Tokenomics: The New Currency of AI, Explained
What even is a token, when you get down to it?
I just got my credit card statement, with a line item reading “AI plan, unlimited.” Unlimited. The most misbehaved word in the telecom industry, resurrected for AI twenty years later.
That takes me back to my first cellphone plan as a teenager: unlimited nights and weekends, some fixed number of daytime minutes. Unlimited, except my mother was the one paying the bill, and somewhere in the fine print sat a “fair use” policy, the kind that starts throttling you past a certain threshold without ever telling you what the threshold was.
AI is doing exactly the same thing today, with an even uglier word for it: tokenomics.
OK… Mathieu, another word ending in -omics, exactly what corporate vocabulary was missing.
I know. But this one’s worth five minutes, because most leaders I talk to still think they’re paying for a model. They’re paying for tokens. That’s not the same conversation, and it changes what you look at the day the bill starts to worry you.
A token is not a word
First thing to unlearn: a token is not a word. It’s a fragment, sometimes a whole word, sometimes part of one, sometimes just a space or a punctuation mark. “Anticonstitutionally” doesn’t make one token, it makes five or six depending on the model. Even “cat” with a leading space and “cat” without are two different ones, to the machine.
That sounds like a plumbing detail. It isn’t. It’s the unit everything else gets calculated on: price, context window size, latency. When a vendor announces a 200,000-token window, they’re not telling you how many pages you can send it. They’re telling you how many of these fragments it can hold in memory before it starts forgetting the beginning of your conversation.
The unlimited plan that never really is
Back to my teenage cellphone plan, knowing upfront the comparison limps a little.
A telecom plan sells a simple promise: past a certain number of minutes or gigabytes, either it costs more or it slows down. An “unlimited” AI plan, say, does much the same, except the unit being rationed isn’t minutes, it’s tokens, and the number shows up nowhere on the invoice. Somewhere in the terms of service sits a cap on requests or tokens, past which you get bumped to a less capable model, or you wait your turn.
And here’s where the comparison limps for real: nobody negotiates their cellphone plan business line by business line. Your organization’s AI usage, on the other hand, varies from one team to another by a factor of a hundred, sometimes a thousand. “Unlimited” hides that variance just as well as it hides the cap.
The asymmetry nobody sees coming
An output token costs three to five times an input token, at pretty much every vendor. The reason is simple: generating text takes more compute than reading it. A model that “thinks” before answering bills that reasoning at the output rate, even though you never see those tokens go by.
And models have no memory. They fake it. Every turn of a conversation resends everything, from the start: the system prompt, the full history, the results of tools already called. Ask the same question a second time in an already-long thread, and that second time costs more than the first, even though the question is identical, because the model rereads the whole conversation just to answer you. On an agent chaining tool calls, that’s no longer growth that adds up, it’s growth that multiplies.
Why leaders need to understand this measure
You don’t need to know how to count tokens yourself. But if you’re the one signing the contract, or approving the budget for an agent project, there are three reasons to grasp at least the principle.
The first is that “unlimited” is a pricing decision, not a technical description. Behind every all-you-can-eat plan, someone made a bet on your average usage, and that bet has a ceiling. The day your organization goes past that bet, you won’t find out by reading the contract, you’ll find out when responses start slowing down or degrading.
The second is that tokens are the one unit that runs through every architecture decision you make. An overly verbose system prompt, a conversation that never gets summarized, a document pasted whole into context: these are spending choices, made by someone in your organization, almost always without anyone noticing at the time.
The third, and the most important: the day someone asks you to approve an AI budget, the right question isn’t “how much does the model cost.” It’s “how many of these tokens are actually necessary for what we’re trying to do.” Nobody else is going to ask that question for you.
The currency changed. It’s worth knowing what you’re spending before you know how much.
Have a good week!