ASML’s big bet on Mistral: what it means for AI in Europe... and beyond

Yesterday, ASML announced a landmark investment of €1.3 billion in Mistral, the French AI start-up that has become Europe’s most visible challenger to U.S. giants like OpenAI and Anthropic. The partnership is significant on several levels. It ties together two of Europe’s leading technology firms: ASML, the Dutch maker of the world’s most advanced chip equipment, and Mistral, a fast-growing AI company known for its open-source models and focus on business applications.

While it is tempting to celebrate the boost to “European sovereignty”, the benefits go far beyond this.

Introducing Mistral

Mistral is a French artificial intelligence company, founded in 2023, that develops large language models (LLMs), similar to OpenAI’s GPT models, but with two distinctive choices.

First, it is open source by default. Unlike OpenAI or Anthropic, which keep their models closed, Mistral releases its models openly. This means businesses, developers, and researchers can download, adapt, and fine-tune them for their own purposes.

Second, it focuses on smaller, specialized models. In addition to training large, general-purpose models, Mistral emphasizes smaller, efficient models that can be fine-tuned on a company’s own data for specific workflows. These models are cheaper to run, require less computing power, and can often perform better when tailored to a narrow domain.

Mistral is often seen as Europe’s flagship AI start-up, both because of its rapid progress and because it represents a European alternative to US (OpenAI, Anthropic, Google) and Chinese (Alibaba, DeepSeek) players.

The reality of Enterprise AI, today

Broadly speaking, LLMs are applied in two ways within companies:

1. Personal productivity. Through ChatGPT, or, as we increasingly see, variants developed by companies for internal use, which employees can use to optimize their tasks. In this case, the breadth of the model is important: it needs to handle all kinds of questions, so general purpose models, such as ChatGPT, works best.
2. Tailored AI solutions to optimize specific workflows. In these cases, domain specialization is what usually matters most: the input/output expected from the model is limited to a narrowly defined objective – for example the processing of claims in the insurance industry. Here the model doesn’t need to know anything about unrelated fields (quantum mechanics, say). In such cases, it makes sense to specialize a model, and a smaller model may therefore suffice: they are cheaper to use, faster, require less computing power, and can perform even better when fine-tuned.