Infrastructure: from giga-datacenters to the silicon in your pocket

When we talk about AI, we often imagine an abstract concept, a sort of « magic cloud » floating around the web and answering our questions. In reality, AI has a brutal materiality. It is steel, concrete, copper, ultra-pure silicon, megawatts of electricity and millions of litres of water.

In this chapter, we are going down into the basement of AI: the place where physics, geopolitics and law collide.

The cloud provider paradox: geography does not make the law

When you store your data or run your models at one of the cloud giants, the first question your security officer asks is usually: « Where are the servers? » If the answer is « In Paris or Frankfurt! », you probably think: « Good, it is in Europe, my data is protected by the GDPR. »

And that is where the trap is.

This is what we call the digital sovereignty paradox: the physical location of a server is not enough to protect your data, because the nationality of the company that owns the datacenter overrides geography.

• ▸If you use AWS (Amazon), Azure (Microsoft) or Google Cloud, even with your physical server sitting in Marseille or Paris, those companies are American.
• ▸They are therefore subject to the extraterritorial laws of the United States, notably the CLOUD Act and section 702 of FISA.
• ▸In plain terms: an order from an American judge or intelligence agency can in theory compel Microsoft or Amazon to hand over access to the data hosted on their servers, no matter that the server sits in France.

This is why the battle of the « sovereign clouds » is raging. On one side, French and European players (Scaleway, OVHcloud, Outscale, Hetzner) guarantee full legal protection against American law, but are struggling to line up the computing power required. On the other, the American giants are pouring in billions to reassure the market with « sovereign » or « trusted » offerings.

The monstrous cost of an AI datacenter

Building a datacenter for generative AI has nothing to do with building a classic datacenter meant to host websites. A traditional server draws a few kilowatts per rack. A state-of-the-art GPU farm built for AI draws megawatts across a single building.

⚡ Energy: the gigawatt war

Electricity needs have become so colossal that the International Energy Agency estimates datacenter consumption could double by 2030. To feed these calculation factories, cloud giants now sign contracts directly with nuclear plants, or buy entire solar farms. Access to abundant, stable and low-carbon energy has become the number one factor deciding where a datacenter gets built.

And this is why, and forgive the national pride, France is a rather serious country on these matters: nuclear power gives us stable, low-carbon energy, which makes it very interesting ground for building datacenters.

And France is really not bad: more than 95 % of our electricity comes from low-carbon sources, about two thirds of it nuclear. Which makes us a net exporter of electricity, helping our neighbours cut their own emissions. Let us compare, over the same period:

There are two things to read in this chart. First, Sweden does even better than we do, but on a much smaller total output and above all one that leans heavily on hydro. Second, and this is the part that matters for a datacenter: Germany and the United Kingdom have invested massively in wind and solar, yet remain at 19.9 % coal for the former and 26.4 % gas for the latter. Electricity that is low-carbon and dispatchable, available at night with no wind, remains a rare commodity.

This is one of the reasons why Mistral signed such a large contract with Microsoft for computing power. Microsoft was short on compute, and that created an opportunity for a company like Mistral to become more than a model maker: a service provider. Some will see a very shrewd strategic move, others a « betrayal » of sovereignty, adding American involvement on our own soil. What happens when your monopolistic customer asks you for something?

The reality is more complex than that. Who else was ready to invest that much in France or in Europe? How do you turn down a chance to keep a hand, even partially, on this market? I am thinking of writing an article about these strategic questions and my personal view of what ought to be done. Do remind me if the topic interests you :)

💧 Water: the cooling headache

Another problem: these chips get hot. Very hot. Classic air cooling, the big fans, no longer suffices for high-density chips. Datacenters are shifting massively towards liquid cooling, where fluids circulate in direct contact with the processors.

But that requires complex infrastructure, and sometimes impressive volumes of water for the cooling towers. Managing the environmental impact of that water has become a major political issue for the local authorities hosting these centres.

