The Token, the Stack, and the Body: What AI Actually Costs an Operator of One.
Everyone is selling you the opportunity. Nobody is pricing the entry fee.
That’s the whole con, and it isn’t even a deliberate one. The opportunity is real. The opportunity is enormous. AI filmmaking, AI production, AI-native everything — the upside is exactly as big as the loudest voices say it is. But somewhere in the gold rush, an entire category of cost went quiet. Not hidden. Just unmentioned. And the people walking into this work fastest are the ones who never got told what it actually costs to be here.
I got told. I got told by the invoices, by my bank, and by my body.
Recently I read a sharp piece by Michael J Domanic, Head of AI at Section, about enterprise inference costs. Smart article. He’s watching CFOs open their first real Claude bill and ask the only question a CFO ever asks, which is some version of how the hell did we spend this much. His answer is good. Set the strategy. Budget AI as a percentage of salary. Set token caps. Educate your people. Measure before and after. Track your outliers. Give it twelve months. Guardrails for an organization that suddenly discovered its experiment has a meter running.
I read all of it nodding. And then I closed the laptop and realized he was describing a world I don’t live in.
Because every single one of his moves assumes an org. Headcount to govern. A dashboard to watch. A salary line that absorbs the learning curve. Domanic is governing a system. I don’t govern the system. I am the system. I’m the CFO and the operator and the line item and the body in the chair, all of it stacked into one person who also has to deliver the work.
So this is the chapter he can’t write from inside an enterprise. This is what the meter looks like when there’s only one of you.
There are three taxes. Most people are panicking about the cheapest one.
The first tax is the token.
It’s the only one anyone’s talking about.
When an AI film project lands, I don’t guess at the cost. I pull receipts.
I track usage on everything. Every project, every job, the token burn is logged so that the next estimate isn’t a finger in the wind. When something new comes in, I go back through the archive and find the closest matches on two axes — complexity and timeframe — because a simple ten-day job and a brutal three-day job are different animals, and averaging them tells you nothing. I find the projects that actually rhyme with this one, and I build the estimate backward from what really happened.
Then I add contingency. Twice.
Fifty percent on time. Fifty percent on compute. A ten-day project carries a five-day cushion. A two-thousand-dollar compute estimate carries an extra thousand. That isn’t padding, and it isn’t pessimism. It’s the only honest way to price a process where the unknowns live around every corner. AI-generated work cannot be estimated accurately yet. Anyone who tells you they can quote it to the dollar is either lying or hasn’t been burned enough times. The contingency is how a studio of one stays solvent and credible when the tool itself is unpredictable.
This is the tax everyone’s writing about. The token tax. It’s variable, it’s attributable, you can estimate it, you can contingency it, you can bill it. It’s the manageable one.
It’s also the cheapest one.
The second tax is the stack.
It runs whether you do or not.
Token spend has a start and a stop. The project ends, the meter stops, the cost is captured.
The stack has no off switch.
GPT. Claude. Figma Weave. Adobe. Canva. Google. Every month, whether there’s a project on the table or not, the subscriptions must be paid. This is the keep-the-lights-on cost. The electricity. In the AI space, it isn’t optional — you cannot show up to serious work without a serious stack, and you cannot let that stack go dark between jobs.
That last part is the part people miss. You’d think you could pause the subscriptions in the gaps. You can’t. The in-between time is exactly when you learn, when you test, when you stay ready for the thing that hasn’t arrived yet. A dark stack means you walk into the next project rusty, a step slow, behind the tools. So the subscriptions buy two things at once — capability and readiness — and you pay for both even on the days you’re billing nothing.
Here’s where it gets hard. The token tax traces cleanly to a project. The stack doesn’t. No single job owns my Claude seat. There’s no client to hand the Figma Weave invoice to. So right now it’s overhead — the cost of doing business — and I recover it the blunt way, spread across my day rates and project rates, amortized like rent.
It works. It’s also unresolved, and I’m not going to pretend otherwise in my own books. The first tax is variable and attributable. The second is fixed and ambiguous. You feel every dollar of it, and you carry it alone.
The third tax doesn’t have an invoice.
It has a body. It’s my body, which has no line item at all.
Some of it you can almost count. The courses. The workshops. The paid sessions have a price and the price is real. But that’s the visible edge of something much larger and entirely uncounted — the hours spent learning. The hours testing what doesn’t work yet. The travel to and from. The 3am alarm to sit through a ninety-minute workshop on advanced AI filmmaking that you paid to attend. You pay in money. Then you pay again in sleep.
This is the tax that doesn’t reset. The token cost ends with the project. The subscriptions at least leave you holding something. But this one just accumulates — physical, mental, additive, no off-switch, no ledger to recover it against.
And here is the line that separates my world from Domanic’s entirely. In an enterprise, the learning curve is salaried. His operators get paid to test. Paid to play. Paid to wake up and figure it out on the company’s clock. The experimentation is a funded activity.
We’re not paid to learn. We’re not paid to play. We’re not paid to experiment. The entire nut is on us. We absorb the cost of becoming someone who can do this work, and then we absorb the cost of doing it — same person, two ledgers, no salary bridging the gap.
It’s not a hard number. It’s wear and tear.
Price the entry fee before you walk in.
So when the conversation is all opportunity — and right now it’s nothing but opportunity — understand that there’s an entry fee nobody quotes you. Three taxes, not one.
The first is variable and attributable. You can estimate it, cushion it, bill it. The second is fixed and ambiguous. You carry it and bury it in your rates. The third is invisible and unrecoverable. It doesn’t show up anywhere except in you.
Domanic mapped the cost of AI at scale, and he mapped it well. I’m mapping the cost of AI at one. And the part that should stop you is this: the token tax everyone’s losing their minds over is the least of it. The expensive ones don’t generate invoices. They generate fatigue.
I’ve been doing this work long enough now to feel all three at once. The meter, the stack, the body. And lately the third one is the one I notice.
It’s the wind, and I’m the canyon.
And the thing about wind on stone is that it takes and it makes in the same motion. It wears you down and it carves you out. The cost and the craft are the same gust. Nobody photographs a canyon and calls it damage. They call it beautiful. They just don’t think about the years of weather it took to get there.
I’ve spent the last few years building the cost architecture this article describes — the estimation model, the contingency logic, the honest way to carry overhead you can’t cleanly bill. If you’re running AI-facing projects and want someone who’s already been there, I work directly with agencies, brands, studios, and operators. Let’s talk.
And if you want the rest of how I think — the production, the creativity, the storytelling, the way all of this actually gets made — that’s what I’m writing at Some Assembly Required. Subscribe and stay close to the work.