The narrative around AI and cost savings is hitting a wall.

Bryan Catanzaro, VP of Applied Deep Learning at Nvidia, told Axios this week: “For my team, the cost of compute is far beyond the costs of the employees.” Coming from an executive at the company whose chips power most of the world’s AI infrastructure, that admission carries weight.

The same week, reporting from The Information revealed that Uber’s chief technology officer had burned through his entire 2026 AI budget months before the year is over, wiped out by token costs from running AI agents at scale.

These aren’t edge cases. They reflect a broader pattern playing out across enterprise AI deployments right now.

The Variable Cost Problem Nobody Planned For

Human workers have predictable costs. You hire someone at $80,000 a year and you know what you’re paying. AI agents don’t work that way. Every query has a cost. Every task completion, every document processed, every customer interaction handled by an agent adds to a bill that compounds with scale.

When companies ran pilots with 10 or 50 users, the numbers looked great. Then they scaled to thousands of users and thousands of daily agent interactions, and the economics changed completely.

Global IT spending is projected to hit $6.31 trillion in 2026, up 13.5% from last year according to Gartner. A large portion of that increase is going directly to AI compute. The chips, the tokens, the cloud inference costs. For many organizations, this is now the fastest-growing line in the technology budget.

What the Research Actually Shows

An MIT study on AI automation viability found that AI would be economically viable to deploy in only 23% of roles where vision tasks are the primary workload. In the remaining 77% of cases, it was still cheaper for humans to do the work.

That doesn’t mean AI isn’t worth deploying. It means the business case requires more precision than most companies have applied.

The organizations getting genuine returns from AI right now tend to share a few characteristics: they identified specific, high-volume processes with clear measurable outputs; they built in cost monitoring from day one; and they didn’t treat “AI everywhere” as a strategy.

This Isn’t Failure. It’s a Calibration Problem.

There’s a difference between AI not working and AI being deployed without a cost model. Most of what we’re seeing falls into the second category.

Companies rushed to deploy agents because the productivity gains looked compelling in demos. What they didn’t model was the operational cost at scale, the hidden complexity of integrating agents into real workflows, or the supervision burden that doesn’t disappear when you automate a task.

Uber’s CTO blowing through his AI budget isn’t evidence that AI agents fail. It’s evidence that enterprise AI deployment requires the same financial discipline as any other major technology investment. You need to know your cost per transaction, your break-even volume, and your ROI threshold before you scale.

Nvidia’s Catanzaro framed it as a feature, not a bug, noting that companies can scale work with compute in ways they couldn’t with headcount. But that only holds if the work being automated is generating more value than the compute costs to run it.

What This Means for Business

If you’re evaluating AI investments or already running agents at scale, these numbers are a signal, not a reason to pull back.

The companies that will come out ahead are the ones treating AI deployment like a financial instrument. That means modeling the cost curve before you commit, benchmarking your cost per outcome against the human alternative, and building in regular ROI checkpoints as you scale.

The businesses that will struggle are the ones that moved fast based on demos, piled on subscriptions, and assumed the savings would be obvious. For many of them, the compute bill has quietly grown to the point where it rivals or exceeds what they were paying the people they replaced.

Data literacy plays a direct role here. Understanding your token costs, your agent completion rates, your cost-per-workflow, and how those numbers change at different usage volumes is now a core business skill. It’s not IT’s job to track this in isolation. Finance, operations, and leadership all need to be looking at the same numbers.

Strategic AI deployment isn’t about having the most agents running. It’s about knowing exactly which processes justify the compute cost and which ones don’t yet.

---

If this is the kind of problem agents can help with, the free Working With Claude field guide is the practical next step. Thirty-two pages, no fluff. Get the free guide.