What Tokenomics Is and Isn’t – Tokenomics Foundation

What Tokenomics Is, and What It Isn't

Executive summary

Thirteen conversations produced broad agreement on how wide the scope is, and almost no agreement on where to go deep. That is the whole finding.

It is not counting tokens. The one near-unanimous exclusion. Supporting evidence offered: an audit of one platform’s own AI spend found only about a quarter was direct model consumption. The rest was GPU and supporting infrastructure.
The closest thing to a shared definition is “relating cost to output.” Seven distinct definitional shapes were offered, and they are predictable by where the speaker sits: hardware talks tokens per watt, finance talks attribution and unit cost, engineering talks caching and routing, product talks pricing and margin.
The three-zone map draws directly contradictory advice. One member said pick one lane and create gravity within twelve months. Another rejected picking lanes at all and pushed the model into a loop. Both are credible. Both cannot be followed.
How this relates to adjacent practice is unresolved. The sharpest objection heard was that this is existing discipline with a new vocabulary. It was partly conceded once the new levers, personas, and buyer evidence were laid out, and partly not. Several boundary tests were proposed in individual conversations, but no single line survived contact with all thirteen.
Culture was deliberately left out of the draft definition, on the reasoning that the cultural layer already exists in adjacent practice and does not need restating here.
The value zone is the weakest-understood part of the map, conceded from multiple directions, with one member saying plainly it will not be solved in months.
Nobody confused tokenomics with crypto. Zero times in thirteen calls. The internal name worry did not materialize with members.

The bottom line

The definitional question is closer to settled than it feels. Members converge on a broad, cost-to-value scope. What is genuinely open is where to go deep first, and that is a governing board decision, not a research problem.

This page synthesizes what members, prospective members, analysts, academics, and internal leads actually said about tokenomics across thirteen onboarding and pre-board conversations. Quotes are verbatim from transcripts and deliberately unattributed. Nothing here is settled doctrine; it is the raw definitional terrain the technical steering committee has to resolve.

Read this first

The core finding

Almost nobody defines tokenomics as counting tokens, but almost everybody defines it differently. The disagreement is not about whether the scope is broad. It is about which edge of the broad scope matters most, and whether the pieces can be separated at all.

The elephant problem

The most repeated framing across the calls was the parable of blind men describing an elephant. Everyone is touching something real. Nobody has the whole animal.

They’re all kind of right, but they’re also all not seeing the whole picture, and that’s very much what I’m running into as I talk to different people, some in hardware, some in software, some at the frontier model layer, some in agent world. They’re all describing it differently.

Somebody will be like, tokenomics, it’s just about counting tokens. And somebody else says tokenomics, it’s about how AI costs impact business value. It’s all over the board.

The pattern is predictable by where the speaker sits:

Vantage point	How tokenomics gets defined
Hardware, data center, neocloud	Token throughput per watt. How much intelligence can I manufacture from the energy and capital I have?
Hyperscaler and model platform	Buy, manage, optimize. Procurement, credits, terms, observability, model portability.
Enterprise finance	Relating cost to output. Attribution, unit cost, budgeting, forecasting, capitalization.
Enterprise engineering	Caching, routing, context, quantization, model selection. The configuration levers.
Business and product leadership	Pricing models, monetization, labor substitution, margin impact.
Solution providers and analysts	An operating discipline for cost, quality, and value visibility, enough to govern it.
Academia	A ratio. Value created over resources consumed, evaluated at multiple resolutions.

What tokenomics is: the definitions on offer

Seven distinct definitional shapes appeared. They are not mutually exclusive, but they imply very different first deliverables.

1. Cost related to output

The most common formulation, and the closest thing to consensus. Not cost alone, and not value alone, but the relation between them.

Relating cost to output. I think at the end, that’s the big thing.

2. A ratio, at multiple resolutions

Value divided by resources consumed, computed at whatever altitude you care about: a token, a sentence, an answer, a workflow, a job, a role, a department.

It doesn’t matter what it is, it’s a value, just definition of value, and divided by the resources you put into it.

Tokenomics to me is economics at different levels. It depends on what level you’re concerned with. No matter which level you’re looking at, we need to start with the bottom, which is one single token.

This framing is the only one offered that actively reconciles the other six. It treats the competing definitions as different resolutions of the same ratio.

3. Constrained optimization, not cost minimization

Your organization’s Pareto frontier, not the generic benchmark published with every model release.

The economic side of it is the constraint solving, like the Pareto optimal frontier for my problem. Not for somebody else’s problem, not for the generic benchmark I see every time a new model is released.

