Token AI: why artificial intelligence costs are reshaping corporate budgets

AI tokens have become one of the most underestimated costs of enterprise AI adoption. Within months, they turned into a budget line item no CFO had anticipated. Uber burned through its entire 2026 AI budget in the first four months, while Royal Bank of Canada reported a 500% surge in token consumption over a six-month period. The problem is not just the amount spent, but the speed at which costs can balloon when AI is deployed across complex, multi-step processes that are hard to track.

What an AI token is

A token is the smallest unit through which AI models—such as those behind ChatGPT, Claude, or Gemini—read and generate text. It does not necessarily correspond to a full word: it can be a character, part of a word, or a punctuation mark, depending on the language and the tokenization system used. Before processing a request, the model breaks the text into these basic elements and generates numerical vector-based representations, which it uses to build a coherent response.

This technical step has a direct economic consequence: every interaction with an AI model is measured and billed based on the number of tokens processed, both input and output. The longer and more complex the conversation, the more tokens are consumed and the higher the cost. This weighs even more when AI does not simply answer a single question but performs complex multi-step activities.

Why token-based billing changes the rules

The mechanism that surprised many CFOs is structural: with traditional software licensing, costs scale with the number of employees, whereas with tokens they scale with usage intensity. Agentic AI, which autonomously executes multi-step tasks, multiplies consumption because every search, draft, and feedback cycle generates new tokens—often without direct oversight. A healthcare company accumulated over 6 million dollars in unplanned costs by consuming one trillion tokens in six months, before the finance team managed to identify the source.

Model switching also significantly impacts the bill. When a team moves from a lighter model to a frontier one for quality reasons—often without involving those who manage budgets—the cost per token can increase by 10 to 100 times, with differences reaching up to 25 times more expensive output between a standard and a premium model of the same family. In this scenario, monthly spending fluctuations of 40% or more are now considered normal, even with a stable workforce.

The paradox explaining why costs rise while prices fall

The price per million tokens has dropped by roughly 100 times over the past three years, yet corporate AI spending continues to grow faster than prices fall. The reason has a name from 19th-century economics: the Jevons paradox. According to this principle, when a resource becomes more efficient and cheaper, overall demand does not decrease—it tends to multiply, because efficiency enables previously unthinkable uses.

Applied to AI, this means that a task requiring 500 tokens in 2023 may now consume 50,000, because agentic systems no longer produce a single response but chains of reasoning, checks, and multiple model calls. Gartner estimates, based on industry analysis, that despite inference costs dropping nearly 90% by 2030, enterprise AI will not automatically become cheaper, because consumption will grow faster than price reductions.

The Uber case: when the budget collapses in four months

Uber’s CTO, Praveen Neppalli Naga, stated in April that tokens consumed by agentic coding tools—mainly Cursor and Claude Code—exceeded every spending forecast. Before limits were introduced, individual engineers generated monthly token bills ranging from 500 to 2,000 dollars, and the CTO himself burned 1,200 dollars in a two-hour internal demo. To bring spending back under control, the company set a 1,500-dollar monthly cap per employee for each tool and introduced an internal dashboard to monitor consumption in real time.

The KPMG survey: companies that don’t know how much they spend

A KPMG survey shows that only 26% of companies truly understand how much they spend on AI, while 50% have very limited visibility and 22% discover costs only upon receiving the invoice. A 2025 poll confirms the same difficulty: 85% of companies miss AI spending forecasts by more than 10%, and nearly a quarter underestimate them by 50% or more. Around three hundred companies addressed AI cost issues in their April and May 2026 quarterly reports, compared to 93 the year before.

The problem is not limited to Uber. The CEO of Royal Bank of Canada stated that the bank increased token consumption by 500% in just six months. Another company spent 500 million dollars in a single month because no usage limits had been set.

From technology to governance: the role of CFOs and boards

For many companies, the first month of high bills was dismissed as an exception; the second—when spending did not return under control on its own—forced the real question about the value generated relative to what is being spent. Individual monitoring alone is not enough to solve the problem, because acting on that data would mean daily micromanagement that is unsustainable over time.

There is also a regulatory front entering public financial statements: for listed companies, token spending may fall under SEC disclosure requirements in the Management’s Discussion and Analysis section of quarterly and annual reports, if the company tracks token consumption as a management indicator relevant to investors.

The three levers to bring costs back under control

Effective responses share a common logic: treating AI as a finite and strategic resource, as suggested by Deloitte, with the same rigor historically applied to energy and allocated capital. The approaches observed move along three directions:

Request-level cost attribution, to link spending to specific workflows, teams, and use cases, following the same chargeback logic already applied to cloud computing.
Governance and spending limits, such as Uber’s monthly caps per employee, the use of smaller models for simple tasks, and approval flows to exceed limits.
Alignment between technology and finance, to build a shared process for spending control, with real-time visibility into consumption and explicit criteria for which use cases deserve budget priority.

A long-term direction that goes against cost-cutting

Despite the alarms, token prices per million dropped by about 80% between early 2025 and early 2026, while enterprise consumption grew 13-fold over the same period. Gartner predicts that over 40% of agentic AI projects will be cancelled by 2027, largely due to high costs and lack of measurable returns. The overall signal is that companies that continue treating tokens as an after-the-fact expense—rather than a resource to plan—will discover the real cost of their AI adoption only when it is too late to govern it.

Token control is now a matter of budget, governance, and strategy, no longer a technical detail.

Sources: