Pay-per-use billing was supposed to control AI costs, however, it is making them out of control. Here’s why infrastructure could save AI from its own bubble.
In the spring of 2026, Uber admitted to having exhausted its entire annual budget ofartificial intelligence from mid-April, four months after the start of the financial year. The episode left its mark, because it concerns a company born in digital, familiar with these tools and which one would have thought immune to such a slippage. Even more revealing, its leaders recognized that they were unable to link this surge in consumption to a tangible improvement in the services provided. The problem is therefore not only that AI is expensive: its cost is simply no longer linked to any value.
This observation refers to a more structural cause than it seems: the economic model that has become dominant with the rise of generative AI, billing based on usage, known as “token based”.
The original defect: a price that ignores value
In this model, the cost is strictly indexed on the volume of text submitted to the model, and never on the result obtained. A short prompt may trigger a high-value decision, while a long, costly query may produce none. However, the price takes no account of this difference: the value created and the invoice follow two parallel trajectories which never come together.
To regain control, companies most often resort to token quotas per user. However, this safeguard introduces an additional distortion: an employee with legitimate needs finds himself restricted in full activity, while another exhausts his envelope in uses without real impact. We then administer scarcity where value should be managed, and the only possible response amounts to deferring the request without assessing its merits. The quota, in fact, says nothing about the real usefulness of the uses it restricts.
A model designed for the supplier, and yet loss-making
Originally, this billing method primarily serves the monetization of large suppliers, rather than aligning with the real interests of their customers. The sector also makes no secret of it, intelligence being called upon to become a resource billed on consumption, in the same way as electricity.
Even more striking, this mechanism does not benefit those who practice it in the long term. OpenAI admitted losing money on its most expensive subscription, usage far exceeding its forecasts, and posted revenues of around $3.7 billion for 2024 for nearly $5 billion in losses. To date, no major player has demonstrated the profitability of this model. The consequence is therefore self-evident: such a balance is not tenable, and a rebalancing will eventually occur.
Change unit of measurement
If consumption constitutes a poor basis for invoicing, another exists, much more classic in the provision of IT tools: infrastructure. Rather than counting the tokens, we index the cost on the computing capacity mobilized.
The change is then profound. Users access AI without volume constraints, at constant cost, and the expense becomes predictable again: the company sizes it, plans it and controls it, exactly like storage capacity. The counterpart is assumed: we no longer manage usage limits, but performance. In case of high traffic, processing may take five minutes rather than two. However, between occasionally slowing down a service and sending a colleague back the next day, the first choice preserves usage and experimentation. Added to this cost control is a related benefit: keeping models and data in a controlled environment also means keeping control of sensitive information.
Maintain the right to change your mind
What posture should we adopt today? First, do not lock yourself into an exclusive solution: in such a changing field, what is required this quarter can be exceeded the next. A model is not judged on a test bench, but on what it brings to those who use it. There then remain open and reversible solutions: the only assurance of changing your mind without paying a high price. It is no coincidence that innovation around AI remains largely driven by free software: it is a sign that openness remains the healthiest trajectory.
In many ways, the trajectory of AI is reminiscent of that of the Internet in the early 2000s: a real revolution, outsized investments, and a bubble waiting to burst. Some players will disappear, but the technology will establish itself over the long term, until it becomes a convenience present in all uses. As long as standards are not stabilized, prudence dictates only one thing: giving yourself the means to reorient your choices quickly, and without penalty. Controlling the cost of AI is not to restrict innovation, but rather the condition for making it sustainable.
