b><span>go in</span></li>
<li class="tk-link" aria-hidden="true"><i>through chips</i></li>
<li class="tk-node tk-node-2"><b>Intelligence</b><span>is created</span></li>
<li class="tk-link" aria-hidden="true"><i>consumed efficiently</i></li>
<li class="tk-node tk-node-3"><b>Business value</b><span>measurable, at the other end</span></li>
</ol>
<p class="mp-strip"><b>Draft</b><span>Version 0.5.2, dated September 1, 2026. It will move.</span></p>
<figcaption id="tk1-cap">The definition from his slide as three steps. The marker walks them in order; the last one is the one the business cares about.</figcaption>
</figure>

## Cost per outcome beats cost per token

Token counts are not the important value metric, he said. Even cost per token isn't right. Cost per task is better, cost per outcome is better still, and what you finally need is business value from the total cost of AI, not just the cost of tokens.

A token count tells you how busy the models were. It can't tell you whether the work was worth doing. My [TokenOps note](/notes/tokenops-agent-cost-curve.html) has the other half of this problem: why the cost of one task can grow much faster than the number of turns it takes.

<figure class="fd-fig tk-ladder" aria-labelledby="tk2-cap" data-desc="Five ways to measure AI, as a staircase that steps right as the metrics get better, from the one he says to skip to the one to aim for. Token count: not the important value metric. Cost per token: even this is not right. Cost per task: better. Cost per outcome: better still. At the bottom, in the marker colour: business value from the total cost of AI, not just the cost of tokens. The rungs light up one after another, top to bottom.">
<p class="tk-big">From counting tokens to business value</p>
<ol class="tk-rungs">
<li class="tk-rung" style="--i:0"><span class="tk-tag">Skip</span><b>Token count</b><span>Not the important value metric</span></li>
<li class="tk-rung" style="--i:1"><span class="tk-tag">Not right</span><b>Cost per token</b><span>Even this is not the right one</span></li>
<li class="tk-rung" style="--i:2"><span class="tk-tag">Better</span><b>Cost per task</b><span>What one piece of work cost</span></li>
<li class="tk-rung" style="--i:3"><span class="tk-tag">Better</span><b>Cost per outcome</b><span>What one result cost</span></li>
<li class="tk-rung tk-rung-top" style="--i:4"><span class="tk-tag">The point</span><b>Business value</b><span>From the total cost of AI, not just tokens</span></li>
</ol>
<figcaption id="tk2-cap">Five metrics, the one to skip first and the goal last. The steps are his; the one-line glosses are mine.</figcaption>
</figure>

## The bill is more than the model

The total cost of AI includes everything under the models. He listed what agents quietly pull in: idle compute, storage, network, data, and often licensing and software. Then labor: your time, your colleagues' time and the company's time to manage all of it, plus what it does to hiring and staffing plans over time.

A token meter sees none of that below the line. My own [Meterlex](/projects/meterlex.html) prices every model call from the session logs, which is exactly the visible slice. Useful, and only the top of the answer.

<figure class="fd-fig tk-berg" aria-labelledby="tk3-cap" data-desc="The cost of an agent stack as a block above a waterline and a set of blocks below it. Above the line, in cobalt: the models, which is the tokens, the part a token meter sees. Below the line, on grey with dashed outlines: idle compute, storage, network, data, licensing and software, and labor. Labor means your time, your colleagues' time and the company's time to manage all of this, plus the effect on hiring and staffing plans. The sizes are not to scale: the talk gave no proportions.">
<div class="tk-top"><b>Models</b><span>The tokens. What a meter sees.</span></div>
<p class="tk-water">Above the line a meter sees it. Below, it does not.</p>
<ul class="tk-below" aria-label="Costs below the models">
<li><b>Idle compute</b></li>
<li><b>Storage</b></li>
<li><b>Network</b></li>
<li><b>Data</b></li>
<li><b>Licensing and software</b></li>
<li class="tk-wide"><b>Labor</b><span>Your time, your colleagues' time, the company's time to manage it, and the effect on hiring and staffing plans</span></li>
</ul>
<figcaption id="tk3-cap">Not to scale. The talk gave no proportions, only the list.</figcaption>
</figure>

## The business counts something else

The business, he said, does not care how many agents you have, how many tasks they finish, how many lines of code you write, how many PRs you ship or even how many features. It cares about revenue, outcomes, and how the organization changed. And that side is less cool than what you can ship with an agent this afternoon, which is why it's talked about less.

