Uber and AI: When Token Billing Goes Off the Rails

The Illusion of Cost Control Through Usage-Based Billing

The initial promise of usage-based billing in the field of artificial intelligence was to control expenses. However, this approach now seems to contribute to an uncontrollable surge in costs. Infrastructure could be the key to preventing AI from becoming trapped in an economic bubble.

In April 2026, Uber revealed that it had consumed its entire annual budget dedicated to artificial intelligence just four months into the fiscal year. This event surprised many observers, as Uber is a company deeply rooted in digital technology, supposedly immune to such financial slip-ups. What is even more striking is that Uber's executives could not associate this increase in spending with any tangible improvement in the services offered. The problem thus extends beyond the mere high cost of AI: it is a cost that does not translate into identifiable added value.

This phenomenon highlights a structural issue related to the business model that has emerged with the rise of generative AI: usage-based billing, often referred to as "token billing."

A Business Model Disconnected from Real Value

Within this model, costs are directly linked to the volume of text processed by the AI model, without considering the results achieved. A simple prompt can trigger a high-value decision, while a longer and more expensive request may yield nothing significant. However, pricing does not reflect this distinction: the generated value and the cost follow parallel trajectories that never intersect.

To regain control, companies often implement token quotas per user. But this solution introduces a new distortion: an employee with legitimate needs may find their work limited, while another may exhaust their quota without producing significant results. Thus, scarcity is imposed where value management is needed, and the only possible response is to defer demand without assessing its relevance. The quota does not reflect the actual utility of the usages it restricts.

A System Favoring Providers, but Not Viable in the Long Term

Originally, this type of billing was designed to maximize profits for large providers, rather than to meet the real needs of their clients. The industry does not hide this fact: artificial intelligence is intended to become a resource billed by consumption, much like electricity.

However, this approach has proven not to be sustainably profitable for those who apply it. OpenAI has admitted to losing money on its most expensive subscription, with usage far exceeding forecasts. For 2024, the company anticipates revenues of $3.7 billion but losses nearing $5 billion. No major player has yet demonstrated the profitability of this business model. The conclusion is clear: such an imbalance cannot persist, and an adjustment is necessary.

Towards a New Unit of Measurement: Infrastructure

If consumption is not an adequate billing basis, an alternative exists, more traditional in the realm of computing tools: infrastructure. Instead of counting tokens, costs could be indexed to the computing capacity used.

This change would be radical. Users could access AI without volume constraints, with constant costs, making expenses predictable. Companies could thus size, plan, and control their spending, just as they do for storage capacity. The trade-off is clear: it is no longer about managing usage limits, but about managing performance. In times of high demand, processing may take longer, but this would allow for the preservation of usage and experimentation, rather than deferring tasks to the next day. Moreover, keeping models and data in a controlled environment would better protect sensitive information.

The Importance of Flexibility and Openness

What strategy should be adopted today? It is crucial not to lock into a single solution: in such a dynamic field, what works today may be obsolete tomorrow. A model must be evaluated not in a laboratory, but based on what it brings to users. Open and reversible solutions are therefore essential, as they allow for changing one's mind without excessive cost. It is no coincidence that innovation in AI is largely driven by open-source software: this shows that openness remains the healthiest path.

In many ways, the evolution of AI resembles that of the Internet in the early 2000s: a true revolution, massive investments, and a bubble ready to burst. Some players will disappear, but the technology will establish itself permanently, becoming an omnipresent commodity. As long as standards are not stabilized, caution dictates that one should have the means to quickly and penalty-free redirect choices. Mastering the cost of AI does not hinder innovation; rather, it is a sine qua non condition for long-term sustainability.