The White House quietly approved a classified $9 billion budget request from the NSA and CIA to buy Nvidia’s top-tier Grace Blackwell chips, building AI infrastructure for national security. Past underfunding and the extreme security requirements of air-gapped classified networks had left spy agencies' computing capacity years behind. The White House is also allowing the agencies to reshuffle $800 million in existing funds to address the hardware crunch, clearing the way to run cutting-edge language models.
To bridge the immediate software gap, White House Chief of Staff Susie Wiles authorized the NSA to bypass Pentagon supply chain restrictions and procure Anthropic’s latest model, Mythos. Trained in April 2026, Mythos excels at offensive and defensive cybersecurity — finding and weaponizing vulnerabilities — so it was kept from public release. Because Mythos runs efficiently on older chips, the deal lets intelligence agencies deploy it directly on existing classified servers, sidestepping the wait for new chip-equipped data centers.
The NSA is now negotiating a final contract with Anthropic. Talks had stalled over usage restrictions: the Department of Defense insisted on a clause allowing the military to use the model for "any lawful purpose." Since intelligence agencies are legally barred from spying on U.S. citizens, the new contract will drop that clause and instead include a specific exemption ensuring Mythos is never used to analyze or monitor Americans' data. The agencies hope the deal will serve as a template for similar classified AI arrangements with OpenAI down the line.