The frontier AI market is consolidating into closed camps, with labs choosing partners, cutting rivals, and restricting access to their strongest models. Examples: Salesforce made Anthropic‘s Claude the default model in its CRM and Slack. OpenAI terminated Cursor’s API access on November 12 after SpaceX acquired the company, citing prior contract breaches. Even ostensibly open-weight models impose gates: Z.ai released GLM-5.3-Flash under MIT license on August 26 but put its flagship behind a $10 billion host-revenue review two days later.

Anthropic started this pattern: its strongest model, Mythos 5, ships only through Project Glasswing, a program that rations capabilities to trusted companies. Its sibling Fable went dark outside the US for three weeks, then returned to US-only 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 frontier models, using export restrictions and nationality screening.

Enterprise buyers face governance challenges: strict Zero Data Retention policies, data sovereignty mandates, and wariness of sending IP through prompts. When platforms hardcode a single provider, buyers need leverage to negotiate changes. The era of permissive open-source weights is narrowing, with creators introducing free-to-pay conversion thresholds for large cloud hosts. Open source now means open until you scale.

Governments treat frontier AI as critical sovereign infrastructure. Pushing back is a $5 trillion countervailing force: Nvidia, leveraging its market cap to invest in open ecosystems. Nvidia has spent $13 billion on Hugging Face, $7 billion on Poolside’s model factory, and a $26 billion commitment to Nemotron, among others. Open weights are Nvidia‘s structural defense against proprietary labs and sovereign restrictions. The key takeaway: access, not price, is the new scarcity at the frontier.