If you’re a bank there’s constraints, and they’re non-negotiable. So you can’t drive cost to zero.

4. A token supply chain, from first principles

Treat the token as a manufactured good and ask for its bill of materials.

Almost like from first principles, what did it take to produce the token? What am I paying for? What are my options for producing one? I can buy them from someone. I can generate them. I can create them from scratch.

It’s no different than how much energy do my factories need, or how much food, how many rations, how much ammunition does every squad need? This is the supply chain for the AI era.

5. Effective work per unit of energy

The production-side view, framed as an efficiency and even a national-competitiveness question.

What is the most effective work you can produce per token? And that is different per company. It’s different per country.

It won’t be who can build the most data centers. It’ll be who can be most effective with what energy they have available.

6. An operating discipline for visibility and governance

The services and practitioner framing: five questions and two lenses.

Effectively just an operating discipline for making AI cost, quality, and value visibility enough to govern it.

The five questions as stated: what is the AI usage creating, what is driving the usage, what does that usage cost, which usage is actually worth the cost, and what decision needs to be made to improve it. The two lenses: financial defensibility and decision quality.

7. Cost-conscious AI, end to end

The broadest version, refusing every layer boundary on purpose.

To me, honestly, all that matters is end to end cost conscious AI, cost conscious agents, different workloads, training, cost conscious inference.

For me, the other layers, hardware, data center management, virtualization management layer, all of that matters to me.

The three-zone map, and the fight about it

The working map presented in most calls has three zones. It is the most tested artifact in the set, and the reaction to it is the single most useful signal in these transcripts.

Production

Converting energy, capital, and hardware into tokens. Data centers, power, cooling, silicon, hardware selection, hardware-level cache offload, self-hosted and open-weights inference.

Consumption

The familiar territory. Allocation, attribution, forecasting, budgeting, unit economics, routing, caching, context, prompt and model efficiency, guardrails.

Value

Business model and pricing impact, monetization, ROI, labor substitution, agentic labor accounting, margin, capitalization.

How members reacted to it

Value first. Several put the value zone first, explicitly as differentiation from market noise. One ranked the three “value, consumption, production, in that order.”
You cannot pick lanes. One member rejected the linear structure outright and pushed for a triangle or circle: “If I have production and consumption, but I’m not able to show the value, then to me it’s not linear. It’s a triangle.” Another said flatly you have to do all three.
You must pick a lane. The exact opposite advice, from a platform vendor: “If you pick too many lanes and you don’t have the impact early, then people are like, oh, this is just another comet in the sky. There are lots of comets. You have to create gravity.”
Production and consumption are inseparable. A different objection: “You have lots of decisions to make on consumption based on what you assume about production.”
Draw the line anyway. The internal rebuttal: “You can push things together and pull them apart. So you do just need to draw a line. We drew a line through this for the sake of making it simple to understand, not because it doesn’t go together.”

Unresolved

“Pick one lane and create gravity” and “you cannot separate the three” are both credible strategic advice from serious members, and they point in opposite directions. This is the first real decision the governing board owes an answer to.

What tokenomics is not

Not counting tokens

This is the one near-unanimous exclusion. The strongest supporting data point offered: an audit of one platform’s own AI spend found only about 25 percent was direct model consumption. The other roughly 75 percent was GPU and supporting infrastructure.

This isn’t about counting tokens. Anybody can count tokens, and counting tokens isn’t even really relevant, because tokens have such a different set of values and costs and outputs and inputs.

I am seeing an increasing number of people who think, oh, I can count tokens, I’m done. Everybody and their mother’s got some kind of token something.

Not a fork of FinOps, and not built on the FinOps framework

We are starting fresh in the sense that we do intend to tap some of that. But right now we’re not touching any of it. We’re starting with blank canvas.

One member endorsed this on grounds of intellectual humility: “The temptation is always to keep close what’s familiar. To your detriment in this case, because the ground’s moving. But you need the humility coming in to know that you know nothing.”

Not a cultural discipline

Culture was deliberately excluded from the draft definition. The reasoning: the collaborative, data-driven, accountability-based cultural practice already exists in FinOps and serves as the substrate.

It is a substrate. It is the foundation that guides the practice of tokenomics.

There’s technical and economic, and the FinOps aspect is the cultural element. So I think it will coexist.

Not cost minimization

It’s not just drive the cost to zero. At all.

And not “more output” either. Outputs were repeatedly distinguished from outcomes: “Your engineers will be sitting there running agents writing code all day long around bad ideas, and that’s just burning cost and not actually driving value.” Also flagged: “Do I have people just trying to get on a leaderboard to show how much AI they are using and producing zero value?”

