The Risk of Single-Model Dependency for European AI Startups

The author argues that European AI startups should prioritize operational resilience by diversifying their model dependencies. Relying on a single API provider like OpenAI or Anthropic creates significant risks, as providers can retire models, change pricing, or alter output restrictions due to compliance decisions.

The availability of open-weights models such as Mistral, Gemma, Google, DeepSeek, and Qwen allows companies to run AI on their own hardware, reducing the gap in quality compared to frontier models. For many specific tasks, smaller, self-hosted models are often more cost-effective and efficient.

The author notes that training foundation models in Europe is generally not financially viable for most startups. However, the EU has established 19 AI Factories and 13 Antennas linked to EuroHPC supercomputers to provide startups and SMEs with necessary compute resources.

The author highlights three main obstacles to AI development in Europe: funding, infrastructure, and talent. He recalls his own experience in 2020 when he had to move his company to the US to secure funding more quickly and easier than in Europe.