June 18, 2026
Tokens Aren't Dollars
This post is part of my Medium blog.
I value the FinOps Foundation. I think JR is doing good work. But the pivot to "Tokenomics" — treating AI spend as a token-counting exercise — deserves pushback. Not because tokens don't matter. Because we're not going to be talking about tokens much longer.
Tokens are what we happen to know about right now. They're the meter on the API call. They're the line item in the billing dashboard. But the focus on tokens is tied to a specific moment: the single-model approach. When your organization says "we're using ChatGPT, everybody use ChatGPT" or "we're using Claude, everybody use Claude," then yes, you're going to be thinking about tokens. One vendor, one pricing model, one unit of consumption.
That's not where this is going.
Tokens from different models aren't interchangeable units — they're different currencies measuring different things. (Image Assist by Anthropic)
Apple's announcements two weeks ago should have told you something. Your laptop can do inference now. If you have a MacBook Pro, you've got 10, 20, maybe more GPU cores sitting there. They're not being used all the time. You can do text generation, speech recognition, a lot of that locally, on hardware you already own, without a token in sight. Or you can go out to ElevenLabs or OpenAI and use their services. The choice exists. The hardware is already there.
As open-source models get better — and they are getting better, fast — and as more storage ships with devices, we're heading toward a layered approach. Some inference happens locally. Some goes remote. The orchestration between those layers is the new infrastructure problem, and it's not measured in tokens. It's measured in compute capacity.
We're going to be talking about hardware again. We have to. And when you have a hardware-based situation, you're not focused on the instantaneous consumption of a cloud AI service. You're focused on the mixture — what runs local, what runs remote, what capacity you need to provision, and how to allocate it.
Here's where it gets interesting. I was talking to someone recently about an experiment they were running at home. They said, "Opus is really good at orchestrating agents, and then I shell out to lower-cost agents for individual tasks." That's a reasonable architecture. I've done that too.
But I actually tend to reverse it. I use a mid-tier model — GLM-5.1 — as my agentic orchestrator. I can't use Opus for orchestration. It's too expensive. The orchestration layer is doing routing, context management, coordination. That's volume work, not precision work. GLM-5.1 does a great job at it. And if I find during that orchestration that there's a specific task that needs a more sophisticated approach, then I'll use Opus for that individual task.
The expensive model goes to the task that needs expertise. The mid-tier model handles the coordination. The individual tasks are where I want the capability — not the whole pipeline.
This is the architecture FinOps for AI needs to understand. It's not one model. It's not one cost structure. It's layers.
And then there's the other problem with treating tokens as a currency. Even if you accept that framework — even if you're measuring tokens — it's a heterogeneous system. When I call Opus 4.8, it's a different currency than when I call GPT-5.4. "Tokens are tokens, Tim. What's your problem? You can measure the efficiency of tokens. You can measure the output. You can look at the benchmarks and compare them to cost."
Except different models have different strengths and different failure modes. Some models were built specifically for the benchmark. The benchmark doesn't tell you what the model will do with your specific workload. And the cost per token doesn't tell you what you're actually paying for, because a token from one model and a token from another model aren't the same thing. They're not interchangeable units of value the way a dollar on AWS compute and a dollar on GCP compute are roughly comparable.
Tokens aren't a currency like dollars. You can't compare them easily across models, providers, or architectures. Building a whole discipline around token accounting is building on sand.
The FinOps Foundation is right that AI spend needs financial discipline. They're right that the current tooling is inadequate. But calling it Tokenomics assumes we know what the unit of measurement is. We don't. We're in a transition from cloud-hosted, single-model, token-metered AI to something more distributed, more layered, and more complicated to account for.
The token is the easy part. The hard part — hardware allocation, the local-remote split, heterogeneous cost comparison, orchestration architecture — is where FinOps actually needs to be thinking. That's where the next wave of cloud financial operations lives.
Counting tokens is better than not counting anything. But it's not the discipline. It's the receipt.
In Essential, the third book in The Condition Set trilogy, the law that requires a human to remain in the decision loop is ninety days from expiring. Rob Coleman runs the agency that oversees every AI system in Canada. The question isn't whether the machines work. The question is whether anyone can tell when they stop working in the public's interest.