Not blanket spend caps

I don’t think everything is equal. I might have some developers working on a really complex program, and they might need a lot more tokens than somebody working on something much more simple. Having a generalized hard limit for everybody is not it. I think we can do more math than that.

Not something you ask humans to optimize per interaction

The cognitive-load objection came up in three separate calls. Tool defaults return users to their last model, and nobody re-derives the cost-optimal choice each time.

We’re not going to sit down and think every single time I do something, am I using the most cost-effective model?

How do we know where to be sending this thing, and what best practices could we put in place so that I don’t have to think about everything I do? It just magically happens.

Not a settled unit of account

A token is not a token. Providers tokenize differently, output token types are not distinguished in most billing data, and a cheap per-token model can be more expensive in practice.

Cost per token looks great, but it consumes twice the tokens.

Most of the billing data we look at doesn’t differentiate output tokens. It just says output tokens. Was it an image, was it text? Was it reasoning tokens? That’s not in there a lot of the time.

Tokens can be made by a letter, a phrase, or a whole sentence. It’s just hard to abstract and explain and sell to somebody. And it’s inconsistent.

Not product building

Stated non-goals for the Foundation itself: it will not build model routing or intelligence layers, and it will not publish model benchmark scores. The stated alternative is to supply standardized metrics so members can benchmark their own workloads.

Not general AI education

I think we need to be careful to stay out of general AI training, like AI concepts.

The precedent cited is the FinOps Foundation’s refusal to teach cloud fundamentals. Counter-pressure exists: multiple members reported that enterprise audiences still cannot answer “what is a token,” and want that baseline covered somewhere.

Worth noting

Across all thirteen conversations, not one participant confused tokenomics with cryptocurrency or blockchain tokenomics. The name risk that gets raised internally did not materialize once with members. The naming friction that did surface was different: “token” sitting inside the word makes people assume it is only about tokens.

Counterpoints and live tensions

“It is FinOps plus”

The sharpest pushback in the set. The argument: the tenets are identical, and the market has been moving up-level for years already.

It is FinOps. It’s FinOps plus. Okay, it’s FinOps with new taxonomy. That’s my opinion. The practice is the same. You have the same tenets, real-time data for decision making, accountability, budget models. All of these are the same terms. It’s just a different taxonomy.

The partial concession, after hearing the levers and personas laid out: “The taxonomy, levers, persona, that is different. I agree.”

The counter-evidence offered elsewhere: the new levers are things FinOps practitioners have no exposure to. Model routing, context engineering, KV and cache-augmented generation, quantization, mixture of experts, low-rank adapters, hardware selection driving token output.

It’s all these things that frankly FinOps people know almost nothing about.

And the buyer evidence: “I’m not seeing the dominant buyer of our services right now be the FinOps professional. And where I do see the FinOps professional bring us into an opportunity, they’re bringing in other people who own this problem.”

The best boundary test anyone offered

Name it versus configure it

FinOps names and measures the concept. Tokenomics configures and optimizes it.

You as a FinOps practitioner would be saying, hey, we should look at prompt caching, and I know what that means, but I’m not going to tell you how to do prompt caching configurations. That’s a tokenomics thing.

It’s still just the person who comes in the room and goes, have we thought about KV caches, here’s a description of KV caches, versus someone who comes in the room and goes, show me how we’re configuring our KV caches.

A second framing worth keeping: the two disciplines as different projections of the same object.

Something like a tesseract. When we look at it from the FinOps perspective we talk about it in a particular way, and when we look at it from the tokenomics perspective we see it slightly differently, but it’s actually just a different projection of the same thing.

Too early, and moving too fast to standardize

The most serious strategic objection, raised most forcefully by a platform vendor weighing whether to allocate people.

I honestly don’t know how valuable that model is in a world that’s moving so fast. If you take six months to build a standard, in six months we might be beyond that.

People are intrigued, but they’ve got so much to get done themselves. And now they’ve got to take 20 percent of someone’s time to come and sit with you and figure this out.

The same speaker supplied the counterweight: “And with the risk that if we don’t, then we might be steering down a standard that we didn’t contribute to.”

The internal counter-argument: “Where there’s mystery, there’s margins. The more complex things get, the more that common language and common layer will be needed.” Also noted: a spec is not a standard. A spec can ship now; standardization takes years.

The authority problem

Named by one member as the single biggest structural challenge in the market, and framed as the Foundation’s biggest opportunity.

If you ask me as someone who sells to FinOps what my biggest challenge is, it’s that many finance people don’t have authority. That’s the biggest challenge in our market, the biggest one.

