For two years, artificial intelligence has established itself as an almost unlimited resource. Accessible chatbots, low-cost APIs, on-demand content generation: the illusion of abundant intelligence has structured adoption. But this phase is coming to an end, because behind the rise in uses, the cost of computing is about to constrain the entire sector.
Demand that exceeds infrastructure
The use of AI models is growing exponentially. At OpenAI, token consumption via API increased from 6 billion to 15 billion per minute in a few months. This dynamic is not linked to a simple increase in the number of users, but to a change in usage.
AI is no longer used for one-off requests. It now orchestrates complete tasks via autonomous agents: code generation, workflow automation, interaction with third-party systems. Each use multiplies the consumption of resources, and an agent can consume several dozen times more compute than a classic chatbot.
This development comes even as infrastructure capacities remain rigid. The construction of data centers, access to energy and the production of semiconductors impose incompressible deadlines, which means that demand exceeds supply.
the return of an economy of scarcity
In this context, signals of tension are multiplying and the rental prices of GPUs, the heart of AI computing, are increasing rapidly. The latest generations of NVIDIA chips are recording significant increases in the spot market, with certain configurations seeing their hourly cost increase by almost 50% in a few weeks.
Infrastructure providers are adjusting their strategy. CoreWeave has raised its prices by more than 20% and now imposes contractual commitments over several years. This means that for companies that have a structuring need for AI, computing is no longer a flexible commodity, but a resource to be secured.
At the same time, AI players are arbitrating in order to reduce uses. OpenAI has suspended certain developments, particularly around video generation, to reallocate its capacities towards uses deemed more critical, such as code or enterprise applications.
The implicit end of free
Until now, the ecosystem has largely subsidized usage, with models accessible at low cost, or even free, to accelerate adoption and capture market share. This logic is now reaching its limits.
The token, a unit of measurement for AI consumption, is now establishing itself as a real economic unit. The more uses become complex, the more the bill increases and the generalization of agents accentuates this phenomenon by transforming AI into an active system, consuming computing continuously.
In this context, rising prices become difficult to avoid, but it places players in a delicate situation, because increasing prices risks slowing down adoption, even though competition remains intense.
Quality of service still unstable
The tension on capacities also results in a deterioration of service. At Anthropic, interruptions are increasing, with an availability rate lower than usual SaaS standards. Some client companies have already started to arbitrate between suppliers to guarantee the continuity of their services.
This point is structuring, because AI is becoming a critical layer of information systems, without yet offering the guarantees of reliability necessary to be part of an industrial deployment. The gap between technological promise and infrastructural maturity remains significant to this day.
An industry that is changing in nature
Beyond economic tensions, it is the very nature of the market that is evolving. Artificial intelligence is not just a software product, it relies on heavy infrastructure, combining data centers, energy and advanced components, the availability and prices of which can vary significantly.
This transformation brings AI closer to industries historically constrained by their resources, where production capacity determines growth, and in this model, the competitive advantage no longer resides solely in the quality of the models, but in access to computing.
Towards a new discipline of uses
For companies, this development requires a change of posture, and AI can no longer be consumed without arbitrage. Each use has a cost, each automation has a compute footprint.
In the short term, this will result in:
- prioritization of high-value use cases
- optimization of queries and architectures
- diversification of suppliers to limit risk
In the longer term, several questions arise: how much are organizations willing to pay to automate their processes? How can we integrate a resource into our economic model for which we do not yet know the true price? And above all, how to arbitrate between performance and cost in a context where each gain in productivity is based on increased consumption of computing? Finally, another more structural question emerges: who, tomorrow, will capture the value, the companies that use AI, or those that control the infrastructure that makes it possible?
The era of abundant AI is reaching its limit. The rise in uses, combined with the physical constraints of the infrastructure, is giving rise to an economy of scarcity. In this new context, computing becomes the central resource, and its cost, the key variable.
The promise of intelligence accessible to all remains. But it will now have to deal with a simpler reality: producing intelligence has a price, and to date, this price is increasing.
