Glassity Blog
What is AI token economics?
AI token economics is the practice of managing what AI costs and what it gives back, measured in tokens. Tokens are the units AI models are billed in. The practice covers four things: how much you spend, who spent it, whether the spending is growing for a good reason, and what you got in return.
Table of contents
If someone has asked you why the AI bill went up and you couldn't answer, you have already met this problem.
AI tokenomics is not crypto tokenomics
In crypto, "tokenomics" means how a digital coin is created, shared and rewarded. That's a different subject.
In AI, a token is a small piece of text. A model splits your request into tokens, writes its answer in tokens, and charges you for each one. One token is about three quarters of an English word, though this changes between models and languages.
So AI token economics is about the cost of running AI models. It has nothing to do with blockchain or coins. The two topics share a word, which confuses a lot of people at first.
Where the term comes from
The Tokenomics Foundation made this a formal discipline. It is a Linux Foundation project, with its own board and working groups. It describes tokenomics as turning energy and money into AI capability, and then turning that capability into business value.
It is a different organisation from the FinOps Foundation, which publishes FOCUS, the standard format for cloud billing. The two work on related topics, but they are separate.
The Tokenomics Foundation splits the work into three questions.
Production: how is the AI made? This covers data centres, chips, electricity and the cost of running models at scale. It matters a lot to AI providers and to companies that run their own models. It matters very little if you buy AI through an API.
Consumption: how is the AI used? This means splitting AI costs between teams, planning future costs, and reducing them. It works a lot like cloud cost management, but for AI.
Value: what did the AI give back? This is the hardest question, and nobody has fully solved it yet.
Most companies that buy AI spend their time on the second question and still have no answer to the third.
Why AI costs don't behave like cloud costs
If you already manage cloud costs, you might expect the same tools to work for AI. Mostly they don't, for four reasons.
The price of the same task changes. A cloud server costs the same today and tomorrow. An AI request doesn't. A longer prompt, a new model version, a retry, or an agent that makes forty calls instead of four can all change what the same task costs. Last month's number is a weak guide to next month.
The bill has no names on it. You can tag cloud resources to show which team owns them. An AI provider's bill arrives as one total, with no team, service or feature. If several teams share one account, they share one line.
There are different kinds of tokens, with different prices. Input tokens, output tokens, tokens read from cache, tokens written to cache, and on some models, reasoning tokens. Output usually costs several times more than input. Cached input often costs much less. If you only see the total, you can't tell whether the bill grew because people used AI more, or because caching stopped working.
The system around the model also costs money. Tokens are the easiest AI cost to measure, but they aren't all of it. The tools that organise requests, store memory, search documents and test results all run on cloud servers, storage and networks. That's cloud spending, and it belongs in the same picture.
What to measure, in order
If you are starting from nothing, this order works well.
- Total AI spend, separate from everything else. This sounds easy. But if you use Amazon Bedrock or Microsoft Foundry, AI costs arrive inside your cloud bill, mixed in with everything else.
- Spend by model. Many companies find they use an expensive model for a task a cheaper one does just as well. This is often the first real saving.
- Spend by token type. Input, output, cache and reasoning. If your bill grows while the number of requests stays flat, the cause is usually longer prompts or caching that stopped working.
- Spend by team, service or feature. This is the difficult step. Without it, you can't show teams what they spend, set a budget someone owns, or decide whether a cost is worth it.
- Cost per unit. Cost per request, per customer, per transaction, or per support ticket closed. Use whatever unit your business already tracks. This is where AI spending starts to make business sense.
- Value. What the spending produced. This is hard, and mostly unsolved.
Most teams get stuck between steps four and five, because you need to know who spent the money before you can measure cost per unit.
Why splitting AI costs by team is hard
There is no "team" field in AI usage data. None of these systems has one.
What exists is an API key. Engineers create keys, usually in a gateway, which is the tool they use to send requests to AI models. Every request is recorded against a key. If a key is named after a team or project, you can see which team spent what. If your keys are called prod-key-1 and test-2, you can't.
So this is mostly about how you name things. And the names have to be in place before the spending happens. You can't go back later and decide which team made a request last March.
How to attribute AI costs to teams when there is no team field →
How your engineers send requests also matters. If they go through a gateway, it records which key made each call, and you get detailed information. If they call a cloud AI service like Amazon Bedrock directly, the bill shows only the account, region and model. It doesn't show the person or team.
Why your Amazon Bedrock bill has token counts but no owner →
How AI cost and AI value connect
Everyone knows what AI costs, because everyone pays the bill. Far fewer companies can say what that spending achieved, or how good the results were. They can't tell what each service got for its tokens, and even less what each team delivered for what it spent. And when a cost has no clear return, it's usually one of the first things cut in the next budget.
Measuring value is hard in general. What was the value of an email an AI helped write? It's hard to say honestly.
In some cases, though, value is simple to measure. If an AI agent finds a cost problem and the fix saves a clear amount on your bill, you can compare the tokens it used with the money it saved. It's a small example, but it's one you can actually check.
Where to start this week
Find out if your engineers send AI requests through a gateway. If they do, the data you need already exists. If they don't, adding one is the most useful change you can make.
Look at how your keys are named. If the names don't mean anything, fix that before another month of spending goes by.
Put AI costs in the same reports as your cloud costs. Most companies will use FOCUS for this, because cloud providers already publish their billing in it.
Don't wait until you can measure value perfectly. Cost, ownership and cost per unit are all possible today, and each one is useful on its own.
Glassity reads your AI usage from the tools you already run, reports the tokens and the cost in the same format as your cloud bill, and lets you ask about both through the Glassity Agent.