When Mark Weatherford, Nvidia’s head of cybersecurity policy, took the stage at the Open Source Summit in Prague, he was not defending just a coalition. He was defending a precise idea: that the security of open source AI needs to be negotiated before regulators turn it into a minefield. The Open Secure AI Alliance, announced in July a few days after the Hugging Face incident and now run by the Linux Foundation, brings together names such as Mistral, Hugging Face, UiPath, Microsoft, Adobe and Dell. Missing, however, are Google, Anthropic, Meta and OpenAI. And those absences matter, especially when it comes to companies that have released open-weight models.

The question put to Weatherford was direct: is the alliance just a symbolic gesture? Will members really share information with rivals in a global AI land grab? His answer was honest enough: success depends on the people who participate, and there will always be market drivers that push a company to ask whether collaboration is worth it. The point, though, is that the alliance lists more than a hundred companies and the admission queue is long. That queue, more than the current member count, is the structural fact of this story.

The Linux Foundation is vetting hundreds and hundreds of companies. Weatherford said it clearly: there is a vetting process, and there are companies, countries or organizations you may not want inside. This changes the nature of the alliance. It is not a simple manifesto: it is a body with an entry filter. For those developing or distributing open-weight LLMs in enterprise or self-hosted contexts, that filter can become a gateway to security tools, threat intelligence and shared best practices. Staying outside does not just mean losing prestige; it means having less voice in defining standards that will later be applied to inference pipelines, vulnerability management and compliance audits.

Asked why OpenAI is not a member, Weatherford replied that there are several large companies that are not signatories and that the Linux Foundation’s vetting queue is still being worked through. TechCrunch had already reported that OpenAI and Google signed the original open letter that led to the group, signed by more than 200 companies. Yet today OpenAI, Google and Meta have released their own open-weight models. The point is not that they do not believe in open source; it is that releasing weights does not automatically mean entering the room where security standards are decided. The separation between distributing a model and participating in open source AI governance is a fracture worth attention.

This fracture has concrete implications for anyone evaluating on-premise deployments. An organization running local LLMs for data sovereignty reasons cannot simply download an open-weight model: it must also manage incidents, verify the software supply chain and demonstrate to regulators that it has adequate controls. The tools that a coalition like the Open Secure AI Alliance chooses to promote — or to neglect — affect the cost and complexity of those obligations. If the center of gravity remains in the hands of vendors like Nvidia, Microsoft and Dell, security frameworks will tend to reflect the architectures and ecosystems those vendors know best. That is not a criticism; it is a structural incentive.

For teams evaluating on-premise deployments, there are trade-offs between local data control and dependence on security standards defined elsewhere. AI-RADAR offers analytical frameworks at /llm-onpremise to assess these aspects, without pointing to a single path as best.

The queue of hundreds and hundreds of companies in vetting can be read as a sign of vitality, but also as a strategic resource: whoever decides who enters also decides, in part, which incidents and which practices become shared priorities. For now, OpenAI remains outside the door, while Nvidia pushes to “bring reality” to the regulatory debate. But reality is also made of filters.