This spring an AI lab built a cybersecurity model so capable that it only gave access to a short list of companies, almost all American. A policy writer argues this is not a one-off. Security worries, the high compute cost of every extra user and growing government involvement all point the same way: the best models will reach a trusted few first, and everyone else later, through someone else’s product.
Many European AI plans quietly assume the opposite. Good enough models, cheap tokens, open API access for all. I think that assumption is the weakest part of our strategy. If the frontier can be switched off for you, then what you run yourself is your floor. Own that floor: open models, your own data, your own infrastructure. Treat frontier access as a bonus, never as the foundation.