AI token costs have gone from a minor technology detail to a board-level budget problem in under a year. Twelve months ago, “token cost” wasn’t a line item on most CFOs’ radar -now it’s showing up in all the wrong places: across the news, internal memos, budget freezes, and earnings calls.
Uber gave its engineers access to Claude Code in late 2025. By April 2026, the company had burned through its entire annual AI coding budget in four months. Its response was to cap spending at $1,500 per employee, per month, per tool, as Bloomberg first reported. Permitting the use of two agentic coding tools together (say Claude and Cursor) and that ceiling doubles to $3,000 a month, per engineer $36,000 a year for one person’s AI usage. It is estimated that Uber has approximately 4,000 engineers which would equate to $144M spend in AI.
Meta’s expense may be bigger. An internal memo warned that AI token costs were on track to hit billions of dollars in 2026. The reason sits in the mechanics of how agents work at scale. When an AI agent connects to a codebase to fix a bug or write software, it processes thousands of tokens of background context with every single prompt. Now put that inside an organization the size of Meta, add a corporate push for engineering to use Claude, and token usage does not just grow, it explodes.
According to Ramp’s AI Index, which tracks real vendor payments across more than 70,000 businesses, the top 1% of AI-adopting companies are now spending roughly $7,500 per employee, per month, on AI, and that spend is growing 14.1% month over month. Executives at AI recruiting startup Mercor have said publicly that AI token costs are starting to approach, or exceed, employee compensation at some organizations.
There is a lesson to be drawn here. If the biggest technology companies in the world can be surprised by this expense, so can anyone. Token consumption has become a genuinely new expense category, one that did not exist on any P&L two years ago. It does not behave like the software costs finance teams have spent decades learning to model and factor into traditional ROI calculations. Understanding why AI token costs behave so differently from traditional software spend is the first step toward getting them under control.
Why AI Token Costs Are So Hard to Predict
Traditional software budgeting runs on a simple assumption: cost scales with headcount. Ten more employees means ten more seats, at a known price. Easy to forecast and plan for as companies grow or license new software platforms.
AI tokens break that assumption. Cost scales with usage, not headcount, and usage is wildly uneven, hard to predict, and in the case of autonomous agents, not even tied to a human sitting at a keyboard. A single engineer running an agentic coding session can burn through thousands of dollars in a month with no warning, because the meter is running whether or not the output is useful. Many models also generate thousands of hidden, invisible tokens just to “think” step by step before answering, multiplying the cost per prompt.
That’s the part most organizations haven’t priced in: every dollar of new agentic AI spend has to come from somewhere. For most CFOs, that’s not a rounding error absorbed by an innovation fund, we are talking about millions of dollars in new spend that did not exist two years ago. This becomes a real trade-off against headcount, other technology investments, or discretionary budget elsewhere in the business.
The Three Tiers of Enterprise AI Spending Risk
Looking across how organizations are actually spending on AI, we can see a clearer pattern emerging. AI cost risk isn’t evenly distributed. It concentrates in specific places, and likely only one of the three is currently under control in a typical organization today.
Business users in sales, marketing, and operations largely run on tools priced per seat: $20 to $30 per user, per month, flat. This is the tier finance already knows how to budget, because it behaves exactly like every other SaaS line item they’ve purchased for the last twenty years.
Developers using agentic coding tools are where the Uber-style stories come from and the entire category just flipped from flat-fee to metered pricing. GitHub Copilot moved to token-based billing in June 2026, and developers have projected bills jumping from $29 to $750 a month, and from $50 to $3,000 a month in heavy agentic sessions. Uber isn’t an outlier here, it is an early look at where the whole category is heading.
Autonomous agents are the tier almost nobody has modeled yet. There’s no concept of a “seat” for an agent. It runs unattended, spins up sub-agents, retries on failure, and consumes tokens the entire time, whether or not it produces anything of value. This is the layer with the least cost data and the least budgeting precedent.
Put simply: the tier CFOs are most comfortable with is also the smallest financial risk. The two tiers actually blowing up budgets are the ones moving away from predictable pricing at the exact moment usage is scaling fastest.
Closing the AI Cost Management Gap
The common thread across Uber, Meta, and the wider Ramp dataset is a lack of visibility. Most organizations don’t have a layer between an employee or agent who opens a tool. Spending caps like Uber’s are a blunt, after-the-fact fix for a problem that really calls for something upstream.
That’s the gap a newer category of tooling often called agent orchestration, or agent gateways is built to address: routing each task to the cheapest model capable of handling it, enforcing spending limits before they’re breached, and giving finance a real-time meter instead of a monthly surprise.
This category is moving fast, and it’s already splitting into a few distinct layers. One layer is a cost controlled gateways which sits between an organization’s applications and its model providers, adding per-team spending caps, automatic routing to cheaper models for simple tasks, and response caching,the same levers Uber reached for manually when it imposed its cap. A second layer is agent orchestration which focuses on coordinating multiple agents working together, with cost visibility increasingly built in rather than bolted on. And a newer, finance-facing layer is starting to treat AI spend the way FinOps teams have long treated cloud spend: attributed by team and feature, budgeted, and reviewed monthly rather than discovered after the fact.
Then there is an architecture approach that is aimed at limiting spend by reducing reliance on large LLMs. Some companies have begun creating a hybrid SLM/LLM architecture where small language models are trained on a narrow set of corporate tasks (e.g. internal policy, customer service, loan document processing, parsing shipping manifests). The economics can be compelling. According to a 2026 enterprise LLM comparison, at one million monthly conversations, hosted frontier LLM costs run $15,000 to $75,000 per month, while a fine-tuned small language model deployed in-house costs $150 to $800 per month at the same volume. Goldman Sachs, for example, manages AI spend at scale through a centralized model gateway that classifies each request by task type and routes it to the most cost-effective model for the job, from lightweight open-source models to frontier LLMs, all within its own security perimeter.
The common thread is that most of today’s tooling was built by and for engineers, to turbo charge their productivity via usage of agents. The layer most organizations are still missing is the one built to answer a CFO’s question directly: which team, which budget, how do we forecast expenses and what is the business use case supported by this spend? That’s the gap likely to matter most as agentic AI spend keeps compounding.
Two years ago, none of this existed as a line item. Today, AI token costs are one of the fastest-growing and least-understood items on the balance sheet, and AI cost management is quickly becoming a core finance discipline rather than a side project. We saw a similar story when cloud first got adopted and costs exploded during that first wave. Organizations planning agentic AI rollouts would do well to have this conversation between engineering and finance now and come up with an approach to manage. Uber and Meta have already shown what happens when you have it after the fact instead of before.
