The Cost of the Cognitive Illusion
The Sora model, developed by OpenAI, was released to the public in November 2025 and shut down after six months. Not due to security issues, but for operational sustainability reasons. The system generated high-quality videos with an estimated energy consumption of 120 megawatt-hours per minute of output. This is not an isolated case: the same architecture that enables detailed descriptions of unseen images (mirage reasoning) requires a compute consumption that is no longer sustainable for the existing infrastructure. The phenomenon is not a technical error, but a symptom of a structural tension: the ability to generate intelligent content is now decoupled from the ability to execute it in real time.
This implies that the AI paradigm is no longer about competition between models, but about competition between computing systems. The most sophisticated model does not win if it cannot be executed. In other words, inference efficiency, not model complexity, determines scalability. This implies that the true frontier of innovation is not in algorithm design, but in the logistical control of computing chips.
The Architecture of the Bottleneck
The scarcity of computing chips is a physical constraint, not a market problem. Nvidia currently holds 80% of the global GPU market for AI, with a technological lead that cannot be reduced in less than three years. This monopoly creates a bottleneck: every attempt to develop an advanced inference model is conditioned by the availability of hardware. The cost of running a model is no longer determined by its complexity, but by its dependence on rare chips.
A recent study conducted by Stanford, UC Berkeley, CMU, and Microsoft Research found that the model chosen to be “78% more economical” in terms of price per token can actually be 22% more expensive. This phenomenon, called Price Reversal, is caused by a failed optimization: low-cost models require more iterations, more temporary memory, and more inference steps to achieve the same result. The actual cost is not in the list price, but in the compute consumption during execution. The operational consequence is that inference efficiency is not a secondary metric, but the decisive factor for economic sustainability.
The Imperfect Symbiosis Between Technology and Power
“Inference compute will shape AI’s future,” said Mustafa Suleyman, CEO of Microsoft AI. The statement is not an opinion, but a market observation: whoever controls the flow of chips controls access to synthetic thought. The $830 million funding for Mistral’s data center, which includes the purchase of 13,800 Nvidia GPUs, is an example of this dynamic. The funding was obtained from a consortium of French and international banks, but the real value lies in the physical control of computing units.
“Frontier models readily generate detailed image descriptions without visual input. We term this phenomenon mirage reasoning.” — Gary Marcus, researcher
The quote reveals a systemic tension: models are not intelligent, but simulate intelligence through patterns. This simulation, however, requires an energy consumption that is no longer sustainable for the existing infrastructure. The data reveals a structural tension: AI is not limited by its cognitive capacity, but by its dependence on scarce physical resources. Expectations of autonomous AI are incompatible with the technical reality of a system based on limited computing chips.
Scenario: The Cost of Thought
By the next election cycle, the cost of running an inference model for an average company will be higher than the cost of developing it. This is not a hypothetical future: it is already happening. Companies that cannot obtain access to low-cost computing chips will be forced to reduce the use of inference models, even if these are more efficient. The systemic cost is not only financial, but of access to synthetic thought.
Who will pay this cost? Not the end users, but the companies that invest in AI. Computing costs are no longer an input, but a strategic asset. Investment decisions will no longer be based on the quality of the model, but on the ability to access chips. The future of AI is not one of an intelligent entity, but of a logistical control system. The real power is not in the model, but in the chip that runs it.
Photo by Shubham Dhage on Unsplash
Texts are autonomously processed by Artificial Intelligence models