The State of Tokenomics 2026

There is a question every AI budget eventually meets, and it is not “how much are we spending?” Most companies can answer that, more or less. The question is the one the CFO asks next: “and what did it get us?” A new survey suggests that, two years into enterprise AI adoption, roughly three in four organizations still cannot answer it with any confidence, and that the ones who can are doing a small number of specific, learnable things.

The State of Tokenomics, published on September 23, 2026 by the Tokenomics Foundation, a Linux Foundation project formed this year to build open standards for AI cost and value management, is the first large survey of its kind: 472 responses across 11 industries, from companies with a combined $4.6 trillion in revenue and a median revenue of $1.8 billion. This is our reading of what the data says, why the value gap it describes is so persistent, and what a FinOps or platform team should take from it. We have drawn on the report's figures throughout, with attribution; the interpretation is ours.

Key Takeaways

• Three in four enterprises cannot confidently prove AI value to their CFO, and only 9 percent are very confident they can, per the State of Tokenomics 2026.
• Visibility is not value: 60 percent of companies with moderately confident spend visibility still cannot produce a measurable outcome, because a total is not an attribution.
• The confident minority meter and attribute spend per team and workload, track concrete outputs rather than tokens, and have a named owner, which correlates with a 3.7x higher chance of proving value.
• Model routing is nearly universal in intent (86 percent evaluating or using) and router users are four times more likely to show CFO value.
• Buyers want transparency, not discounts: 23 percent asked providers for granular data and 19 percent for standards like FOCUS, against 4 percent asking for lower prices.
• A shift toward open-weight models is expected, with heavy-frontier users planning the biggest move, which will make attribution harder as spend spreads across more surfaces.

What Tokenomics Means Here, and Why the Word Is Changing

A clarification first, because the word carries baggage. In this context tokenomics has nothing to do with cryptocurrency. The Tokenomics Foundation defines AI tokenomics as the discipline of converting energy and capital into AI capabilities and consuming that intelligence efficiently across the organization to realize measurable business value. The FinOps Foundation, in its own writing on the subject, puts it more simply: tokenomics is FinOps applied to AI. That framing matters, because it locates the discipline where it belongs, alongside cloud cost management rather than alongside crypto, and it explains why the survey's findings will look familiar to anyone who lived through the first decade of cloud spend. The mechanics are new; the organizational failure modes are not. We covered the underlying unit in our guide to token economics and TokenOps.

Finding One: The Value Gap Is Not a Visibility Gap

The report's central number deserves a slow read. Asked how confident they were that they could connect AI spend to a measurable business outcome their CFO would accept, 39 percent of respondents said not confident and a further 34 percent said only slightly. That is 73 percent, three in four, unable to answer the CFO's question with conviction. Just 9 percent were very confident.

The instinctive explanation is that companies cannot see their spend, and the survey shows that explanation is wrong. Spend visibility was considerably better than outcome confidence: 53 percent were at least moderately confident they could see what they spent. But of the organizations with moderately confident visibility, 60 percent still could not provide a measurable outcome, and even among those with highly confident visibility, only a quarter could. Seeing the bill did not close the gap; it merely made the gap visible.

This is the finding that should reorganize how teams think about the problem, because most AI cost tooling stops at the dashboard. A dashboard tells you the total. A CFO-grade answer requires knowing which team, feature, and workflow that total bought, and what output it produced, which is a question of attribution rather than observation. The survey names exactly this: the two traits shared by the very confident are that their spend is metered and attributable across workloads and teams, and that they track concrete business outputs such as tickets, pull requests, and revenue rather than token counts alone. That is the method we set out in our guide to AI unit economics, and the survey is, in effect, its empirical case.

One more detail complicates the comfortable story that this is a maturity problem that scale will solve. Confidence falls as revenue rises. Thirty-four percent of companies under $100 million in revenue were very confident in their spend picture, against 14 percent of companies between $100 million and $1 billion and 23 percent above $10 billion. Smaller companies see the whole picture more easily. Scale adds vendors, business units, procurement channels, and shadow usage, and each one fragments attribution further. Growth makes the problem worse, not better, unless the attribution is built deliberately.

