Behind the enthusiasm for generative AI, an economic question is beginning to emerge: what if some automations ultimately cost more than an employee?
For two years, generative artificial intelligence has established itself as the new cognitive infrastructure of the global digital world. Companies are experimenting with copilots, autonomous agents, conversational assistants, and automated production chains capable of writing, synthesizing, coding, translating, or analyzing massive volumes of information. The dominant narrative rests on a simple promise: increase productivity while reducing human costs.
But a more discreet debate is beginning to emerge in academic, financial, and industrial circles. What will happen when hyperscalers charge the true cost of generative AI? What if certain tasks ultimately become more expensive to automate than to entrust to a human?
This question is no longer marginal. It now runs through MIT analyses, Sequoia Capital strategic notes, statements from Nvidia executives, as well as reflections on inference economics and the energy sustainability of large models. Could generative AI reveal an unexpected paradox: that of an extraordinarily powerful technology… but economically difficult to scale?
The Myth of Naturally Profitable Automation
In the collective imagination, technological automation mechanically reduces costs. This logic has structured several decades of industrial computerization. Yet generative AI introduces a fundamental break: unlike traditional software, a large language model does not produce a “free” response.
Each query mobilizes extremely expensive GPUs, cooling infrastructures, high-bandwidth memory, energy-intensive networks, as well as massive storage systems and complex software orchestration mechanisms. In other words, generative AI has an ongoing operational cost.
This distinction is essential. In traditional software, the marginal cost of a query tended toward zero. In generative AI, each interaction has a real computational cost. This is precisely what several recent studies are beginning to document.
MIT Questions the Economics of AI Automation
One of the most important works on the subject was published by MIT CSAIL MIT CSAIL – Rethinking AI’s impact: economic limits to job automation
The associated scientific paper, titled Beyond AI Exposure: Which Tasks are Cost-Effective to Automate with Computer Vision?, analyzes the real profitability of AI automation.
The researchers do not simply evaluate the technical capabilities of AI systems. They introduce a much more strategic dimension: the complete economic cost of human replacement. Their conclusion has sparked numerous debates: only about 23% of wages associated with the analyzed visual tasks would be economically profitable to automate today.
This conclusion deserves to be interpreted rigorously. The study does not cover all white-collar jobs, but primarily computer vision tasks. Nevertheless, it reveals a fundamental tension: technological feasibility does not guarantee economic profitability.
In other words, just because an AI can replace a task doesn’t mean it’s economically rational to do so. This nuance upends the dominant narrative around massive replacement of human labor.
Inference: The Hidden Cost of Generative AI
For several years, attention focused on the training cost of models. The media discussed the billions needed to build GPT-4, Gemini, or Claude. Yet another phenomenon is beginning to worry industrialists: the cost of inference.
Inference refers to the moment when the model is actually used. Each prompt triggers extremely intensive computational operations. The more sophisticated the uses become, the more costs explode. Autonomous agents, multi-agent workflows, advanced reasoning, long contexts, persistent memory, connected tools, and RAG architectures significantly increase computational needs. The problem then becomes structural: generative AI is not only expensive to build, it is also expensive to use daily. This reality explains the massive investments by hyperscalers in GPU infrastructures.
The Hyperscaler Question: An Economy Still Under Pressure
Since 2023, major cloud players have been investing historic sums in AI infrastructures. Microsoft, Google, Amazon, and Meta are dedicating tens of billions of dollars to data centers, GPU clusters, specialized chips, and energy networks.
This dynamic has been extensively analyzed by Sequoia Capital. AI’s 600 $ B Question — Sequoia Capital. Then in a second analysis: AI Is Now Shovel Ready – Sequoia Capital
David Cahn poses a question that has become central in Silicon Valley. Will the revenue generated by AI be sufficient to amortize the colossal investments made in infrastructures? The issue is no longer just technological. It becomes macroeconomic. Hyperscalers seem engaged in a race where each must invest massively to avoid being overtaken. But no one yet knows precisely at what horizon this economy will become fully profitable.
Nvidia’s Admission: “The Cost of Computing Exceeds That of Employees“
The most striking signal may have come from Nvidia itself. Bryan Catanzaro, vice president of the company, stated: “The cost of compute is far beyond the costs of the employees.” This phrase was picked up by several economic media outlets: Fortune – Nvidia executive says AI costs more than employees, Axios – AI can cost more than human workers now, Entrepreneur – Nvidia VP says it costs more to use AI than to hire humans.