He sees a pattern from cloud coming back. There it went too expensive, so people pulled workloads back: repatriation. With AI the first signs are there already, in the stories moving from token maxing to how we realize value on the other end.

<figure class="fd-fig tk-count" aria-labelledby="tk4-cap" data-desc="Two cards. The left, on dashed grey, is what builders count: agents, tasks, lines of code, pull requests shipped, features shipped. The right, in cobalt, is what the business counts: revenue, outcomes, and how the organization has changed. Under them a strip carries his warning from cloud: when the bill got too high, people pulled things back, and he sees the first signs of the same with AI as stories move from token maxing to realizing value.">
<div class="mp-two-row">
<div class="mp-two-card tk-dim">
<p class="mp-two-who">What we count</p>
<p class="mp-two-name">The shipping</p>
<p class="mp-two-line">Agents</p>
<p class="mp-two-line">Tasks accomplished</p>
<p class="mp-two-line">Lines of code</p>
<p class="mp-two-line">PRs shipped</p>
<p class="mp-two-line">Features shipped</p>
</div>
<p class="mp-two-vs">vs</p>
<div class="mp-two-card mp-two-peter">
<p class="mp-two-who">What the business counts</p>
<p class="mp-two-name">The change</p>
<p class="mp-two-line">Revenue</p>
<p class="mp-two-line">Outcomes</p>
<p class="mp-two-line">How the organization changed</p>
</div>
</div>
<p class="mp-strip"><b>Cloud, again?</b><span>Too expensive, pull it back. He sees the first signs with AI.</span></p>
<figcaption id="tk4-cap">His two lists, side by side. Nothing on the left is wrong; it's just not what gets the budget signed.</figcaption>
</figure>

## Three layers, and a gap at the end

The Foundation looks at the whole supply chain, in three layers. Production is the data center side: energy, chips, the AI factories and the models. Consumption is how well we use what comes out: the KV cache, prompt engineering, model selection, demand forecasting, even the cost of the harness. Value is not only cost per outcome or intelligence per watt but business impact, including pricing models and margins.

Their first survey, the State of Tokenomics, ran in September. Plenty of respondents were confident they could see their spending. Very few were confident they could show value on it to the person who signs off the AI budget, the CFO. He called that a real disconnect. The talk gave no figures and I haven't read the report, so I won't quote any.

<figure class="fd-fig tk-layers" aria-labelledby="tk5-cap" data-desc="The Tokenomics Foundation's supply chain as three columns. One, production: energy, chips, data centers, AI factories and models. Two, consumption: KV cache, prompt engineering, model selection, demand forecasting and the cost of the harness. Three, value: cost per outcome, intelligence per watt, business impact, pricing models and margins. Under them, the survey's gap as two bars: confident they can see the spend, a long bar, and confident they can show the value to the CFO, a short one. The bar lengths are schematic because the talk gave no numbers.">
<ol class="tk-cols">
<li class="tk-col"><p class="tk-col-n">1 · Production</p><ul><li>Energy</li><li>Chips</li><li>Data centers</li><li>AI factories</li><li>Models</li></ul></li>
<li class="tk-col"><p class="tk-col-n">2 · Consumption</p><ul><li>KV cache</li><li>Prompt engineering</li><li>Model selection</li><li>Demand forecasting</li><li>Cost of the harness</li></ul></li>
<li class="tk-col tk-col-v"><p class="tk-col-n">3 · Value</p><ul><li>Cost per outcome</li><li>Intelligence per watt</li><li>Business impact</li><li>Pricing models</li><li>Margins</li></ul></li>
</ol>
<div class="tk-gap">
<p class="tk-gap-h">First survey: who is confident</p>
<p class="tk-meter"><span class="tk-meter-n">Seeing the spend</span><span class="tk-bar"><i class="tk-bar-see"></i></span><span class="tk-meter-v">Plenty</span></p>
<p class="tk-meter"><span class="tk-meter-n">Showing the CFO the value</span><span class="tk-bar"><i class="tk-bar-show"></i></span><span class="tk-meter-v">Very few</span></p>
</div>
<figcaption id="tk5-cap">Three layers, then the disconnect from the survey. Bar lengths are schematic: the talk gave no figures.</figcaption>
</figure>

