In this article, we look at why AI spending is currently almost impossible to compare across different vendors and what is being done to solve this, following in the footsteps of what already happened with the cloud.
Ask three different vendors what a token is, and you will get three different answers. Yet, it is the unit used today to measure a large part of corporate AI spending, and the one upon which budget decisions worth hundreds of thousands of euros are increasingly based. The problem is not strictly technical. It is that a standard is missing: a shared way of defining and measuring consumption that makes it possible to read a vendor's invoice using the same yardstick as another's. Without a standard, every invoice tells a different story, and comparison is simply impossible. This is not the first time the IT industry has found itself in this situation. It already happened with the cloud, and a solution was found: we will see how later on.
A token is, as a first approximation, a fragment of text: in English, it roughly corresponds to four characters, or a portion of a word. "Understanding" might be split into two or three tokens; "the" is almost always a whole token. In languages with non-Latin characters—Chinese, Japanese, Korean—the ratio changes radically, and the exact same phrase can cost two or three times as many tokens compared to its English equivalent.
Up to this point, this might seem like a technical yet consistent concept. The problem is that every vendor tokenizes in its own way. One vendor's model might count 100 tokens where another counts 140 for the exact same text. There is no universal conversion table, because each tokenizer is built internally by the vendor, using proprietary logic that is rarely documented in detail.
Result: two invoices with the exact same number of "consumed tokens" can represent completely different amounts of work.
Not all AI spending goes through tokens. Many AI features embedded within broader SaaS software—Adobe Firefly inside Creative Cloud, Salesforce Einstein inside CRM—use the concept of the credit: a package of units that a company buys in advance and consumes over time.
To these is added a third metric, perhaps the most familiar to those dealing with licensing: the seat, meaning the named-user license. This is the model used to sell tools such as Microsoft Copilot, ChatGPT Enterprise, or Claude Team—a fixed fee per assigned person, often combined with additional credits for usage exceeding the included threshold. We could also talk about hybrid situations like Claude Enterprise, but for the moment let's stick to the basic metrics; we will return to it later.
The differences between these three metrics are not just terminological, they are contractual. Credits expire, typically at the end of a period (monthly or annual)—if you don't use them, you lose them. Tokens, on the other hand, usually represent a purchased amount that is freely consumed, often until the expiration of the contract itself, even years later. Seats follow yet another logic: the cost is fixed per person, regardless of how much that person actually uses the tool—a model familiar to those managing traditional software licenses, but which in AI practice almost always coexists with a second layer of consumption based on credits or tokens.
This means that two companies with the same volume of AI spending can find themselves in very different risk situations: those working on credits must carefully manage the consumption pace so as not to waste budget already paid for; those working on tokens have more flexibility, but less urgency and therefore, often, less control; those working on seats risk, conversely, not knowing whether they are paying for underutilized user accounts, a well-known problem in the world of software license management that in AI adds on top of everything else. Now that we have clarified the basics, let us resume, as anticipated, the topic of hybrid situations through a concrete example that gives a good idea of how much these three logics can coexist within the same contract. A bundle like Microsoft 365 E7 already includes, in a single seat license, access to Copilot—whose AI features, in turn, consume credits or tokens as they are used. The invoice shows a single cost line, but behind it lie at least two different measurement logics: the fixed one of the seat, and the variable one of actual consumption. Figuring out how much of the overall spending depends on the one or the other factor—and therefore where one can intervene to optimize—requires breaking down that bundle, rather than looking at it as a single block.
In recent years, the cost per single token has plummeted—analysts estimate a drop of nearly 99% compared to first-generation models. It would be reasonable to expect overall spending to have dropped accordingly. The exact opposite happened.
The reason is simple: newer models consume many more tokens to perform the same task. A prompt that once required a few hundred tokens can today—amidst an extended conversational context, more articulate reasoning, and internal verification chains—end up consuming tens of thousands. As the unit cost falls, the quantity required by modern technologies rises as if the two were linked without knowing which is the cause and which is the effect. Somewhat like what happened over the years for processors or storage: today the cost of a megabyte of storage is much lower than in the 90s, yet even the operating system alone would occupy a thousand hard drives of that era today.
Added to this is a second multiplier, often underestimated: agents. Unlike a user who asks a question and waits for an answer, an agent works autonomously, repeats, checks, and calls the model multiple times to complete a single task. An activity entrusted to an agent can end up consuming 5 to 10 times the tokens of an equivalent interactive prompt, and agents, unlike people, never stop working.
Put these three elements together—non-standardized tokenization, different metrics (tokens vs. credits vs. seat licenses), and consumption that grows despite falling unit prices—and the result is that today there is no reliable way to compare AI spending across different vendors.
This is not just a theoretical problem. It means that a company evaluating whether to shift a workload from one model to another, or negotiating a contract renewal, often does so without a real point of comparison, comparing figures that, on paper, look comparable, but in substance measure different things. Those who have long worked in the FinOps world will recognize this pattern. It is the exact same problem that arose with the explosion of cloud spending: invoices from different providers—AWS, Azure, Google Cloud—with completely different cost items, measurement units, and pricing structures.
The answer back then was not a tool. It was a methodology—FinOps, created to align finance, technology, and business around a common way of reading and managing cloud spending—which at a certain point also produced a technical standard: FOCUS (FinOps Open Cost and Usage Specification), created within the FinOps Foundation, with the goal of normalizing cloud consumption data into a common, readable format regardless of the provider. Today, the same path is repeating for AI. In June 2026, the Tokenomics Foundation was launched, promoted by the Linux Foundation—the same organization that hosts the FinOps Foundation—with the stated goal of building a shared standard for the AI economy. Founding members include Microsoft, Google Cloud, Salesforce, ServiceNow, and, of course, Flexera.
It is too early to say how rapidly this standard will mature—the normalization of cloud consumption, with FOCUS, took years to become widespread practice. But the direction is the same: first the common language, then the tools that will use it.
Flexera, among the companies that promoted the birth of the Tokenomics Foundation, has already embraced this concept: in its new AI Cost Management module within Flexera One, it introduced the principle of consumption normalization, applying the same comparison logic available for the rest of IT and cloud spending to AI spending (what we call unit economics in FinOps). Not only that, thanks to the introduction of the AI consumer as the baseline element of analysis (whether user, agent, workflow, service principal, or other), it has also been able to include the same attribution logic in order to know not just how much is spent, but who is consuming.
The stated goal is to offer a 360-degree view of IT spending, including traditional licenses, cloud, and now also AI, in a single coherent read. At WEGG, as FinOps consultants and Flexera technology partners, we share this approach and apply it every day: we want to ensure that companies can embark on an attribution path starting from normalized data, bringing the different measurement units—tokens, credits, seats—back to a common language. Work that translates into very concrete advantages: choosing wisely between different models, negotiating contract renewals on comparable bases, keeping costs under control before they get out of hand, and finally knowing who is spending what, and why.
We will talk about the topic in more detail in the webinar "Governing AI Costs", scheduled for October 27. Go here to register!
Insights
OUR OFFICES
OUR OFFICES
PADUA
Via Arnaldo Fusinato 42, 35137
MILAN
Viale Enrico Forlanini 23, 20134
ROME
Viale Giorgio Ribotta 11, 00144
Copyright © 2025 WEGG S.r.l. • P.I 03447430285 • C.F. 02371140233 • REA 311023
Azienda Certificata ISO 9001:2015 – ITA / ISO 9001:2015 – EN