This statement is significant because it comes from the company that directly benefits from the AI boom. Nvidia is not questioning the strategic interest of AI. However, the company implicitly acknowledges that certain uses remain economically very expensive. This reality is often masked by low-priced consumer subscriptions. A user today pays a few dozen euros per month to access extremely sophisticated models. But this price probably does not reflect the complete infrastructure cost.
The Real Problem: Marginal Cost
In many companies, an already recruited employee has a relatively stable marginal cost. Conversely, an AI agent consumes GPU, energy, bandwidth, storage, and cloud resources with each query.
Each interaction adds a cost. This difference profoundly changes the economics of automation. A human employee can handle ambiguous situations, arbitrate, contextualize, and make decisions without triggering thousands of computational operations. AI, on the other hand, transforms each cognitive operation into computing expenditure.
We are therefore witnessing a partial reversal of the historical software paradigm.
An Economy Still Artificially Stabilized?
However, caution is necessary. Asserting that “hyperscalers subsidize AI” would be excessive without access to their detailed accounting data. But several analysts estimate that current prices may not reflect the complete cost of infrastructures, energy, data center depreciation, and GPUs. Current competition pushes players to rapidly conquer market share and lock in usage patterns to impose their ecosystems. This logic recalls several historical phases of digital technology where growth preceded profitability.
The Energy War Behind AI
The issue of AI’s true cost cannot be separated from energy. Modern generative models require massive electrical power, advanced cooling systems, and complex network infrastructures. The explosion in GPU demand is already creating tensions in supply chains, HBM memory, and the electrical capacities of certain territories. Generative AI is gradually becoming a geopolitical and energy issue.
Companies that massively adopt AI agents will tomorrow need to integrate energy cost, resource availability, regulatory constraints, and digital sovereignty issues.
The Illusion of the Universal Autonomous Agent
For several months, the industry has been talking extensively about AI agents capable of replacing entire functions. Yet several limitations are appearing: hallucinations, reasoning errors, necessary human supervision, high inference cost, and maintenance difficulty.
The more autonomous an agent becomes, the more it consumes context, memory, reasoning, and tool calls. Autonomy therefore simultaneously increases power… but also cost. This economic equation remains largely underestimated.
Toward a Hybrid Human + AI Model?
Another recent work provides interesting insight: Economics of Human and AI Collaboration (arXiv)
Researchers show that the human + AI combination often becomes more economically efficient than total automation. This so-called “human-in-the-loop” approach could become dominant. AI excels in speed, synthesis, assistance, and exploration. Humans retain decisive advantages in contextualization, accountability, arbitration, situated creativity, and implicit understanding.
The future could therefore be less about complete replacement than about augmented cognitive collaboration.
Companies Will Soon Need to Make Different Trade-offs
For two years, many AI projects have been launched in an experimental logic. But a new phase is beginning: that of economic rationalization. Finance departments will gradually ask for the real cost per task, effective ROI, energy cost, infrastructure cost, and human supervision cost. Some automations may then appear less profitable than expected, particularly simple tasks, functions that are inexpensive humanly, and processes requiring much verification.
A Possible Gap Between Large Companies and SMEs
Large companies have volumes, cloud infrastructures, investment capacity, and internal AI teams. They will be able to more easily absorb inference costs. Conversely, some SMEs may discover that subscriptions increase, quotas decrease, and advanced uses become expensive.
The risk is then seeing a new technological concentration emerge around players capable of financing AI at scale.
The Return of the Human Factor
The paradox is fascinating. Generative AI was supposed to reduce the weight of human labor. Yet it could ultimately revalue certain human skills: supervision, judgment, relationship management, contextualized creativity, and complex arbitration.
Humans do not disappear from the system. They change position in the value chain. In several cases, the question will no longer be: “Can we replace humans?”, but rather: “What human + AI balance truly minimizes total cost while maximizing quality?”
An Economic Revolution Still Unfinished
The story of generative AI remains largely open. Costs could decrease thanks to smaller models, specialized chips, software optimization, and hybrid architectures. But several major trends remain: explosion of uses, energy needs, growth of contexts, multi-step agents, and quality requirements. The real challenge will probably not be the disappearance of human work, but the economic redefinition of cognition in the enterprise.
We may be entering a period where each cognitive task must be arbitrated between human cost, machine cost, energy cost, regulatory cost, and reputational cost. Generative AI will then no longer be just a technological revolution. It will become a fundamental question of political economy.
