Over the past year, tech-forward companies have gorged themselves on AI, gobbling up more and more tokens, the base unit of AI use. The scale is staggering. Google alone now processes more than 3.2 quadrillion tokens a month, roughly seven times what it handled a year earlier. Meanwhile, Uber burned through its entire artificial intelligence budget for 2026 by April—four months into the year.
After encouraging heavy AI use through token leaderboards and tokenmaxxing, tech companies are ready to pull back, and rightly so. Companies with no handle on their AI usage see worse returns. One study found that organizations with full visibility into their AI operating costs were five times as likely as others to report established ROI. Companies want more value from their AI, but struggle to get the employees they pushed toward it to change their behavior.
The central problem is that most folks simply don’t know how much energy and computing resources even the smallest prompt consumes. AI became easy to use before its cost became easy to understand.
However, the most effective AI-using companies won’t go on a crash diet. Instead, they’ll use AI tools that give their employees access to clear information about the true cost of each prompt before they run it—a “nutrition label,” so to speak.
Paying Without Knowing the Price
While improvements in AI infrastructure may make systems more efficient, and token limits can prevent the most egregious overconsumption, neither addresses the underlying problem: Employees can’t judge whether a prompt’s computational effort fits their intended outcome.
The issue is not just that employees are using AI for tasks that don’t need it, like checking the weather. It’s also that wasted AI use often accumulates through ordinary, everyday misalignment: unclear requests, failed outputs, repeated attempts, and conversations burdened with irrelevant context.
A single unnecessarily complex AI workflow using multiple tools, repeated attempts, and excess context might run a bill of $2.25. A company might be willing to stomach this cost in the name of “experimentation,” but the same employee running two of these inefficient prompts a day would rack up $99 a month. Repeat this usage pattern across 1,000 employees for one year and that’s $1.19 million from inefficient prompting.
Not only is the cost untenable for companies, the sheer energy usage is untenable as well. While each AI model has its own efficiency level, analyzing publicly available Nvidia chip specifications, industry-standard power usage effectiveness benchmarks, and average estimates on how fast AI models handle data reveals that a single token uses approximately 200 joules of energy. That’s enough to power a 10W LED bulb for 20 seconds. At the highest level of AI complexity, employing a team of multiple AI agents can use more than 1 million tokens, or the equivalent of nearly two days of household electricity.
As AI systems become increasingly autonomous, the opaque nature of prompts’ monetary and energy costs won’t survive for much longer. Companies will demand more transparency from frontier AI models.
OpenAI’s head of enterprise, Alexander Embiricos, noted that their business customers have shifted from asking “What can AI do?” to “Can I audit the efficiency of this model?” AI makers will soon provide the stats to do so.
Token Consumption Labels
An understanding of AI token consumption is as important for the C-suite as it is at the prompt box itself. I believe we’ll soon see visible token estimates for every prompt, akin to AI “nutrition” or “calorie count” labels. Nutrition labels emerged to translate complex vitamin levels, ingredients, and calories into standardized information consumers could quickly understand. An AI equivalent could estimate a request’s token use, cost, and computational intensity.
Just as a shopper weighs calories against nutrients, employees could weigh a prompt’s cost against what it delivers.
The value of this visibility would extend well beyond controlling excessive consumption. Users commonly respond to an unusable output by correcting the request, adding instructions, uploading more material, and trying again. Each attempt consumes additional tokens while carrying an increasingly cluttered conversation forward. Research indicates this weakens the results even when the relevant information remains available.
An AI label could reveal that expansion before another attempt begins, allowing the employee to correct a misunderstood task, narrow unnecessary scope, or proceed confidently when the anticipated work warrants the resources. At the company purchase level, buyers may demand AI tools be categorized under standardized efficiency ratings, similar to the Energy Star system for appliances.
Companies are already discovering this problem through exhausted budgets and maxed-out dashboards. But by the time they identify it, thousands or even millions of individual employee decisions producing those costs have long since accumulated.
As a result, the companies using AI most effectively will soon demand shared usage information at the point of input. They’ll gain visibility earlier in the process and will be able to better track what’s contributing to their AI spend, instead of solely tracking tokens. At the same time, employees will improve at matching computational bandwidth to the importance of an assignment.
Capable companies may ultimately consume more AI, but with a greater share of that use directed toward work that deserves it. In this new era of more conscious AI consumption, understanding what goes into an answer will matter as much as the answer itself.