The White House Just Cut a Secret AI Deal With OpenAI, Google, and Anthropic — And Meta Got Left Out in the Cold

Something big just happened in Washington, and I don't think it's getting nearly enough attention. The White House is finalizing a voluntary agreement with OpenAI, Anthropic, and Google that would give federal agencies up to 30 days to review new frontier AI models for national security implications before they can be publicly released. An announcement could come as early as the first week of August.

Let me explain why this is a watershed moment — and why the fact that Meta is excluded from the deal tells you almost everything you need to know about the current state of AI politics in America.

What's Actually In the Deal

The framework is voluntary, which is an important word to pay attention to. President Trump's June 2 executive order on AI and cybersecurity explicitly prohibits mandatory licensing, preclearance, or permitting for AI model development. The administration went out of its way to include that language to reassure the AI industry that Washington wasn't building a government approval regime for their products.

But "voluntary" doesn't mean toothless. The enforcement mechanism here isn't legal penalties — it's informal pressure, and the administration has already shown it's willing to use that lever aggressively. Export control threats, delayed launch approvals, and direct calls from cabinet officials have all been deployed in recent months. The AI labs know that cooperating with this framework is, in practice, the cost of doing business in the United States at frontier scale.

The specifics: when a lab is ready to release a new frontier model, federal agencies — particularly Treasury, Defense, and Homeland Security — get a 30-day window to evaluate the model's national security implications using classified benchmarks. If they raise concerns, there's a process for addressing them before the public release happens.

Why Meta Isn't In the Room

This is the detail that jumped out at me immediately. The benchmarks used to evaluate these frontier models are classified. That's not a minor bureaucratic detail — it fundamentally shapes who can be part of this framework.

Meta's entire AI strategy is built around open-source releases. Llama models go public. Weights go public. The whole point is that anyone, anywhere, can download and run Meta's models. That approach is genuinely incompatible with a classified evaluation process, because once a model is released as open weights, there's no way to un-ring that bell. You can't give the government 30 days to review something that's already been downloaded 10 million times.

So Meta is out. And that's not just a political footnote — it's a signal about the direction this administration wants to take AI governance. Closed models, controlled by identifiable companies with U.S. government relationships, are the preferred model. Open weights from a giant tech company operating on a global scale are a complication.

The Uncomfortable Truth About "Voluntary"

I want to be direct about something: this deal was drafted in close consultation with the very labs it's meant to hold accountable. OpenAI, Anthropic, and Google had significant input into the standards they'll be evaluated against. That's not unusual in regulatory contexts — industry input is often how workable rules get written — but it's worth being clear-eyed about what that means.

These three companies benefit from a framework that applies to them but not to Meta, not to Chinese competitors like DeepSeek, and not to smaller labs without the resources to navigate a 30-day federal review window. The barriers to entry just got higher for everyone except the incumbents who helped design the system.

That said, I don't think the framework is purely cynical. The national security concerns around frontier AI models are real, and some kind of government oversight process was probably inevitable. The question is whether this particular version of oversight — voluntary, classified, incumbent-friendly — is the right shape for that oversight to take.

What Happens Next

The 60-day deadline from Trump's June 2 executive order expires in early August, which is when we should expect a formal announcement. Watch for how the labs publicly respond — they'll almost certainly frame this as responsible cooperation with legitimate national security interests, and they're not entirely wrong to do so.

But also watch Meta's response. They've been the loudest voice arguing that open-source AI is actually better for American security than closed models — the logic being that open models let the entire research community scrutinize for vulnerabilities and misalignment. This deal implicitly rejects that argument, and I expect Meta to push back hard.

We're watching the first real attempt to govern frontier AI in the United States take shape in real time. It's messy, it's imperfect, and it's being written by the people with the most to gain from writing it. But it's also genuinely happening, and that matters.

What's your experience? Drop a comment below! 👇

Do you think a voluntary framework like this can actually work to keep frontier AI development safe — or does "voluntary" just mean "optional for anyone who doesn't mind the consequences"?