The hardware capable of training large models sits overwhelmingly in a small number of countries, and access to it is priced and allocated accordingly. That is a fact about the world rather than a grievance, and the useful conversation is about what follows from it.
Training and inference are different problems
Training a frontier model from scratch requires capital and energy at a scale that only a handful of organisations anywhere can command. Running one — inference — is enormously cheaper, and fine-tuning an existing open-weight model for a specific language or task sits somewhere in between and is well within reach of a university department or a well-funded company.
Conflating the two produces bad strategy. The question is rarely "can we train a frontier model" and almost always "what can we build with the weights that already exist, and where does our own data give us an advantage nobody else has".
The energy constraint is real
Compute at scale is an electricity business. In grids that already struggle to serve existing demand, siting large training clusters raises questions about who gets the power that go well beyond technology policy. Countries with surplus generation capacity are in a genuinely different position, and a few are treating that as a strategic asset.
Where the leverage actually is
Not in matching other people's training runs. It is in data nobody else has, evaluation that reflects local reality, and deployment into markets that outside companies do not understand well enough to serve. Those are defensible positions. Competing on cluster size is not.