From Tokens to Total Value

I spent a few days in Amsterdam at Tokenomicon FinOps X, and one idea kept surfacing across the talks, workshops and hallway conversations: the economics of AI are too important to be measured in tokens alone. Enterprises are spending more on AI every quarter, yet most still can’t say clearly what that spend delivers. What follows are my observations on where the market stands and where it’s heading.

Production, consumption and value are one conversation

The day started with a chalk talk built around the Tokenomics Mind Map. The exercise was simple. We took the language that now surrounds AI economics and sorted each term into one of the three dimensions defined by the Tokenomics Foundation: production, consumption and value.

Some terms landed easily. Energy, GPU capacity, AI factory buildout and model training belong in production, since they are the structural investments needed to build AI capability. Model selection, prompt routing, token usage and demand forecasting sit in consumption, the day-to-day running of AI workloads. Revenue enablement, pricing models, FTE equivalents and risk reduction clearly belong in value.

The interesting debates came when terms refused to stay in one box. Context caching is a consumption technique, but it directly reduces how much compute you need to produce. Capacity planning spans production and consumption. Model routing is an operational decision that shapes both the cost base and the business outcome. By the end of the session, the room agreed that tokenomics works better as a Venn diagram than as three separate pillars. Decisions made in one area change what’s possible in the others, and treating them in silos is how organizations lose sight of what their AI actually costs and returns.

The state of tokenomics: Spend is visible, value is not

One of the highlights was the reveal of the State of Tokenomics research, based on 472 enterprise respondents across 11 industries representing $4.6 trillion in combined annual revenue. The data tells a story of a market that is maturing quickly, but unevenly.

Start with ownership. In 35% of organizations, tokenomics sits with the CTO, CIO or technology function, and another 26% run it as a shared cross-functional responsibility. That still leaves 12% with no defined owner at all, which is a real risk for a cost line growing this fast.

Then there is the model mix, which is about to shift significantly. Today, 51% of enterprises rely heavily on frontier models. Within 12 months, only 24% expect to still be in that position. Heavy frontier users expect to reduce their reliance by 2.7 points on average as workloads move to open-weight models, and just 17% expect to increase frontier usage. This isn’t a rejection of frontier models, but a sign that enterprises are becoming deliberate about which workloads truly need them.

The most striking finding concerns value. While 39% of organizations are moderately confident in their visibility of AI spend, that confidence drops as company size grows. And three in four companies cannot confidently prove measurable AI business outcomes to their CFO. Proving value and ROI was named the number one challenge by 43% of respondents, while only 7% cited cost or pricing complexity as a major obstacle. In other words, the industry’s problem is no longer understanding the bill. It’s explaining what the bill buys.

Governance reflects that tension. For developer productivity spend in particular, most organizations are still at an early stage, relying on budget caps and token limits. What I found encouraging is that the same organizations are investing in identifying and empowering their internal power users. The consensus is to hold the line on spend while backing the people who get the most out of the tools, and I think that is exactly the right instinct.

If you only optimize for tokens, you might be missing the bigger picture

A recurring theme across sessions was the iceberg. API and token costs are the visible tip, but the Total Cost of AI includes infrastructure, compute, memory, storage, data pipelines, energy, software licenses and labor. Building a CFO-defensible story means pairing that full cost with quantified business value claims at the same level of detail, workload by workload.

That is where the meter comes in. UNICC shared how they structure AI across three execution tiers: private AI on internal cloud, trusted AI with partner HPC in protected jurisdictions, and sanitized AI through approved public APIs with sensitive context stripped out. Tying it together is a centralized AI gateway that applies key-level budget caps, routes requests based on data classification and tracks token traces and cost per key, all aligned to FOCUS. The meter provides attribution. Attribution is what makes value provable.

Agent observability is becoming part of that same picture. Google’s open-source Agent Development Kit, for instance, now offers a BigQuery Agent Analytics plugin that supports OpenTelemetry-style tracing to visualize agent execution flows and tracks the origin of each tool call, whether a local function, MCP server, sub-agent or remote agent. The value for FinOps is clear: once every agent step, model call and tool invocation lands in a queryable dataset, you can connect cost to the specific workflow that generated it rather than to a single undifferentiated API bill. Paired with an agent registry that tells you which agents exist, who owns them and what they’re allowed to do, this is the foundation for governing agentic AI at scale.

Several levers came up again and again for turning that visibility into savings. Context caching avoids paying repeatedly for the same context. Model routing sends each request to the model that fits it best. The research backs this up: 86% of organizations are evaluating or using model routers, and those that do are four times more likely to demonstrate ROI to the CFO.

Using AI to govern AI

Alistair Scott from Infracost gave one of the most memorable proof points of the event. By using AI agents to govern their own AI usage, Infracost grew token consumption fivefold while costs grew only 2.4 times. Their approach follows a three-step pipeline of context gathering, optimization and remediation, drawing on prompt history and model price and benchmark data to choose the right model dynamically. It’s a strong example of a principle I believe will define the next phase of FinOps: the tools we use to manage AI will increasingly be AI themselves.

When compute becomes the constraint

Lundbeck offered a perspective from pharma that brought the production side of the Venn diagram into focus. Molecule discovery work that used to be bounded by lab capacity has become a compute task, with GPU capacity now the bottleneck. As they put it, introducing AI means FinOps breaks twice: once because the workloads themselves are compute-intensive and hard to forecast, and again because traditional FinOps frameworks weren’t designed to govern them.

The rise of the specialist model

Deutsche Bank shared that it is exploring small language models for niche use cases. The reasoning is straightforward: few frontier models have been trained extensively on financial services data. A targeted SLM can be both more accurate and considerably cheaper for domain-specific work. Combined with the survey’s shift toward open-weight models, this points to a future where enterprises run a portfolio of models matched to their workloads rather than defaulting to the largest one available.

The moment it clicked: value lives in the business

The observation that stayed with me longest came from Booking.com, and it reframes the value problem entirely. If you respond to the ROI challenge by standing up a central team and asking it to measure the value of AI across the company, you might want to reconsider.

Booking.com's point was that this is close to unwinnable. A central unit has to reconstruct, from the outside, what "good" looks like in a dozen different functions. That turns into slow investigative work, and the answer it produces rarely convinces anyone. The contextual understanding of what value means already exists inside each business unit. The people running support, marketing, engineering or finance know what a better outcome looks like in their own terms.

The ask should flow the other way. Business owners and line managers need to be the ones who assess and articulate the value AI creates in their area and own the responsibility of continuous valuable returns from AI, with FinOps team providing the cost and usage data to make that assessment credible. And when a unit can't yet document its value, that is not a failure. It's the actual task: work with that team to identify the measurable datapoints that would indicate or attest to value in their specific setup, then start capturing them. Proving ROI becomes a shared exercise in defining the right metric, rather than a central audit imposed after the fact.