AI campuses worth 45 billion euros on one side, 95% of GenAI pilots with no return on the other. The problem isn’t the number of megawatts – it’s what we do with them.
In Hauts-de-France, SoftBank is committing 45 billion euros for 3.1 gigawatts of data centers dedicated to AI – first firm phase of a 75 billion project, announced at the end of May at the Choose France summit. At the same time, 95% of enterprise generative AI pilots produce no measurable return. Two figures, one question: why so much power for such slim results?
The public debate pits those who want to build in the name of sovereignty against those who want to slow down in the name of the environment. Yet both camps share the same hypothesis – more megawatts, therefore more AI, therefore more value. She is the one I want to examine. Because the harshest constraint is neither ecological nor political. It is economical.
Infrastructure hits two walls, not one
The first wall is physical and documented. Global data center power consumption jumped 17% in 2025, to around 485 TWh, and the IEA projects it towards 950 TWh in 2030 (special report “Key Questions on Energy and AI”, 2026). A doubling in five years.
The amounts follow the same slope. Investment spending by the five largest technology groups exceeded $400 billion in 2025 and is expected to further grow by 75% in 2026. To provide context: the EPR2 program, six nuclear reactors, is capped at 72.8 billion euros in 2020 – figures audited and confirmed by EDF at the Nuclear Policy Council in March 2026. A private data center project at the price of a national nuclear program. You had to dare.
The second wall is financial. We talk about it much less. He’s getting closer faster.
Nineteen pilots out of twenty
According to the MIT study published in the summer of 2025 (“The GenAI Divide”, NANDA project), 95% of generative AI pilots in business produce no measurable return, for 30 to 40 billion dollars invested. I prefer to say it up front: its methodology has been criticized for its very narrow success criterion – a measurable P&L impact six months after the pilot. But even taking some margin, the order of magnitude remains uncomfortable. A year later, same picture: according to Domino Data Lab’s annual survey (July 21, 2026) – a publisher’s study, read as such – of 639 AI executives from companies worth more than $100 million, 57% have a return that still does not cover their AI expenses. Unchanged since 2025.
The problem, obviously, is not the available power.
The bill for architectural laziness
My response as a practitioner: cheap power has killed optimization. When computing seems limitless, no one questions whether the task really required the biggest model. A three-page prompt to file an email: no one would charge a client for that, but you charge a GPU without batting an eyelid.
Each waste taken in isolation is invisible. The same context returned for each request, a giant model on a task that a business rule already solved – multiply by millions of requests: a server load, then megawatts, then one more data center. The race for megawatts is not only the price of progress. It is also, in part, the result of architectural laziness.
Software has been through this cycle before: as long as the hardware absorbed everything, the code bloated; the constraint returned, clean code once again became an economic advantage. AI gets there, faster.
The economics of efficiency
The signal already exists, and the IEA figures it: measured per task, AI energy efficiency is progressing “at a rate unprecedented in the history of energy,” with energy per task falling by at least an order of magnitude per year. These gains are devoured by the explosion of energy-intensive uses – agents, video generation. All the more reason not to waste the rest. On the hardware side, the STM32N6 from STMicroelectronics (announced at the end of 2024, produced in volume since) executes neural network inference on a microcontroller: tasks that yesterday required a data center run on the periphery. The edge will not replace the training of large models. It shows the direction: scale the calculation as needed, not the other way around.
This is what I call the efficiency economy. The winner of the decade will not be the player with the most gigawatts. It will be the one that produces the most value per watt and per token consumed. An architectural competition, no budget.
Three levers for a B2B SME
No need for a 45 billion campus. Three sites accessible immediately.
Size the model to the task. Classify, extract, route requests: a small specialized model is sufficient in most of these cases. Reserve the large model for tasks that justify its cost – today, it’s the default setting just about everywhere.
Put the cost per query in the dashboard. You track your acquisition cost to the penny. Do the same with AI: take last month’s API bill, divide it by the number of requests, put the number next to the CAC. Until it is there, optimization remains an intention.
Audit token waste. Overloaded prompts, duplicate context sent, repeated calls for the same result. Treat these feeds like code: a periodic review that makes the fat visible before it becomes a line item.
Change axis
Megawatts will continue to be built – the demand projected by the IEA leaves little doubt. But the right question isn’t who will have the bigger data center – it’s who will need the smaller one to produce the same result. The 95% of drivers without feedback do not lack power. They lack architecture. And no envelope of 45 billion can buy that.
