Meta is now the sole holdout. More than three weeks after President Trump signed Executive Order “Promoting Advanced Artificial Intelligence Innovation and Security” on June 2, 2026, every major U.S. AI lab — OpenAI, Anthropic, Google DeepMind, xAI, and Microsoft — has agreed to submit frontier models for voluntary 30-day pre-release security reviews. Every lab except Meta.
The Trump administration is now openly pressing Meta to comply, according to multiple reports from June 24. The standoff crystallizes a tension that has been building since Meta pivoted to a dual strategy earlier this year: proprietary Muse Spark (via Meta Superintelligence Labs, led by Alexandr Wang) on the frontier side, and open-weight Llama for the broader ecosystem.
Open Weights, Open Question
The specific friction point is Meta’s open-weight strategy. The Executive Order’s review framework asks developers to grant the federal government — specifically the NSA and the newly formed Cybersecurity and AI Safety Institute (CAISI) — 30 days of pre-release access to “covered frontier models.” For companies shipping API-only models like GPT-5.5 or Claude Fable 5, this is bureaucratically manageable. For Meta, which has built the world’s largest open-weight AI ecosystem — 1.2 billion Llama downloads as of early 2026 — a 30-day review delay cuts directly against the speed and openness that define the Llama brand’s competitive positioning.
The Llama ecosystem now encompasses over 25 cloud partners (including NVIDIA, Databricks, and Snowflake), tens of thousands of fine-tuned variants on HuggingFace, and inference engines like llama.cpp and Ollama that have lowered the barrier to local deployment. A government review process — even a voluntary one — introduces friction that Meta’s open-weight playbook was never designed to accommodate.
The Broader Crackdown Context
Meta’s resistance arrives at a moment of intensifying government scrutiny of AI model releases. On June 12, the administration issued an export control directive that effectively blocked Anthropic from deploying Claude Fable 5 and Mythos 5 to users in adversarial nations. On June 22, Anthropic ended free access to Fable 5 entirely. The message from Washington is clear: frontier AI is now a national security asset, and the era of unrestricted release is ending.
For Meta, the calculus is more complex. The company has bet tens of billions — reports suggest up to $145 billion in 2026 capex — on AI infrastructure, including its custom MTIA chips and a sprawling partnership ecosystem. The open-weight Llama brand is central to that investment thesis. Ceding control to a government review process, however voluntary in name, risks undermining the very openness that makes Llama strategically valuable.
What Comes Next
The Executive Order explicitly states it “does not create any mandatory licensing or preclearance requirement for AI models” — the framework is voluntary by design. But the political pressure is real. With Anthropic’s models already caught in the export-control net and OpenAI, Google, and xAI having signed on, Meta’s continued refusal looks increasingly isolated.
Whether Meta ultimately bends — or whether the Llama open-weight strategy becomes a new front in the battle between AI innovation and national security — remains the open question of the moment. For the millions of developers who rely on Llama as the infrastructure layer of the open-weight world, the answer matters enormously.