AI token costs are not flat fees, and the gap between power users and average users could be 20 to 1
When companies price AI access as a uniform per-seat subscription, they are assuming something the usage data may not support. Nikesh Arora puts a number on the disparity that should give any CFO pause.
The per-seat model for enterprise AI pricing rests on a quiet assumption: that consumption, while not perfectly uniform, is roughly predictable across a workforce. Nikesh Arora argues that assumption is wrong, and the mechanism he identifies is not a rounding error.
The problem is not that AI usage varies. Every enterprise software deployment produces some variance. The problem, as Arora frames it, is the magnitude. The employee who understands how to query, prompt, and chain AI tasks effectively is not consuming modestly more than a casual user. That person may be consuming 20 times more, measured in the tokens that actually determine what compute costs.
Arora puts it plainly: “The risk is your smartest employee who knows how to use AI really well could be using 20 times the tokens that an average employee uses.” The word “risk” is doing real work in that sentence. The observation is not that power users generate more output, which would be a straightforward argument for the tools. It is that their usage creates a cost structure that is difficult to model in advance and nearly impossible to recover through flat-fee pricing once the contracts are signed.
The risk is your smartest employee who knows how to use AI really well could be using 20 times the tokens that an average employee uses. Nikesh Arora
The practical consequence is asymmetric. A company that licenses AI seats on a per-user basis and then watches a cohort of skilled users run at 20 times the token rate of their colleagues faces a cost profile that was never priced into the deal. The vendor’s economics hold. The buyer’s do not. And the buyers who will be most exposed are the ones who did the most to encourage genuine adoption, who trained staff, rewarded experimentation, and pushed the tools into consequential workflows.
There is a further wrinkle. The employees driving the highest token consumption are, by Arora’s framing, the smartest users: the people who have internalized how to get the most from the technology. A company cannot simply throttle those employees without also throttling the output that most justifies the AI investment in the first place. The cost problem and the performance asset are the same population.
None of this makes AI adoption inadvisable. What it makes inadvisable is pricing it the way the industry currently tends to price it, as though a seat is a seat and consumption will average out. Arora’s 20x figure, if it holds in practice, means that the distribution of token usage inside a skilled organization is not bell-shaped and tractable. It is skewed, and the skew runs in the direction of the firm’s most capable people. Budgeting for the mean, in that environment, is budgeting for a number that does not exist.
The contracts being signed today by enterprise buyers are mostly not built around this reality. At some point, the gap between the pricing model and the actual cost structure will surface in renewal negotiations, in budget overruns, or in decisions to cap the users who drive the most value. How the industry responds when that happens will say something about whether the current per-seat model was a considered pricing strategy or simply the path of least resistance when the tools were new enough that nobody had seen the usage tail.