White House Super Intelligence Accord: who signed, what it means, and the open-source AI debate

White House Super Intelligence Accord: The Open Source Wildcard

The White House dinner with major AI lab CEOs—Dario Amodei, Mark Zuckerberg, Greg Brockman, Elon Musk, Jensen Huang, and Sundar Pichai—produced the White House Accord on Super Intelligence, a commitment to internal safety controls, independent audits, and board-level oversight for AI safety. But the real tension isn't what got signed. It's who wasn't in the room and what happens to open source in the next six to twelve months.

The attendee list itself reveals a fault line. Jensen Huang and Nvidia back open source models; Anthropic sits on the far end, skeptical. OpenAI signed the open source letter but remains focused on frontier pacing, a position somewhat at odds with the openness agenda. Notably absent: George Hotz and Linus Torvalds, who represent the actual open source movement—which by definition has no business seat at a policy table.

The cybersecurity case for restriction

Max Kahn at Semi-Analysis flagged a material risk: the gap between open source model release and the discovery of zero-day exploits has collapsed. Models like Mistral can now be weaponized for hacking at nearly frontier speed. The timeline matters. Nobody at the White House dinner will push back hard on open source in the next few months because Trump's closest advisors have been evangelizing for it. But six to twelve months out, a coalition for restriction could form—not on economic grounds, but on cybersecurity.

The mechanics of control exist at multiple layers: inference restrictions, data center-level monitoring, chip-level safeguards. A gaming PC won't be sufficient to run frontier-grade attacks; you'd need serious GPU infrastructure, probably at rack scale. That hardware requirement opens a KYC angle: data centers can be monitored or gated; distributed laptops cannot.

The oligopoly scenario

There's a plausible path where everyone at the dinner table ends up against open source, but profits stay distributed. If models become closed source but multiple companies (OpenAI, Anthropic, Google) compete fiercely for enterprise and consumer access, and if front-end applications splinter across providers, revenue distribution could look like a competitive oligopoly rather than a single-winner outcome. That was Jensen's original worry—a world where one company captures all the value. A shifted world where five companies share it fairly evenly might command consensus.

The open source community wouldn't survive that shift. But the companies would sign a letter about cyber risks and move on.

The enforcement problem

Banning open source entirely is mechanically difficult. A coordinated refusal by major hosting providers (GitHub, Hugging Face) could choke distribution. But a determined adversary can download weights locally and run them on owned hardware. That gap—between corporate-level enforcement and actual enforcement against nation-states or organized crime—is the real unresolved tension.

What's notable is that this dinner focused on pacing the frontier and safety frameworks, not open source policy. The accord itself doesn't address it. The signal, instead, comes from the absence of open source advocates and the presence of executives who stand to benefit from a closed-source world—provided the rules are written to reward all of them.