Commentators across Europe are voicing concern over the announcement of the partnership between Microsoft and Mistral AI, a French AI startup. This collaboration grants Mistral access to Microsoft’s infrastructure, including supercomputers. This will let Mistral both expedite the development of models and provide the capacity to run inference.
Some critics claim that European legislators have been played. Mistral was particularly vocal during the recent deliberations on the AI Act, pushing for loosening the regulations on foundation models and general-purpose AI. The start-up argued that overly stringent rules would damage “European champions” and make it difficult to compete with big tech companies from other regions. Now that the goal has been achieved, Mistral can stop pretending and instead enter into a partnership with the big tech company, its former supposed rival. This is a win for Microsoft who further diversifies the portfolio of models at its disposal.
It does seem rather coincidental that news of this partnership emerged shortly after the AI Act negotiations were finalized. However, what’s perhaps more significant is that this partnership underscores the fact that infrastructure poses the primary bottleneck in AI development.
Speaking at a panel discussion in the European Parliament on 19 February, Andreas Mundt, President of Germany’s Federal Cartel Office, said that the EU’s hope against Big Tech lies with small and medium-sized companies that have very high-quality specialized data, which “can be used to build specialized AI models [that] could outperform foundation models in their respective domains”
The case of Mistral shows that competing with big tech requires more than the creation of “specialized models.” If European companies have to depend on partnerships with big tech to develop their models, no real competition is possible. Without the Public Digital Infrastructure and without breaking monopolies in cloud infrastructure, any discussion of European companies challenging big tech’s position and European digital sovereignty sounds nice but is almost meaningless. Digital sovereignty limited to the application layer is not real sovereignty.
