In April, Uber’s chief technology officer sat down to demonstrate the company’s AI coding tools and burned through $1,200 in tokens in two hours. Praveen Neppalli Naga wasn’t trying to make a point about cost.
However the number did stick, because it captured something finance teams across the industry were only beginning to grasp: a developer’s laptop had quietly become one of the least predictable line items on the company’s books.
By the time Naga made that demo, Uber had already blown through its entire 2026 AI budget, only a third into the year.
Average bills ran $150 to $250 per engineer per month; heavy users hit $2,000. “I’m back to the drawing board, because the budget I thought I would need is blown away already,” Naga told The Information. Uber has since capped spending on agentic coding tools at $1,500 per engineer per month.
That story has become something like the industry’s cautionary tale, and it’s the backdrop against an article from Gorilla Logic. The tech company’s article argues that AI token costs have moved from an IT footnote to a board-level problem in under a year.
“Twelve months ago, ‘token cost’ wasn’t a line item on most CFOs’ radar — now it’s showing up,” the article states.
The shift matters because it breaks a pricing logic finance departments have relied on for decades. Traditional software costs scale with headcount: more seats, more revenue, a predictable curve. Agentic AI doesn’t work that way. A single employee, or a single autonomous agent working unsupervised on a complex task, can consume enormous volumes of tokens in one sitting.
Gorilla Logic estimates that letting engineers use two agentic coding tools could run $3,000 per engineer per month, or $36,000 a year. Scaled across a 4,000-person engineering organization, that’s roughly $144 million in annual spend for one category of tool.
Meta, according to the article, has told employees internally that its own token costs could reach billions of dollars in 2026. Gorilla Logic’s Chief Growth Officer, Bob Graham (article’s featured photo), frames the broader lesson bluntly: “If the biggest technology companies in the world can be surprised by this expense, so can anyone.”
Token consumption, he notes, is a genuinely new expense category, one that didn’t exist on any P&L two years ago and doesn’t behave like the software spend finance teams spent decades learning to model.
Other data points around the industry back up the scale of the shift. KPMG’s Q2 2026 Global AI Pulse survey of more than 2,100 senior leaders found that only 7 percent of organizations have established measurable AI ROI, and 42% report only partial visibility into what they’re actually spending.
The FinOps Foundation’s 2026 State of FinOps report found that 73 percent of enterprises say their AI costs have exceeded original projections. And Goldman Sachs has projected total AI-related spending will top $800 billion globally in 2026, a figure that only makes sense once you understand how much of it is now usage-driven rather than subscription-based.
Gorilla Logic’s articles sorts corporate AI exposure into three tiers. At the bottom sit business users in sales, marketing, and operations, mostly on seat-based tools running $20 to $30 per user per month, and above that sit the agentic coding and engineering tools, where costs scale unpredictably with usage rather than headcount. And at the top is the kind of unmonitored, autonomous agent activity that produced Uber’s blowout: work that happens continuously, invisibly, and without a natural ceiling.
The article points to two emerging responses. One is the growth of AI cost-management infrastructure — agent gateways and orchestration layers that can route routine tasks to cheaper models, impose spending caps, and give finance teams real-time visibility into who is spending what. The other is architectural: pairing smaller, task-specific models with frontier models, so companies aren’t paying premium rates for work that doesn’t require a frontier model’s full capability.
What’s changed, in other words, isn’t just that AI has gotten more expensive. It’s that the cost has become genuinely difficult to see coming, and even harder to attribute once it arrives. Uber’s own leadership has acknowledged as much, with Naga’s line about being “back to the drawing board” was echoed weeks later by COO Andrew Macdonald, who admitted the company still can’t draw a clear connection between its AI spend and any specific feature shipped to riders.
For CFOs, that’s the harder problem sitting underneath the dollar figures. Controlling AI usage outright is achievable with caps and dashboards. Understanding which teams are driving the spend, whether that spend maps to any measurable business outcome, and whether the model paying for the work is even the right one. That’s the part nobody has fully solved yet, and it’s why token costs are now showing up in board meetings instead of IT budgets.
