Generative artificial intelligence: the Autorité issues its opinion on the competitive functioning of the sector

The generative AI sector

Definition

According to the European Parliament, artificial intelligence refers to any tool used by a machine "to display human-like capabilities such as reasoning, learning, planning and creativity". Generative AI refers to AI models capable of generating new content such as text, image, sound or video.

A growing priority for public authorities

The generative AI sector is attracting growing interest around the world.

In France, the government launched a national AI strategy in 2018, for which almost €2.5 billion of the "France 2030" plan has been earmarked. In March 2024, the French AI Commission (Commission de l'IA) launched by the Prime Minister presented a report containing 25 recommendations, suggesting in particular to make France a major centre for computing power, to facilitate data access and to establish global AI governance.

At European level, most of the provisions of the AI Act (which will soon be published in the EU Official Journal) will be applicable from 2026. Although published before the rise of generative AI, the Digital Markets Act (DMA) and the Data Act will have an impact on the sector.

A series of initiatives on AI have been adopted globally, such as the Bletchley Declaration in the United Kingdom in November 2023 at the AI Safety Summit. The next global summit will take place in France on 10 and 11 February 2025. Other initiatives have been taken by the G7, the United States, the United Kingdom and China, for example.

How the sector works

There are two key phases in generative AI modelling:

• training: the initial learning process of a model (often called "foundation model", which includes LLMs), during which its parameters, known as "weights", are determined. Training requires both significant computing power and a large volume of - generally public - data. The training phase may be followed by fine-tuning, during which the model is adapted to a specific task or a specialised dataset (e.g. legal or health-related data). Fine-tuning is generally based on a smaller, proprietary dataset and may involve human expertise;
• inference: the use of the trained model to generate content. The model can be made accessible to users via specific applications, such as Open AI's ChatGPT or Mistral AI's Le Chat, or APIs for developers. The computing power required depends on the number of users. Unlike many digital services, the marginal cost of generative AI is not negligible, given the cost of the computing power required. New data that was not used for training may be added during the inference phase, in order to ground the model in recent data, such as news articles.

The participants in the value chain

The operators in the generative AI value chain are:

• major digital companies: Alphabet and Microsoft are present across the entire value chain, while Amazon, Apple, Meta and Nvidia are present only at certain specific layers;
• model developers: for example, start-ups or AI-focused research labs, such as Anthropic, Hugging Face, Mistral AI and OpenAI. They have often formed partnerships with one or more digital giants, such as OpenAI with Microsoft and Anthropic with Amazon and Google. For the distribution of their models, they may adopt either a proprietary or open-source approach.

At the upstream level, several types of operators are involved:

• IT component suppliers, such as Nvidia, develop graphics processing units (GPUs) and AI accelerators, which are essential components for training generative AI models;
• cloud service providers, including digital giants, known as "hyperscalers", such as Amazon Web Services (AWS), Google Cloud Platform (GCP) and Microsoft Azure, cloud providers such as OVHCloud, as well as specialist AI providers such as CoreWeave. The necessary computing resources may also be provided by public supercomputers (such as Jean Zay in France).

At the downstream level, many operators are marketing new services based on generative AI to the general public (like ChatGPT), companies and public authorities and/or integrating generative AI into their existing services (like Zoom).

Public link

Please use the above public link if you want to share this noodl on another website.