What is AI, concretely?

I was starting to write an article about my methodology for rolling out AI in a company's projects, when a conversation I had recently at an event in Paris started echoing in my head:

« You know, Benjamin, understanding everything that makes artificial intelligence possible, or even what artificial intelligence itself is, is already a problem. So talking about strategy… »

And there was my mistake: I was writing you a methodological article about a concept that perhaps was not clear yet to the reader, or worse, a concept you and I did not define in quite the same way! So why not start over from the beginning?

Let us talk about what AI is. This article is meant to be accessible to everyone:

• ▸it is neither a theory book for researchers (who already know the subject),
• ▸nor a programming tutorial.

It is a plain-language article, but one that can also be useful for developers ;), that will give you a solid base for understanding the stakes around this topic, the players and the trends, so you can make your own decisions.

With that said, let us start talking about AI!

AI is made of layers

Your Gemini on your phone, or your Citymapper finding you the best route, are in fact made of several components which, put together, make the applications you know. Each piece will get its own chapter to go into a bit more detail. Here, I am just going to explain broadly what they represent. For those who cannot bear long reads, or my writing style, you will already have a good overview! 😉

Here are the different components you have at hand to build and run your AI applications. Each major part comes with its own commercial, autonomy or industrial stakes.

Let us quickly describe what we have here with a simple example. Hold on tight, we are going to move very fast over concepts we will dig into in the coming chapters!

Punch it, Chewie — light speed!

Our running thread

You are at your company, which has a strict data protection policy. Your manager asks you for a summary of the commercial strategy you have just received from another department. The document is long and you do not really have the time. Your company has not given you an AI tool (or only one that is not very good). So, like most employees right now, you are going to use a tool you know: ChatGPT.

Red card! You have just done Shadow AI: using an external AI and handing it company data (part of the company assets) to OpenAI, which may use it to train its own models.

Right, back to our little story. So you open your browser and go to the ChatGPT website to talk to your AI.

1. The application layer

Here you are on the application layer: the site holding the chat interface, the tools that check you are who you claim to be, grant you access (and charge you for your subscription). Beyond AI itself, this application is what changed everything when LLMs arrived. The simpler it is to use, the better for users. That is why you must think carefully about what your user will actually see: the "100% chat" trap is already pretty obvious in the various companies I come across.

You upload your file and say: "Give me a summary in 5 bullet points."

2. The software layer

At this point you think: "Right, the AI is working now!" But actually… not yet! First, your file is handled by the software layer (regular code) before it ever meets an AI model. The text is extracted, then cut into small pieces (what we call chunking).

Why? Because every AI model has a size limit on what it can read at once: its context window (a bit like the working memory of your brain when you read a page). If your document exceeds that capacity, or if the application wants to save time and tokens, the software will orchestrate the work first: cut up the big block, summarise each piece, or fetch only the key passages (using a technique called RAG) so that only the essentials reach the model. It is genuine teamwork between traditional code and AI!

3. The AI model

Now the text reaches the model (LLM) you picked in the interface.

This is where we are genuinely talking about AI. And there are plenty of kinds! I am talking about LLMs here because it is the best-known type of AI, but there are other algorithms that perform extremely well in their own domains. And within a single family you have different models: this is where GPT, Claude, Mistral, Qwen, Kimi and the rest come in.

4. The infrastructure: where does all this run?

While our models are working, let us ask the question: where do they actually run? Your document did not stay in your browser! It was sent somewhere to be analysed: to a cloud provider, in a data centre.

Now, there is something you need to know: even if your physical server sits in France, your data remains subject to the law of your provider country of origin. If you use Azure (Microsoft), AWS (Amazon) or Google Cloud, your data falls under US extraterritorial law, notably the CLOUD Act (Clarifying Lawful Overseas Use of Data Act) or section 702 of FISA. In plain terms: US justice can in theory demand access to data handled by a US company, regardless of whether the servers are in Paris, Amsterdam or Frankfurt. And without even telling you! Geography is not everything, the nationality of the tech giant matters too.

There are many of them. The best known are AWS (Amazon), Google Cloud and Azure (Microsoft), plus French players such as Scaleway, OVHcloud and Outscale. But those are just brands. Data centres, on the other hand, are anchored in a physical territory. As we speak, your document could be in Paris, Amsterdam or New York. In fact part of your application can be in one place and your AI model in another (say, your site on Azure in Europe, but the AI writing the summary in a US data centre). And of course, all that infrastructure needs electricity to run and water to cool!

5. The hardware

Your document was really long, it is still running, and the chips are heating up a little… Yes, the hardware! The chips (GPUs, CPUs and so on) that everything rests on: your document, the AI models, your application… The cloud providers had to buy them.

Today this market is dominated by one major player: NVIDIA. But NVIDIA does not manufacture its chips directly: it designs them. The ones producing them are foundries such as TSMC (mainly located in Taiwan). And the wildest part is that the ultra-complex machines these foundries use to etch your chips are built by a single company in the Netherlands: ASML! Not to mention the raw materials needed to make all this hardware.

And there we are, the task is done! The chunks of text were processed one last time by the model to produce your 5 bullet points, and that is that!

A more realistic view

As you will have gathered, if we wanted to represent this more realistically than our initial breakdown:

Different machines run your AI and your application, all orchestrated by software and hosted at a cloud provider. And at every level, we have questions of performance, trust, sovereignty and so on.

If you want to understand the stakes around all these subjects, do read the coming chapters! And above all, send me your feedback so I can sharpen my writing.

Have an excellent day, and do not forget to have fun!