The value of AI is spread across a chain of several layers, from chips to applications. Understanding this chain means understanding where value is created and where European sovereignty stands.
4.1Semiconductors (the hardware base)
This is the most concentrated and most strategic layer. NVIDIA controls more than 90% of the market for generative AI chips (GPUs) and is a de facto gatekeeper: Mistral trains its models on thousands of Nvidia H100 and L40S GPUs. Competitors: AMD, and the in-house chips of the cloud giants (Google's TPUs, AWS's Trainium).
Upstream, manufacturing depends on TSMC (Taiwan) and cutting-edge lithography on the Dutch company ASML, a major European asset. Emerging players (Cerebras, Groq) are pursuing alternative architectures. French/European position: SiPearl (first units expected around 2027-2028) and Kalray, still far from Nvidia's scale.
4.2Infrastructure and energy
Data centres consume growing amounts of electricity; energy is becoming a central constraint on AI development. France highlights one asset: abundant, low-carbon electricity (nuclear).
4.3Cloud and computing power
American hyperscalers dominate: AWS, Microsoft Azure, Google Cloud, Oracle. A new generation of AI-specialised "neoclouds" is emerging (CoreWeave, Nebius).
French/European position: OVHcloud and Scaleway (Iliad group), as well as Mistral Compute. The issue is sovereignty: a European cloud is not subject to the US CLOUD Act, a decisive argument for banks, hospitals and public administrations bound by the GDPR.
4.4Data and training
Models need annotated data and massive human work for tuning and alignment (RLHF). The emblematic player is Scale AI, whose founder Alexandr Wang joined Meta in 2025 to lead its superintelligence efforts. Other players: Surge AI, and large-scale outsourced annotation.
4.5Foundation models
The media heart of AI. Three major geographical poles stand out:
- United States: OpenAI, Anthropic, Google DeepMind, xAI, Meta (Llama).
- France / Europe: Mistral (open-weight, ~$14bn in 2026); AMI Labs (LeCun, world models); the Cohere–Aleph Alpha tie-up (Canada/Germany), valued at ~$20bn in April 2026.
- China: DeepSeek, Alibaba (Qwen), Moonshot (Kimi), ByteDance.
4.6Applications and deployment
The layer in contact with end users: vertical SaaS, integrators, consulting firms. It is here, in the application layer, that the largest number of entrepreneurial opportunities arise (see Module 5).
Sovereignty: where does Europe stand?
France launched the France 2030 plan (up to €109bn announced for AI), including ~€10bn through Bpifrance. At European level, the EuroStack initiative targets €300bn of computing, AI and cloud capacity "under European control" by 2035. The fact remains, however, that in the most strategic layer (chips), dependence on Nvidia and Asia is still almost total.
Sources
Mistral AI Secures $14B Valuation (pulse24.ai); Sovereign AI (softwareseni.com); France's AI Sovereignty Push (introl.com).
Finished this module?
Your progress is saved on this device.