Insight / signal
AI token spend is a payroll decision. Most companies are making it wrong.
If you cannot explain what your token spend is producing, you do not have an AI strategy. You have a bill.
Most companies are treating their AI token bill like a technology cost. It is not.
It is a payroll decision wearing a developer invoice’s clothes.
That distinction matters more than almost anything else in how you should be thinking about AI right now.
The confusion is easy to see. When Eric Siu and Neil Patel discussed OpenClaw founder Peter Steinberger’s reported $1.3 million annual token spend, the number landed like both a flex and an anomaly. One person spending that much on AI tokens sounds absurd until you ask what the spend is actually buying.
He is not buying software seats. He is not buying traditional headcount either.
He is buying execution capacity that scales without proportional hiring.
That is not infrastructure spend. That is payroll replacement.
The market is starting to catch up to this distinction, and not gracefully. New data from Ability.ai suggests organisations are burning through around 13 times more AI tokens than last year, while most of that spend still runs through systems with no measurable ROI tracking. The Pragmatic Engineer made the same point in late April: lots of teams are consuming tokens at scale without a coherent framework for whether the spend is generating value proportional to cost.
They are not measuring. They are just spending.
That should make every founder, operator, and CMO flinch a bit.
The maths gets ugly fast
Average cost per million tokens across major providers reportedly fell from about $10 to $2.50 in a year, based on Ramp enterprise data cited by Investing.com. On the surface that looks like great news.
I do not think it is, at least not in the comforting way people want it to be.
Cheap tokens have hidden a lot of bad discipline. They have made it easy to treat AI usage like a low-stakes experiment rather than an operating decision. But subsidised pricing windows do not stay open forever. As demand rises and providers stop underwriting adoption so aggressively, the businesses that built their rhythm on untracked token burn are going to get a nasty surprise.
The budget shock is one part of it.
The bigger issue is that most teams still cannot answer the obvious follow-up: what exactly are we buying with this spend?
Taste does not come in an API call
This is the bit that gets brushed aside because it sounds soft, even though it is the hard bit.
Experience and taste beat raw AI output in creative and strategic work. Consistently.
A big token budget does not create advantage on its own. Better judgment does.
The teams getting disproportionate value from AI are not winning because they can afford more model calls. They are winning because they know which outputs are worth shipping, which arguments are hollow, which positioning lands, and where the model is confidently producing slop.
That is why the Brian Chesky example keeps resurfacing. Player-coach leaders who stay close to the work tend to build better systems than people who only manage from a distance. Small teams with real taste and AI leverage can outperform much larger teams because leverage amplifies judgment.
AI amplifies what you bring to it.
It does not replace the thing you are bringing.
That is where a lot of token-spend conversations fall apart. Companies are buying model access and expecting good output by default. But output quality depends on the judgment inside the instructions, the approval loop catching bad calls, and the operating layer holding the process together.
Without that, you are not building an AI workforce.
You are just burning money on probabilistic autocomplete.
The real decision
You are not buying AI tokens.
You are buying a workforce that runs on your behalf.
That means your token budget should be treated like any other headcount decision. With rigor. With measurement. With accountability. With a clear view of what work is being done, what value is being created, and where the waste sits.
If a team member cost you six figures a year and you could not explain what they produced, you would have a problem.
Token spend deserves the same standard.
The businesses that win this next phase will not be the ones with the biggest AI bill. They will be the ones with the clearest operating model behind it.
They will know:
- what tasks AI is handling
- what good output looks like
- where human judgment still matters most
- how failures get caught
- how spend maps to outcomes
- when a workflow is worth scaling and when it is just noise
That is the difference between strategy and novelty.
If you cannot explain what your AI token spend is producing, you do not have an AI strategy.
You have a bill.
The reckoning will come when token economics tighten and finance teams start asking harder questions. When that happens, the companies with taste, judgment, and orchestration discipline are going to look smart.
The ones who just bought more tokens will not.
If you are wrestling with that shift, this is exactly the sort of operating problem I care about most.