Salesforce made Claude the default model across Agentforce and Slack this August, under a partnership branded Claudeforce. Three months earlier, OpenAI had terminated Cursor’s API access after SpaceX acquired the coding-tool company, citing prior contract violations. Those two events, on their own, look like ordinary vendor moves. Together with a run of similar decisions across the industry, they describe something structural: frontier AI providers are choosing sides, and enterprises are being asked to pick one too.
Tom Tunguz, the venture investor who writes the Tomasz Tunguz blog, laid out this thesis in a recent post. His argument is that access, not price, has become the scarce resource at the frontier. Anthropic now routes its strongest model, Mythos 5, only through Project Glasswing, a vetting program limited to trusted partners. A sibling model was unavailable to anyone outside the United States for three weeks this summer, then came back American-only and more expensive. OpenAI ran the same play, handing its government editions of GPT-5.6 to a short list of partners well ahead of any public endpoint.
Open weights were supposed to be the release valve. That is narrowing too. Z.ai shipped its GLM-5.3-Flash model under a permissive MIT license in late August, then put its flagship model behind a licensing requirement two days later: any host doing more than $10 billion in trailing revenue has to clear a security review before commercial use. The pattern recurring across labs is a free tier for experimentation and a gated tier once a company matters enough to be worth gating.
Nvidia is the one large actor pushing the other direction, and its motive is not altruism. The chipmaker has committed capital to open-model efforts, including a five-year, $26 billion pledge tied to Nemotron and separate investments in Hugging Face and Poolside. Nvidia sells compute regardless of which lab’s weights run on it, so an ecosystem of open models that anyone can host protects its market from being cut off by any single closed lab’s access decisions. That is a hedge against customer concentration, not a bet that openness wins on principle.
The practical effect for enterprise buyers is a governance problem that used to be theoretical. Zero-retention data policies, sovereignty requirements, and reluctance to expose proprietary prompts already push companies toward fewer vendors. Hardcoded model defaults inside SaaS platforms compound that: switching providers now means renegotiating a contract, not changing a config flag. A buyer standardizing on one lab’s model today should treat that choice as a multi-year commitment with real switching costs, not a reversible technical decision, and should ask any SaaS vendor exactly which models are swappable versus baked in before signing.
For startups, the cost of landing on the wrong side of a whitelist is existential rather than inconvenient. Cursor lost API access to a frontier model with roughly a day’s notice once its ownership changed hands, a reminder that a startup’s dependency on a single lab’s goodwill is a liability its cap table can trigger. Founders building on a rationed model like Mythos 5 or a partner-only variant of GPT-5.6 should treat that access as revocable at the lab’s discretion, and should be building toward a second qualified model provider now, before a change of control, a compliance dispute, or an export restriction forces the migration on someone else’s timeline.
Tom Tunguz, “The Price of Entry to the Frontier,” published August 13, 2026.