Finding Two: Ownership Multiplies Everything

If the value gap has a single strongest predictor in the data, it is whether anyone owns the problem. Eighty-eight percent of respondents said tokenomics had a defined owner: 35 percent named the CTO, CIO, or a technology function, 26 percent said it was shared across functions, 9 percent named the CEO or executive team, 6 percent an AI leadership role, and 5 percent finance. Twelve percent had no owner at all.

The consequence of that last group is the starkest line in the report. Organizations with defined ownership were 3.7 times more likely to be able to show value to the CFO. Among the 12 percent with no owner, not one could connect AI spend to a CFO-acceptable outcome. Not a reduced share; none.

This is the FinOps lesson arriving on schedule. Cloud cost became manageable only when someone was accountable for it and the practice was shared rather than delegated to finance alone, and the survey shows AI following the identical path: technology leads, shared ownership is close behind, and finance-led ownership is rare. The organizational design in our guide to FinOps for AI, a named practitioner with engineering, finance, and product each holding a role, is not a nicety; on this evidence it is the difference between being able to answer the CFO and not.

Finding Three: Routing Is the Lever Everyone Is Reaching For

Eighty-six percent of enterprises reported that they were evaluating or already using a model router, software that directs each request to an appropriate model rather than sending everything to one. The adoption is spread across company sizes and split between purchased tools, with OpenRouter and LiteLLM the most cited, and homegrown builds, which remain common; the largest enterprises often run several at once. The report is candid that much of this is early-stage evaluation rather than mature deployment.

What makes routing more than an optimization detail is the second number: organizations using routers were four times more likely to be able to show CFO value. We would read that correlation carefully. Routing does reduce cost, often dramatically, because sending routine work to a model a tenth the price of a frontier model is the largest single saving available, a point we make in our guides to LLM token costs by model and LLM cost optimization. But the likelier explanation for the 4x is that routing forces the instrumentation attribution depends on. You cannot route a request without classifying it, tagging it, and measuring what it cost on the model it ran on, and those are precisely the per-request records that turn a total into an attributable answer. Routing may be less a cause of CFO confidence than a marker of the discipline that produces it.

Finding Four: What Buyers Want From Providers Is Not a Discount

Asked what they wanted model and token providers to offer, respondents gave an answer that should embarrass any procurement team still leading with price negotiation. Twenty-three percent asked for more transparency and granular usage data, 19 percent for standards such as FOCUS, 17 percent for attribution and tagging, and 13 percent for efficiency guidance. Only 4 percent asked for cheaper prices. Without being prompted, 7 percent wrote FOCUS, the open billing data specification now generated natively by the major clouds, into a free-text field.

The report's own gloss is sharp: enterprises are not asking providers to charge less; they are asking for a bill they can explain to a CFO. The provider that offers usage data broken down by workload, tagged to the customer's own dimensions, in a standard format, gives its customers the raw material for the attribution they cannot otherwise build, and that costs the provider almost nothing. For buyers, the practical move is to make FOCUS-format export and per-request metadata a procurement requirement for AI vendors the way it has become for cloud vendors, a shift we discussed in our guide to the FOCUS standard.

Finding Five: The Open-Weight Shift Will Make Attribution Harder

Today the frontier labs have the workloads. Ninety-six percent of respondents use a frontier model provider directly and 87 percent use one through a cloud token service such as Bedrock, Vertex, or Azure Foundry. On a ten-point scale from fully open-weight to fully frontier, 51 percent placed themselves at 8 to 10 today. Asked where they expected to be in twelve months, only 24 percent expected to still be there, and the heaviest frontier users planned the largest shift, dropping an average of 2.7 points. Just 17 percent expected to move further toward frontier models. Nearly one in five already uses at least one of DeepSeek, Qwen, Kimi, or GLM.

Whether that shift arrives on the schedule respondents expect is an open question; surveys of intent often run ahead of behavior. But the direction has a cost-management consequence the report touches on in a different section. Half of AI consumers already spread their usage across four of the six procurement channels the survey lists, from direct provider APIs to cloud token services, embedded AI in tools like Cursor and Databricks, rented GPUs, private hardware, and edge deployments. Open-weight models pull more spend into the last three, where cost is GPU-hours and power rather than tokens, and where the invoice says even less about which team used what. Thirty-eight percent of respondents already factor energy consumption into their tokenomics, typically those running their own or rented hardware. The attribution problem is about to gain new dimensions, and teams that have not solved it for tokens will find it harder still for GPUs.

