You're Not Paying for AI. You're Running Unmetered Infrastructure.

Uber's CTO said it plainly: "I'm back to the drawing board, because the budget I thought I would need is blown away already."

This wasn't a startup that moved recklessly. Uber has 5,000 engineers, a mature FinOps practice, and professional procurement. They adopted AI coding tools aggressively — Claude Code adoption went from 32% to 84% of their engineering org in three months. By April 2026, the entire annual AI budget was gone. The calendar still said Q2.

One company. Entire annual AI budget. Four months.

The same pattern showed up at Microsoft, which quietly revoked its own developers' Claude Code licenses after token bills spiraled past what anyone had modeled. One organization reportedly ran up a $500 million Claude bill after forgetting to set usage caps. These aren't cautionary tales about recklessness. They're the predictable outcome of buying AI the wrong way.

The Pricing Model Nobody Planned For

For two decades, enterprise software had a comfortable pricing structure: pay per seat, sign a multi-year contract, model costs as headcount scales. It was linear, predictable, and easy to budget. Your CFO could approve it on a spreadsheet. Your procurement team could benchmark it against last year.

AI broke that model and almost nobody noticed until the bill arrived.

Token-based pricing is consumption infrastructure, not a software license. The price per unit has been falling — down roughly 67% year over year on blended enterprise API costs between early 2025 and early 2026. That number gets cited in every vendor deck. What doesn't get cited is the other half of the equation: volume has grown faster than price has dropped. Total spend equals price times quantity. The first variable is falling. The second is growing at a pace no traditional budget cycle accounted for.

Agentic workflows are the multiplier. A simple chatbot query uses around 1,000 tokens. A RAG-based lookup uses 10,000. A multi-agent workflow that reasons, calls external tools, verifies outputs, loops back for corrections, and hands off to another agent? That can consume 500 times more tokens than the chatbot the same team was running 18 months ago. One analysis put the cost of a single complex orchestrated interaction in 2026 at 30 times what the same workflow cost in 2023.

Per-token costs fell 98% since early 2024. Enterprise AI bills are still rising. That is not a paradox. It's math.

The Agentic Tax Nobody Put in the Business Case

Here is what a single agentic interaction actually costs at the token level in 2026: a system prompt, user instructions, retrieved documents from your knowledge base, outputs from tool calls, conversation history from earlier in the session, and intermediate reasoning steps. Before the model generates a single word of response, you've already consumed tens of thousands of tokens. At frontier model pricing, that input context alone costs more than an entire 2022 chatbot conversation.

Now multiply that by every agent running in the background. The monitoring agents watching your logs. The compliance agents scanning document updates. The developer assistants running autonomously while engineers are in meetings. These workloads don't wait for a user to press enter. They run continuously, consuming compute against every event and data change, regardless of whether a human asked for anything.

One healthcare enterprise burned through more than $6 million in unplanned AI costs over six months from a single deployment — before the finance team even understood what was driving the bill.

When per-developer consumption is rising 18 times over nine months, the ROI case built on last quarter's numbers is already wrong.

What the Cloud Taught Us That We're Ignoring

We've seen this movie before. Early cloud adoption followed the same arc: teams provisioned what they needed, bills arrived that nobody expected, finance got involved, and a whole discipline called FinOps emerged to bring structure to what had become unmanaged infrastructure spend.

By 2026, 98% of FinOps practitioners are responsible for managing AI spend. That number was 31% just one year ago. The function that spent a decade governing cloud waste is now being handed a cost structure it has no established playbook for — consumption-driven, token-based, and architecturally volatile in ways that cloud compute never was.

The difference with AI is that the consumption compounds in ways cloud didn't. A cloud VM you forgot to turn off costs you one VM. An AI agent running in an unconstrained loop costs you every step of every reasoning chain, plus the tool calls, plus the retrieval overhead, plus the context that gets re-sent on every iteration. Gartner's 2026 AI Hype Cycle report projects that 40% of AI agent projects will be cancelled by 2027 — not for technical failure, not for poor market fit. Just economics.

The governance discipline exists. Companies just haven't applied it yet.

The Question Leaders Are Avoiding

OpenAI's CEO called the ROI question "the most fair criticism right now of AI." He acknowledged that customers are burning through their entire 2026 budgets and that cost concerns have become the second-most common issue he hears. When the person selling you the product concedes that the value question is fair, the value question deserves a serious answer.

Only 12% of CEOs say AI has delivered both cost and revenue benefits. Despite that, companies have committed $740 billion in AI-related capital expenditures this year. Deloitte found that 50% of enterprise leaders are spending between 21% and 50% of their entire digital transformation budgets on AI.

The pressure to use AI is coming from boards, from CEOs, from investors who want to see the strategy. What isn't coming from the top is the financial architecture to govern it. The deployment decision preceded the cost model at Uber. It preceded the cost model at most enterprises scaling agents right now. That sequence — ship, then measure — works at chatbot scale. At agentic scale, it's a structural problem that compounds with every renewal cycle.

What Actually Changes the Math

The economics of AI are not permanently broken. They just require governance that most organizations haven't built yet.

The highest-leverage change is also the simplest: stop routing every task to frontier models. There is roughly a 180x price difference between routing a task to a budget-tier model versus a frontier reasoning model. Most enterprise workloads — document summarization, routine triage, boilerplate code generation — do not require the most powerful model available. Organizations that route intelligently based on task complexity consistently cut costs by 60-80% without measurable quality degradation.

Prompt caching, context window management, per-user budget caps, and hard agent loop limits follow from there. These are the same governance moves cloud FinOps teams made starting in 2018. They work. Most agentic deployments just haven't implemented them because the team that deployed the agents and the team that owns the budget are different teams, talking monthly, if at all.

AI will get cheaper. The Gartner projection that inference costs drop 90% by 2030 is probably directionally right. But cheaper per-token costs won't solve the agentic consumption problem, because agentic models require more tokens per task as they get more capable, and enterprises will deploy them more broadly as they prove useful. The volume will keep growing. The governance has to come first.

The bill for ignoring that is already sitting in someone's inbox.