Anthropic May Keep Supplying Claude to the NSA Despite Being Flagged as a Supply Chain Risk by the Pentagon

Why It Matters

The deal underscores a pragmatic shift in U.S. AI policy, balancing national security needs against supply‑chain concerns, while setting a template for future government‑industry AI contracts.

Key Takeaways

• Anthropic's Mythos runs on legacy Nvidia chips, fitting NSA's constraints
• White House chief of staff personally cleared the NSA‑Anthropic agreement
• Contract bans model from processing data of U.S. citizens
• $9 billion AI chip fund pending congressional approval to end hardware shortage

Pulse Analysis

The National Security Agency’s AI shortfall highlights a broader hardware bottleneck that has forced the government to look beyond the industry’s flagship providers. Nvidia’s latest Grace Blackwell processors, while powerful, are in short supply, leaving the NSA unable to run cutting‑edge models from OpenAI or other vendors. Anthropic’s Mythos, designed to operate on older GPU generations, emerged as a stop‑gap, offering the agency a functional, classified‑network‑compatible solution without waiting for new silicon. This pragmatic choice illustrates how legacy‑compatible AI can gain strategic relevance when supply chains falter.

The White House’s direct involvement—chief of staff Susie Wiles signing off on the arrangement—signals a willingness to override prior risk designations when national security imperatives arise. By stripping the controversial “any lawful use” language and inserting a clause that prohibits processing data of U.S. persons, the contract attempts to reconcile security concerns with civil‑rights safeguards. Officials hope this framework will serve as a repeatable template for future engagements with AI firms, streamlining approvals while maintaining oversight.

Long‑term, the $9 billion AI chip investment, still awaiting congressional approval, aims to eliminate the hardware scarcity that drove the NSA to this compromise. Funding new chip production could diversify the supply chain, reduce dependence on a single vendor, and enable broader deployment of advanced models across government agencies. The episode may accelerate policy discussions around AI export controls, data governance, and the balance between innovation and security in the rapidly evolving artificial‑intelligence landscape.