Finding Six: Pricing, Margins, and the Bill Nobody Modeled

The last finding turns the lens outward. Fifty-two percent of respondents had already changed their own pricing or were considering it because of AI cost, 30 percent reported no change, and 18 percent said it was too early to tell. The report describes companies moving toward usage-based and outcome-driven models to pass through cost variability, alongside real pressure on margins from AI costs that traditional planning did not anticipate. This is the same dynamic we examined in our piece on AI gross margin and SaaS profitability: a feature whose cost scales with usage rather than with seats breaks the assumptions that SaaS pricing was built on, and the answer is to know cost per customer before setting the price, not after.

What Enterprises Say Is Hardest

Asked to name their biggest tokenomics challenges, respondents produced a ranking that is worth reproducing in full, because the order is the argument.

Read from the bottom up. Cost and pricing complexity, the thing most AI cost content is about, was the smallest challenge on the list at 7 percent. Efficiency and optimization, the thing most AI cost tooling is about, was second smallest at 9 percent. The top of the list, at 43 percent, was proving value, followed by attribution at 27 percent. The survey population is not struggling to understand token prices or to find savings; it is struggling to explain what the spend produced. That reframes what an AI cost practice is for. Optimization is necessary and it is not the goal. The goal is a defensible answer to the CFO, and everything else is in service of it.

What a FinOps Team Should Do With This

The report describes governance as immature, resting mostly on budget caps and token limits, with monitoring in place and ROI deferred to later. The data suggests a more specific sequence, drawn from what the confident minority actually do.

• Instrument attribution before spend grows. Attach team, feature, environment, model, and where relevant customer to every AI request at the application layer, because provider invoices cannot reconstruct it afterward. This is the single trait most strongly associated with proving value, and it is far cheaper to build into new code than to retrofit.
• Name an owner and share the practice. The 12 percent with no owner produced zero organizations able to show CFO value. Assign one, and structure the work so engineering, finance, and product each hold a role rather than delegating it to a dashboard.
• Adopt routing, and treat it as measurement. Route by task difficulty to cut cost, but recognize that the per-request classification routing requires is also the attribution data you need. Build the two together.
• Measure outputs, not tokens. Pair spend with concrete output metrics the business already values: tickets resolved, pull requests merged, cases closed, revenue influenced. Cost per output is the number a CFO accepts; cost per million tokens is not.
• Demand transparency from vendors. Make FOCUS-format export and per-request usage data a procurement requirement. The survey shows buyers want it far more than discounts, and providers can supply it cheaply.
• Plan for GPUs and open weights now. If a shift toward open-weight and self-hosted models is coming, attribution will need to cover GPU-hours and energy, not just tokens. The method is the same; the meters are different.

Together these are the operating model in our FinOps for AI guide, and they are the capabilities Opslyft's cost visibility was built around: attributing token, GPU, and agent spend to the teams and features that generate it, alongside cloud, so the answer to the CFO's question exists before it is asked.

Reading the Report With Appropriate Care

A few honest caveats. This is a self-selected survey, which tends to over-represent organizations already engaged with the topic, so the true share unable to prove value across all enterprises may be higher. Forty percent of respondents came from technology and software companies, whose AI usage and maturity differ from the rest of the sample. The publication is marked Release Candidate 1.0, so figures may be refined. The correlations it reports, ownership at 3.7x and routing at 4x, are associations rather than demonstrated causes, and we have tried to read them as such. None of that weakens the central finding, which is consistent across every cut of the data: most enterprises cannot yet connect AI spend to an outcome, and the ones who can are doing identifiable things.

The Bottom Line

The State of Tokenomics 2026 is, underneath its six findings, one finding stated six ways. Enterprises can see their AI spend well enough and understand their AI prices well enough; what they cannot do is explain what the spend produced. The gap is not closed by cheaper tokens, which almost no one asked for, or by better dashboards, which most already have. It is closed by attribution, ownership, and measurement of real outputs, the three things the confident minority share and the 12 percent with no owner entirely lack.

That is a familiar story to anyone who watched cloud cost management mature, and it points to the same conclusion. AI cost is becoming a discipline, with a foundation, a standard forming around it, and a body of evidence about what works. The organizations that treat it that way now, before the open-weight shift spreads their spend across GPUs and edge deployments, will be the ones with an answer when the CFO asks. The rest will be in the 73 percent.