Companies' AI Token Costs Now Exceed Employee Salaries, Leaving Bosses Stunned

As the title suggests.

This is truly no joke.

This is a genuine complaint voiced by many companies recently, after vigorously, or even fully, embracing AI.

A month ago, Uber's CTO revealed that since the company deployed CC (presumably a custom AI system or tool), over 90% of its engineers have been using it at a high frequency.

However, while AI usage is astonishing, the resulting cost bills are equally staggering.

According to the CTO, Uber's AI budget, originally intended to cover the entire year of 2026, was completely depleted in just the first four months of this year.

On average, the monthly AI usage cost per employee ranged from $500 to $2000.

Coincidentally, Nvidia's VP of Deep Learning also recently revealed in an interview that for their team, compute power costs now far exceed employee costs.

Seriously... we're talking about a statement from an executive at a hardware company that specializes in selling compute power.

Furthermore, while browsing forums, I've seen numerous cases of teams' token bills skyrocketing, AI budgets spiraling out of control, and ROI becoming unsustainable.

At first glance, doesn't this seem a bit counter-intuitive?

It's important to remember that for decades, the software industry has operated with near-zero marginal costs, and we've grown accustomed to the logic and intuition that digital products become cheaper the more they are used.

However, in the age of AI, things appear to be different. Compute power consumption exhibits entirely distinct economic characteristics; it's more akin to a high-consumption industrial technological raw material than infinitely replicable code copies.

Why is this happening?

Some argue that tokens are inherently expensive, especially for top-tier models and tools, which is unavoidable. Moreover, the working mechanism of modern AI also dictates its extremely high token consumption.

With the popularization of Agents, AI has long ceased to be a simple Q&A tool. Instead, it has evolved into a dynamic system capable of autonomously breaking down tasks, performing iterative deep reasoning, repeatedly invoking tools, and continuously self-correcting. This infinitely looping work mode causes token consumption to grow exponentially, far exceeding linear estimates based on common understanding.

When it comes to practical application, I believe the first thing to understand is where your tokens are actually being burned: are they being wasted on inefficient volume generation and redundant human overlays, or are they being spent on the cutting edge where they can truly reshape value?

This reminds me of a rather amusing piece of industry news from a while ago.

Some companies, in their eagerness to quickly demonstrate their commitment to transformation amidst the AI wave, crudely adopted token consumption as a KPI to measure employees' embrace of AI. This, in turn, spawned the so-called "Tokenmaxxing" trend...

Things like internal AI token usage leaderboards, incorporating token consumption into performance reviews, treating it as a productivity metric, or even an employee's identity tag... well, frankly speaking, isn't this a bit hasty?

How is this any different from the laughable practice of some companies years ago that used lines of code or commit counts as KPI metrics for employees?

For example, when Uber's CTO observed their team members using AI, they noticed a phenomenon: the token consumption difference for the same person using the same tool on the same day could be as high as tenfold.

Uber's Head of Operations also stated that there seemed to be no direct correlation between internal token consumption and the actual value output of product features.

What do these observations indicate? I believe those within the company understand it better than anyone.

This, therefore, exposes a problem.

Many companies or teams, despite being equipped with the most advanced large models, still adhere to outdated thinking and models in their workflows and management. This kind of extensive investment, lacking precise targeting, is essentially a manifestation of technical debt, using expensive compute costs to cover existing compatibility issues.

Therefore, if we view the phenomenon of tokens being more expensive than people from this perspective, it's less a technical problem and more a growing pain in the process of AI transformation. Eventually, it will force companies to shift from merely adopting tools to genuinely overhauling their processes.

Anyone can burn tokens. The real challenge, and what many companies and teams will need to consider in the future, is who can burn fewer tokens to accomplish more valuable tasks, effectively "burning tokens on the cutting edge."

Only when the human-machine collaboration paradigm undergoes a fundamental change, ensuring every unit of compute power is precisely utilized in irreplaceable value-creation stages, will this equation truly balance out.

At this stage, many companies may find this very difficult to achieve.

Note: This article has been included in the open-source GitHub repository "Road to Coding" https://github.com/rd2coding/Road2Coding. It contains my curated self-study roadmaps + knowledge point summaries for 6 major programming directions (roles), interview hot spots, my resume, several hardcore PDF notes, as well as insights and reflections on a programmer's life. Feel free to star it.