All of this pushes cloud providers to improve their energy efficiency and their cooling systems, such as the adiabatic cooling used at Scaleway's Paris 2 datacenter. Some even go as far as giving their customers tools to compute environmental impact. Because yes, AI is a heavy consumer of energy, and it is currently under close scrutiny: companies need as much monitoring and steering capability as they can get, for their sustainability strategy, but also quite simply for their spending.

It is not all about tokens ;)

The silicon tax: where does the money really go?

When a cloud provider spends billions building an AI datacenter, what do you think costs the most? The walls? The electricity? The network?

No. It is the chips.

In a datacenter optimised for AI at very large scale, the computing hardware (the graphics cards) costs roughly twice as much as the building itself. For a 1 gigawatt campus project, count around 10 billion dollars for the structure and infrastructure, and 20 billion dollars just to fill the racks with chips.

Why such an imbalance? Quite simply because the market is dominated by a near-monopoly: NVIDIA. The company does not merely sell a chip at tens of thousands of dollars apiece, it sells the complete architecture: the ultra-fast interconnect (NVLink) and the software layer (CUDA) on which almost every AI model on the planet rests.

This situation has even given birth to a new category of players: the neoclouds (CoreWeave, Nebius, Lambda). These are not general-purpose cloud providers but specialists doing one single thing: buying tens of thousands of NVIDIA GPUs, and renting them out by the hour to AI companies.

For the detail on the chips, I will tell you about it in the next chapter. A little mystery is always good for keeping you attentive ;)

The big trend: hybrid AI

Faced with insane infrastructure costs, network latency problems and confidentiality concerns, one question arises: are we condemned to run everything in enormous remote datacenters? The answer is no. We are watching a shift towards on-device AI.

The reign of Small Language Models

The « ever bigger » era is reaching its economic limits. We have realised there was no need for a giant 700-billion-parameter model hosted in the United States to fix a typo or summarise a work email.

Thanks to ultra-optimised, compact models (SLMs, Small Language Models, such as the lightweight versions of Mistral, Llama or Phi) and to the arrival of dedicated chips in our devices (NPUs, Neural Processing Units), AI is moving straight into our smartphones and computers: Copilot+ PCs, Apple Intelligence.

The problem is that using SLMs demands far more engineering care than throwing everything at a big model. Which is why the practice is not yet widespread.

The hybrid model: the right compromise

Some think the future of infrastructure will be neither 100 % cloud nor 100 % local, but hybrid:

• ▸Locally, on your phone or PC: for 80 % of everyday tasks (autocomplete, photo retouching, searching your own documents). It is instant, free of API costs, it works offline, and your data never leaves your device.
• ▸In the cloud: for the 20 % of complex tasks that demand deep reasoning, analysis of terabytes of data or heavy video generation.

Today the market does seem to be heading towards this approach, with a lower share for the local side. And we are increasingly seeing a third strand gain importance: the private cloud / on-premise, hosting on servers inside the company itself, for workloads that are too sensitive. It is generally on this kind of subject that technologies like Socle AI become more than an accelerator: a necessary step, because the whole AI infrastructure then has to be built in-house and wired together to deliver value.

A cloud provider: expertise that is not only in the bricks

So far we have talked about walls, chips and energy. But making servers work together, maintaining them with all the software layers specific to that hardware, is no picnic. It is a precise expertise, not to be underestimated: good devops, those grumbling engineers in dark rooms, are in fact the ones without whom nothing would work at all.

« A Rafale pilot is impressive. Rather less so if the aircraft never takes off, because nobody installed or maintained its engine. »

That is also the job of a cloud provider: supplying the tools, the services, and therefore the ability to interact with those hardware resources. And each has its own positioning, from Outscale, Dassault's cloud provider with an institutional reputation, to Scaleway, the fashionable provider that moved fast on AI.

In short

Infrastructure is the physical frontier of AI. It reminds us that a single click on « Generate » sets off a global supply chain: chips etched in Taiwan, sold by an American company, installed in a water-cooled datacenter in Europe, and powered by a local power plant.

Understanding this layer means understanding the real costs, the real ecological limits and the genuine sovereignty stakes of artificial intelligence. And since everything always comes back to the chips, let us go and look more closely at who knows how to make them.