The follow-on: heads of AI transformation typically do not have authority either. “They’re supposed to do exploration, get interesting technologies in front of the right stakeholders, but they don’t have authority.”

The practitioner may not exist yet

They’re not really classic FinOps people, who my people are, and who are now trying to learn the token side.

When we get to the purest of the practitioners in this space, and I’m caveating that we’re not sure there is a tokenomics practitioner in the way there’s a FinOps practitioner as a role, they tend to be principal engineers.

The internal position is to define personas by activity across the three zones and let the practitioner role emerge, explicitly to avoid the FinOps precedent: “In FinOps land, we invented the practitioner and told them how to influence FinOps. And then the business spent the next five years still trying to realize that FinOps is everybody’s job.”

The cost of that choice, named honestly: “Maybe the one thing we don’t get from not defining the practitioner is that it’s harder for us to pin down who this is for.”

Bleeding edge versus long tail

A meaningful share of the market is still selling cloud cost management internally. One anecdote from a public-sector attendee at a recent event: “Wait, we came to learn about cloud FinOps. Why are you talking about this other thing?” Regionally, reported inbound skewed heavily toward Southeast Asia and the UK, with comparatively little North American AI cost inquiry, and much of it at a 101 level.

There is a really long tail of companies still just coming to adoption. And when you think about the tokenomics stuff, that’s the bleeding edge.

The value zone is the least understood

Conceded from multiple directions. Internally: “The risk with that area is we don’t really understand it. We’re not even sure what the map looks like.” From a member: “It’s the end picture we are striving for, but I’m not sure we’re going to solve it in the next months.” And a hard scope objection from a platform vendor, drawing the line the other way: “That’s the value of AI, but it’s not the token thing.”

Also: value may not be standardizable into one method. “Every company is a bit different in how to attribute costs to value, so I don’t think there’s a one-size-fits-all approach.” Value itself is plural: revenue, cost avoided, time saved, customer satisfaction, and in public sector, lives saved.

The empirical basis does not exist yet

I don’t even think we’ve gathered empirical data to say, well, if you did this with this, the results would be close, or close enough for your job. So, are we overspending based on quality? I don’t know.

Related and unsolved: budgeting. “I don’t know how much I’m going to spend next month or the month after or next week even. I’m turning things on for developers. Now I’m turning it on for entire business units that are not developers. I have no idea what it’s going to cost.”

Naming

“AI economics” was independently proposed as more holistic, and internally there is appetite to treat the two terms as synonyms. The honest assessment of why the current term wins: “Tokenomics is more fun. It’s just fun to say. That’s why everybody’s on it.” And: “It’s changing week over week, so I think we have a chance to define what the right term is.”

How technical should this be?

One member argued for maximum depth: “I’m in the camp that we should make the people dealing with this as technical as possible.” Others observed a hard ceiling: “They’ll get so far, but then they need to know what questions should I be asking of the person doing the tokenomics?” The compromise that emerged internally is a concept-versus-instance rule: teach the concept, not this week’s release.

Do you need to know which of these have come out and what models they apply to, or do you want to just understand what the technique is? There might be a new one tomorrow. I’m not going to teach you that.

A working definition, offered to be argued with

For TSC debate

Draft

Tokenomics is the discipline of relating the full cost of AI to the value it produces, across production, consumption, and impact.

It covers all AI and AI-adjacent cost, not token spend. Compute, hardware, storage, database, energy, data center, and the licensed and embedded AI inside software you already buy.
It is technical and economic. It borrows its cultural practice from FinOps rather than restating it.
It is constraint-based. The objective is your organization’s optimal frontier under your non-negotiable constraints, not the lowest cost and not the highest capability.
It is multi-resolution. The same value-over-resources question applies to a token, a workflow, a role, and a department, and it must reconcile upward from a common unit.
It operates before the spend, not only in reporting after it. Model, hardware, and architecture choices are in scope.

The open questions this doc does not answer

One lane or all three? Credible members argue both, forcefully.
What is a token, and can any definition be made to hold across providers who will not change their tokenizers?
Does tokenomics own the value zone, or does it stop at the consumption boundary and hand value off?
If there is no practitioner role, who is the audience for the framework, the training, and the certification?
Does the discipline converge back into FinOps? One board-level framing was that this is a compressed problem to be solved now, and if the job is done right it eventually reads as one discipline.
Does the Foundation serve the bleeding edge or the long tail, and can it credibly serve both at once?
Is “tokenomics” the durable name, or is “AI economics” the term the market settles on?