## What the community is building

Work is going on across all three. He named caching optimization, model routing and gateways, and Big T, a riff on Big O for tokens. On the Foundation's site Big T writes token growth as T of n times k times a: n is requests or input size, k is model calls per request, a is agent depth. The page lists more levels between them; I've only checked the two ends and the default:

| Level | What it is |
|---|---|
| T of 1 | A cache hit, no model call at all |
| T of n | The healthy default: one call per request |
| T of infinity | Unbounded: a loop with no termination condition |

The page's advice is to bound depth and add budgets. The launch announcement adds two more items of planned work: token cost telemetry in FOCUS, the FinOps cost-reporting format, and value measurement that starts with the share of work completed without human involvement.

All of it, he said, cascades into one job: take what your AI did and show the business in terms it recognizes. Revenue. Cost avoidance. Capacity gain. Labor augmentation. Speed to market, and not just code to market.

<figure class="fd-fig tk-cascade" aria-labelledby="tk6-cap" data-desc="Work under way on the left flows into the words the business recognizes on the right. On the left, three streams of work: caching optimization, Big T which is Big O for tokens, and model routing and gateways. An arrow labeled cascades into points to the right, to five phrases a CFO recognizes: revenue, cost avoidance, capacity gain, labor augmentation and speed to market, which he said means getting new capabilities out quickly, not just code to market.">
<div class="tk-cas">
<div class="tk-cas-in">
<p class="tk-cas-h">Work under way</p>
<ul>
<li>Caching optimization</li>
<li>Big T: Big O for tokens</li>
<li>Model routing and gateways</li>
</ul>
</div>
<p class="tk-cas-arrow" aria-hidden="true"><span>cascades into</span></p>
<div class="tk-cas-out">
<p class="tk-cas-h">What the business hears</p>
<ul>
<li>Revenue</li>
<li>Cost avoidance</li>
<li>Capacity gain</li>
<li>Labor augmentation</li>
<li>Speed to market<span>not just code to market</span></li>
</ul>
</div>
</div>
<p class="mp-strip"><b>The job</b><span>Show the impact of your AI in terms the business already uses.</span></p>
<figcaption id="tk6-cap">The technical work on the left only matters if it ends up in one of the five words on the right.</figcaption>
</figure>

## Try it

The Foundation's conference in San Diego, Tokenomicon with FinOps X, runs June 7 to 10, 2027. Before then:

- Watch the five minutes: [the keynote on the AGNTCon livestream](https://youtu.be/iNV5-YkPHW4?t=2381), starting at 39:41.
- Read the Foundation's page: [tokeneconomics.com](https://www.tokeneconomics.com) has the definition, the three layers and Big T.
- Read the launch: [the Linux Foundation's announcement](https://www.linuxfoundation.org/press/linux-foundation-launches-the-tokenomics-foundation-to-define-the-economics-and-roi-of-ai-value).
- Read the cost side in detail: [TokenOps, why a 40-turn agent costs 16×, not 4×](/notes/tokenops-agent-cost-curve.html).

Next time you open an AI bill, write one outcome next to it in the CFO's words. If you've managed a real cost per outcome, I'd like to hear how you defined the outcome.