Frontier AI access tightens as labs pick partners

The pattern of closing access began with Anthropic, which now ships its strongest model Mythos 5 only through Project Glasswing, rationing capabilities to trusted companies. Its sibling Fable was unavailable outside the US for three weeks and now runs only on US-based inference. OpenAI mirrored this by launching government-facing GPT-5.6 variants to a small group of trusted partners before any public API. Governments are also deciding who may run the frontier, treating it as critical sovereign infrastructure with export restrictions and nationality screening.

This closure is countered by Nvidia, whose $5t market cap funds open ecosystems. Its investments—$13b for Hugging Face, $7b for Poolside's model factory, and a $26b commitment to Nemotron—serve as a structural defense against proprietary labs and sovereign restrictions. The article suggests that open weights ensure foundation models cannot be entirely gated, but even these now have free-to-pay conversion thresholds, meaning open source means open until you scale.

For enterprise buyers, the take-it-or-leave-it approach from platforms like Salesforce creates a governance challenge. They must negotiate for leverage to change providers when models are hardcoded, amid strict zero data retention policies and data sovereignty mandates. Access, not price, is becoming the new scarcity at the frontier, with whitelists and blacklists shaping who gets to use the strongest models.