The Short Answer
Tokenmaxxing is unmeasured AI consumption: teams are encouraged to use AI everywhere, nobody attributes the spend to a workflow, and the cost compounds far faster than the productivity it buys. One operator reports token costs doubling every 45 days against roughly a 5% productivity gain. The fix is instrumentation plus substitutability — see the meter per workflow, and be able to move that workload to a cheaper model without a rewrite.
This is a genuine failure mode of usage-based pricing, and worth naming honestly. Per-seat licensing is still the wrong shape at scale, because you pay for every badge whether or not it logs in.
But usage-based only wins if you can see and steer consumption. Owning the stack is what makes that possible: cost attribution per agent, and model routing that treats the model as a swappable component.
What is tokenmaxxing?
Tokenmaxxing is the habit of pushing AI usage across an organization without measuring whether the output justifies the consumption. The term comes from investor Chamath Palihapitiya, who warned on CNBC in July 2026 that finance leaders are structurally blind to it.
His framing is specific: "CEOs and the CFOs, in my opinion, probably have no idea how much tokenmaxxing is going on inside of their organizations. I suspect what'll happen is one day you're going to have a miss, and EPS will be off by a few pennies, and the CEO will say to the CFO, 'What happened?'"
The mechanism is not waste in the ordinary sense. Nobody is being careless. Each individual decision — use a bigger model, add a reasoning step, widen the context window, run the agent on every ticket instead of the hard ones — is locally reasonable.
What makes it dangerous is that the cost of those decisions accrues in a line item nobody owns, on a meter nobody reads, until a quarter closes.
How fast can enterprise token spend actually grow?
Faster than budget cycles are designed to catch. Palihapitiya has described asking his CTO about token spend and being told costs were doubling every 45 days while productivity was up about 5%.
Run that forward. A cost that doubles every 45 days is roughly 8x over a year. A 5% productivity gain is not 8x anything, and no efficiency narrative survives that spread.
He has separately said his own company's AI spending was trending past $10 million a year, which he called "very scary" for a startup of that size.
The number matters less than the shape. Traditional software spend is a step function you negotiate once a year; agentic AI spend is a continuous function driven by how your staff and your agents behave this week.
| Spend type | Changes when | Finance sees it | Failure mode |
|---|---|---|---|
| Per-seat SaaS | Headcount changes | At renewal | Paying for badges that never log in |
| Metered tokens, uninstrumented | Any engineer changes a prompt | At quarter close | Compounding variance nobody owns |
| Metered tokens, instrumented + routable | Any engineer changes a prompt | Same day, per workflow | Requires you to own the layer that meters it |
Why can't finance teams see AI token spend?
Because in a rented stack, the meter belongs to the vendor and reports at the wrong grain. An invoice tells you the organization consumed some quantity of tokens last month. It does not tell you that 60% went to one agent, retrying a step that fails silently on malformed inputs.
Three specific blind spots recur:
No workflow attribution. Spend is aggregated per tenant or per API key, not per agent, department, or business process. You cannot manage what you can only see in total.
No unit economics. Almost nobody can state the cost per resolved ticket, per document processed, or per student interaction. Without a denominator, "AI cost is up 40%" is unreadable — it might be excellent news.
No feedback loop to the people spending it. The engineer who widens a context window or upgrades a model tier sees the latency change immediately and the cost change never.
Is per-seat pricing the safer alternative?
No — and this is worth being precise about, because tokenmaxxing is sometimes read as an argument for going back to fixed per-user licensing. It isn't.
Per-seat AI pricing is structurally wrong at scale for a different reason: it bills headcount rather than use. At roughly $60 per user per month, 5,000 staff costs $3.6 million a year whether or not anyone opens the tool, and enrolling the five-thousandth user costs exactly what the first one did.
Per-seat doesn't remove the variance — it hides it inside someone else's margin, then re-prices it at renewal when the vendor's own consumption math catches up. We ran the full arithmetic in the true cost math of per-seat AI.
The honest position is that usage-based pricing is the correct shape and it is only an advantage if you can see the meter. Both statements are true at once. The rest of this post is about the second one.
What should you instrument to control token spend?
Four things, in this order:
1. Attribute every call. Tag each request with the agent, the department, and the workflow that triggered it, so spend rolls up to an owner rather than to the organization. Cost analytics and per-interaction transcripts are an existing platform capability, not a project.
2. Define a unit economic per agent. Cost per resolved ticket, per appeal drafted, per record reconciled. This converts the AI line item from a mystery into a margin you can defend or kill.
3. Set the routing policy explicitly. Not every request needs the frontier model. Route by task difficulty and reserve the expensive tier for the calls that measurably need it — the pattern behind our open-source cost-optimizing model router, which cuts agent costs substantially by sending each request to the cheapest capable model.
4. Alert on rate of change, not on total. A budget threshold catches the problem a month late. A doubling rate catches it while it is still a configuration mistake.
How does owning the stack change the cost equation?
It changes who holds the two levers that matter: visibility and substitutability.
Visibility, because the meter runs inside your infrastructure, so attribution is a property of your own logs rather than a feature request to a vendor.
Substitutability, because the model becomes a component instead of a dependency. That is what converts the open-weight price collapse into money you actually keep — open-weight models now track 3–6 months behind the frontier on capability, a gap that has held steady for over 18 months, while sitting far below frontier pricing.
The June 2026 figures make the spread concrete: DeepSeek V4 Flash at $0.054 per million input tokens and $0.242 output — roughly 150x cheaper on output than the top frontier tier — and GLM 5.2 leading open weights at 51 on the Artificial Analysis index.
None of that arbitrage is available if moving a workload to a different model means rewriting the application. The value of a model-agnostic architecture is not ideological; it is the option to take a 10x price cut on a Tuesday.
What should a CFO ask this quarter?
Five questions, each answerable with a number:
- What did we spend on inference last month, broken out by agent and department?
- What is our cost per unit of work for the three highest-volume AI workflows?
- What is the rate of change in token spend — and does anyone get alerted on it?
- If a model 10x cheaper met the bar for our highest-volume workflow, how long would switching take?
- Who owns the AI line item the way a manager owns a cost center?
If those questions don't have answers today, the exposure is not that AI is expensive. It is that AI spend is currently unmanaged, and unmanaged compounding costs are exactly the kind that surface as a surprise with a date on it.
The organizations that avoid that outcome are not the ones spending least. They are the ones that own the meter, attribute every call, and treat the model underneath as a component they can replace.