Using Artificial Intelligence has become remarkably easy. Just open an app, type a request, and wait a few seconds. From the outside, it seems almost trivial: a prompt enters, a response comes out. But this apparent simplicity masks a much heavier, costly, and complex industrial reality. Behind every query are data centers, networks, electricity, cooling systems, specialized chips, and billions of dollars in investment that no one is deploying purely out of technological romanticism.
The popularization of generative AI has created a dangerous illusion: that access is cheap because the user’s effort is also minimal. But just because a prompt is convenient doesn’t mean its structural cost is irrelevant. The International Energy Agency estimates that the electrical demand of data centers went from about 485 TWh in 2025 to a trajectory pointing to 950 TWh by 2030, and that the energy consumption of AI-focused data centers grew by 50% in 2025 alone. This is not just a technological anecdote, but a new industrial layer with direct impacts on power grids, investment, and profit margins.
The problem isn’t the prompt, but the scale
It’s important to clarify a point that’s often exaggerated on social media. Not every individual query consumes an enormous amount of energy by itself. Epoch AI estimated in 2025 that a typical ChatGPT query with GPT-4 was around 0.3 Wh, well below some of the earlier estimates that circulated for months. More recent academic reviews continue to place the energy cost per query within a range of a few watt-hours depending on the model, length, and infrastructure used. The real issue is not the single gesture but the massive and sustained scale of usage.
This completely changes the perspective. When a tool shifts from handling thousands of queries to hundreds of millions or billions, the cost per request ceases to be an academic curiosity and becomes a critical business variable. The IEA emphasizes that AI is now the main driver of data center electricity consumption growth and that the U.S. and China will account for nearly 80% of the total increase through 2030, while Europe will also continue to grow strongly.
That’s why the race is no longer just about models but about the infrastructure capable of supporting them. Andy Jassy, CEO of Amazon, explained in his 2025 shareholder letter that the company plans to invest around $200 billion in capex in 2026, much of it tied to AWS and AI demand. Microsoft declared in January 2026 a quarterly capital expenditure of $37.5 billion, with approximately two-thirds dedicated to short-lived assets like GPUs and CPUs. Alphabet, for its part, projected a capex between $175 billion and $185 billion for 2026, driven by AI infrastructure needs.
A giant industry disguised as simplicity
The popularization of AI has been largely a visual mirage. Users see a clean interface and a text box. Operators see something else: energy availability, high-speed networks, chip supply chains, liquid cooling, long-term contracts, and capacity planning. The difference between these perspectives explains why the market will eventually adjust. No company invests hundreds of billions to provide capacity indefinitely as a free good.
From the cloud infrastructure sector, this tension is quite clear. In a recent analysis on AI and architecture, David Carrero, co-founder of Stackscale (Aire Group), summarized it as: “AI tests the architecture at its most fragile points: cost per use, latency, and control. When inference becomes daily and critical, predictability and governance capabilities are required—not just the speed to deploy a pilot.”
The phrase holds more depth than it seems. During the initial enthusiasm phase, many companies assumed that connecting models, experimenting, and growing was enough. But daily inference, RAG systems, traceability, feedback loops, and integration into critical environments turn AI into a structural load, not just an experimental toy. When that happens, costs cease to be an abstract billing line and become a key architectural decision.
Energy, GPU, and data centers: the real bottleneck
The other major constraint is physical. Carrero warned in another analysis about data centers that AI racks can now exceed 70-80 kW per rack, far above traditional environments, forcing a rethink in electrical distribution, supply, and cooling. He also noted that, where 10-15 kW densities were common before, the market is now moving toward more than 40 kW per rack in certain deployments. This cannot be resolved with a simple software layer.
The IEA reinforces this view with global figures: in their baseline scenario, the electricity needed to power data centers will grow from 460 TWh in 2024 to over 1,000 TWh by 2030, with renewables covering almost half of the additional demand, but also with natural gas, coal, and nuclear playing significant roles. Frankly, AI is driving an industrial expansion that cannot be separated from the energy debate.
Moreover, location matters. The U.S. continues to lead in volume, and China in growth speed, while Europe seeks to boost capacity without sacrificing sovereignty, costs, and regulations. In this context, AI is less like an app and more like a capital-intensive industry — similar to other technological revolutions where initial competition is about building capacity, followed by monetization.
Real monetization has yet to fully arrive
This is perhaps the most fundamental point. Much of the market still invests for positioning, future share, and scale effects. But that cycle isn’t endless. The history of technology tends to repeat itself: first comes aggressive expansion, then the pressure to demonstrate returns. When it’s time to monetize genuinely, prices change, priorities shift, and tolerance for unprofitable uses diminishes.
This doesn’t mean AI will stop or that the current model is immediately inviable. It simply means that the market will tend to discriminate between high-value uses and trivial consumption, between flashy pilots and sustainable deployments. What seems like cheap magic today will probably come to resemble a costly digital utility—governed by energy costs, capacity commitments, and much stricter business models.
In this process, AI will keep growing. But it will gradually stop seeming free.
Frequently Asked Questions
Does each AI prompt consume a lot of energy by itself?
Not necessarily. Recent estimates, such as Epoch AI’s for GPT-4, place a typical query at around 0.3 Wh. The main issue isn’t the individual prompt but the massive volume of usage and the infrastructure needed to support it.
Why is it said that AI isn’t magic but infrastructure?
Because behind every AI service are data centers, GPUs, networks, electricity, cooling, and billions in capital investment. Major providers are spending record sums to expand capacity.
How much is AI-related electricity consumption growing?
The IEA estimates total data center energy use increased by 17% in 2025, with AI-focused data centers growing by 50% that same year. Additionally, they project total data center demand will nearly double by 2030.
What role does private or hybrid infrastructure play in this scenario?
According to David Carrero, when inference becomes daily and critical, factors like cost per use, latency, and control become increasingly important, driving interest in private and hybrid architectures for certain AI workloads.