I watched an operations lead stare at a dashboard while a counter climbed faster than anyone could explain. Ten minutes later she whispered, “Someone just burned through a quarter-million dollars,” and the room went quiet. You can feel the instant shift from giddy experimentation to alarm.
I’ve spent years watching companies flirt with shiny tech, and I’ll tell you straight: you don’t need permission to be skeptical. You also don’t have to wait for an auditor to tell you your playground is a bonfire.
At Uber’s briefing someone said token bills were outpacing the output.
When Dara Khosrowshahi told investors it was “getting harder to justify” some AI projects, he was naming a pattern you already sense: token consumption rising, measurable value lagging. Executives treated models from OpenAI, Anthropic’s Claude and internal tools as if CPU cycles were free. You know the ritual—leadership wants evidence of adoption, teams want to show engagement, and someone starts weaponizing usage metrics to prove ROI.
How much does enterprise AI actually cost?
Short answer: more than a demo. The headlines say an anonymous client accidentally charged $500,000,000 (€465,000,000) in a month to access Claude, according to Axios—an extreme example that reads like a cautionary tale. The Wall Street Journal found banks and other firms seeing employees generate $100,000s per month (≈€93,000+) in token bills by feeding premium models trivial questions or endless chat loops. Those numbers stack quickly when multiply by scale.
At Meta an internal leaderboard became a public embarrassment.
Meta killed its token-burning leaderboard after it leaked and revealed a “Token Legend” who burned 281 billion tokens in a month—compute enough to re-create Wikipedia many times over. Amazon followed suit and removed its scoreboard after staff started assigning pointlessly expensive tasks to maintain rank. This is human behavior: give people a scoreboard and some will game it. The scoreboard became a carnival ride that spun with corporate dollars.
What are tokens and why do they matter for my budget?
Tokens are the units of compute you pay for when you call a model—OpenAI, Anthropic, Amazon Bedrock and bespoke internal LLMs all bill for those units. Asking a 175B-parameter model a simple factual question costs more than hitting a lightweight endpoint. If teams habitually use high-tier models for basic prompts, tokens multiply and so does your bill. You can treat tokens like napkins at a restaurant: small waste per person, catastrophic across thousands.
At a finance review the CFO asked for hard limits, not hopeful slides.
After the initial frenzy—when executives were comfortable saying “we’re experimenting”—the books started to matter again. Shareholders noticed that “tokens burned” is not a strategy. Internal audits revealed employees querying premium models for idle conversation and trivial formatting tasks. One company’s incident forced a policy change: throttles, role-based access, and cost-aware prompts.
You can do this without firing off a memo. I recommend three practical levers: set per-user or per-team quotas on Anthropic, OpenAI and Bedrock usage; route routine tasks to smaller models or retrieval-augmented systems; and add billing alerts tied to finance workflows so an extra $10,000 ($9,300) in unexpected spend trips a review.
There’s a cultural angle too. Reward actual outcomes—faster closes, fewer support tickets, cleaner code reviews—not raw token consumption. Otherwise you’ll only measure activity, not value.
The press will keep printing dramatic cases—the WSJ, Financial Times, Axios—but don’t mistake spectacle for inevitability. Companies spent freely because the easiest metric to point to was consumption. When consumption becomes a headline, boards start asking about the price tag.
You can keep experimenting and still run a ledger. I’ve seen teams cut bills by switching divertible work to cheaper models and by instrumenting query templates that reduce token churn. It takes discipline and a dashboard that everyone trusts.
So here’s the provocation: if your leadership keeps treating token burn as a vanity stat, are you measuring progress or burning capital to validate a narrative?