Per-token AI pricing has dropped 98% since 2022. Enterprise AI bills have tripled in the same period. That is not a pricing problem. It is an operations failure.

Meta sent an internal memo to 6,000 employees in late June warning that AI usage costs were approaching billions. The company tracked consumption on an internal leaderboard called “Claudeonomics.” Employees consumed 73.7 trillion tokens in roughly 30 days. Meta is now dismantling the leaderboard and replacing it with a centralized monitoring platform called AI Gateway to track usage and spending across teams in real time.

Uber hit the same wall earlier. The company exhausted its entire 2026 AI coding budget by April after roughly 5,000 engineers pushed token consumption past every projection the finance team had modeled. Monthly costs per engineer ranged from $500 to $2,000 for power users. Uber responded with a $1,500 per month cap per tool. Walmart, Amazon, and Cisco followed with similar controls.

The FinOps Foundation’s 2026 State of FinOps report found that 73% of enterprises reported AI costs exceeding original projections.

The 4,500x problem

The gap between the cheapest and most expensive AI models available today is roughly 4,500x. Using a frontier model for every task is like booking a private jet for every commute. The trip gets done. The budget does not survive the quarter.

This is what the industry is calling “token maxing.” Organizations give employees access to the most capable model for every task, with no routing logic, no tiering, no cost visibility. An engineer running autocomplete suggestions consumes a fraction of what an engineer orchestrating parallel agents across a codebase will consume. Same tool. Same person. Same workday. Wildly different invoices.

Agentic AI workflows consume 5 to 30 times more tokens than a standard chatbot interaction. When organizations deployed these tools, they budgeted for the chatbot version. They got the agent version. Nobody updated the forecast.

Average annual enterprise AI budgets grew from roughly $1.2 million in 2024 to $7 million in 2026. Per-developer usage is up 18.6x in under a year. The pricing model changed, the consumption pattern changed, and the finance models stayed frozen in 2024.

This is cloud spend all over again

Anyone who managed cloud infrastructure between 2018 and 2022 recognizes this pattern. Companies moved workloads to AWS and Azure. Monthly bills exploded. Not because the cloud was expensive per unit, but because nobody governed how it was used. Engineers spun up instances with no shutdown policies. Dev environments ran 24/7 for workloads that needed 8 hours. The unit cost was fine. The consumption was ungoverned.

That era produced an entire discipline: FinOps. Cost allocation. Usage policies. Spend alerts. Tagging requirements. None of it was technically difficult. All of it was operationally necessary.

AI is repeating the same cycle, compressed into months instead of years. The companies that already have model routing policies, usage tiers, and spending dashboards are operating within budget. The companies that handed every employee a frontier model with no guardrails are the ones writing emergency memos.

What the spending caps actually reveal

Meta did not cap spending because AI is too expensive. Meta capped spending because it had no system to route usage to the right model for the right task at the right cost. The AI worked. The operations around it did not exist.

Uber did not burn its budget because Claude Code costs too much. Uber burned its budget because 5,000 engineers used it without any consumption framework, and the finance team was still modeling costs based on per-seat licensing assumptions from a pre-agentic world.

The organizations that built the operations layer before scaling access are not in this story. You do not hear about their budget emergencies because they do not have budget emergencies. They have model routing. They have tiered access. They have cost visibility at the team level. They treated AI spend as an operations problem from day one.

The question that matters now

Every enterprise increasing its AI budget this year should be asking: do we have the operations layer to govern how that budget gets consumed?

Not which model is best. Not which vendor to choose. Not whether AI is worth the investment. Those questions have answers. The question that does not have an answer at most organizations is simpler and more urgent: who owns AI cost governance, and what system are they using to enforce it?

The model is not the line item that will surprise your CFO. The missing operations layer is. And unlike the model, that one is entirely within your control.