Recently, burning tokens at our company costs more than hiring people. The boss is stunned — well, as the title says
And no, I'm really not joking
This is exactly the complaint coming from a lot of companies lately — especially those that have aggressively, or even fully, embraced AI
A month ago, Uber's CTO revealed that since the company rolled out Claude Code internally, over 90% of its engineers are using it at high frequency
But while AI adoption is staggering, the bill that comes with it is just as staggering
According to the CTO, Uber's AI budget — originally meant to cover the entire year of 2026 — was completely burned through in just the first four months of the year
For heavy users, monthly AI usage runs between $500 and $2,000 per engineer
Coincidentally, NVIDIA's VP of applied deep learning also recently revealed in an interview that for their team, compute costs have now far exceeded employee costs
Wait... let that sink in. This is coming straight from the mouth of an executive at a company that literally sells compute hardware
Not only that — browsing forums, I've also seen plenty of cases of teams with token bills blowing up, AI budgets spiraling out of control, and ROI that just doesn't cover the spend
At first glance, doesn't this feel a bit counterintuitive?
For decades, the software industry's marginal cost has trended toward zero. We've all grown used to the logic and intuition that digital products get cheaper the more you use them
But in the AI era, that's no longer how it works. Compute consumption now has a completely different economic character — it behaves more like a high-consumption industrial raw material than an infinitely reproducible copy of code
Why?
Some say tokens are just expensive — especially with top-tier models and tools, there's nothing you can do about it. And the way modern AI works inherently drives enormous token consumption
Because with the rise of agents, AI is no longer a simple one-question-one-answer tool. It has become a dynamic system that can autonomously break down tasks, loop through deep reasoning, repeatedly call tools, and continuously self-correct. This unbounded looping pattern makes token consumption grow exponentially — far beyond the linear estimates our intuition defaults to
On the practical side, though, the first thing I think you need to figure out is: where exactly are your tokens going? Are they being burned on low-efficiency token-farming and redundant human stacking, or on work that genuinely reshapes value?
That reminds me of a hilarious piece of industry news from a while back
Some companies, desperate to quickly prove their commitment to the AI wave, simply and crudely turned token consumption into a KPI for measuring how "AI-embraced" an employee is. And with it came the so-called "Tokenmaxxing" trend...
Internal AI token usage leaderboards, folding token consumption into performance reviews, treating it as a productivity metric, even as an employee identity badge. Um... forgive my bluntness, but isn't this a bit rash?
How is this any different from the old joke about companies using lines of code or commit count as employee KPIs?
Uber's CTO, while observing how their teams use AI, noticed a phenomenon: the same person, using the same tool, on the same day, could vary in token consumption by more than tenfold
Uber's COO also pointed out that there seems to be no direct correlation between internal token consumption and actual product value output
What does all this tell us? I'm sure the people inside know better than anyone
Which exposes a bigger problem
Many companies and teams may have deployed the most advanced large models, but their workflows and management still run on old thinking and old models. This kind of poorly-targeted, brute-force investment is, in essence, a form of technical debt — using expensive compute to paper over pre-existing compatibility problems
So if you look at "tokens costing more than people" from this angle, it's less a technology problem and more the growing pains of AI transformation. Sooner or later, it will force companies to shift from merely adopting tools to genuinely reworking their processes
Anyone can burn tokens. The real question is who can burn fewer tokens to accomplish more valuable work — putting tokens exactly where they count. That's probably what many companies and teams will need to figure out going forward
Only when the human-machine collaboration paradigm fundamentally changes — when every unit of compute is precisely aimed at irreplaceable value creation — will this ledger finally balance out
And right now, most companies probably can't pull that off
I'm a frontend developer based in China, writing Dual-Track Dev — a blog about Chinese developer tools and trends for a global audience
Top comments (0)