- Anthropic, labeled a ‘supply chain risk’ by the Pentagon, will likely keep supplying Claude to the NSA, per New York Times reporting via The Decoder.
- White House Chief of Staff Susie Wiles personally approved the arrangement.
- Intelligence agencies lack enough Nvidia Grace Blackwell chips to run current OpenAI models; Anthropic‘s new Mythos runs on older chips, making it the only short-term fix for NSA classified networks.
- The contract bars Mythos from processing Americans’ data; the ‘any lawful use’ language that derailed earlier talks is reportedly not part of the deal.
What Happened
Anthropic — labeled a “supply chain risk” by the Pentagon — will likely keep supplying Claude to the National Security Agency, The Decoder reported on Thursday, citing the New York Times. White House Chief of Staff Susie Wiles personally approved the arrangement.
Why It Matters
The contract resolves a months-long friction between Anthropic’s responsible-deployment policies and U.S. national-security procurement. The Pentagon had previously classified Anthropic as a supply-chain threat because the company refused to release its technology for “any lawful use” — Anthropic’s standard contractual term that restricts uses it considers high-risk. The NSA contract works around that constraint with a specific carve-out.
The operational driver is straightforward: U.S. intelligence agencies do not have enough Nvidia Grace Blackwell chips to run current frontier models from OpenAI and others. Anthropic’s new Claude Mythos model runs on older chips too — making it the only short-term fix for NSA’s classified-network requirements. The chip-scarcity constraint is severe enough that the White House approved the arrangement personally.
Technical Details
A contract now being finalised includes a clause barring Mythos from processing Americans’ data — addressing the privacy concern that has dogged AI-NSA partnerships. The controversial “any lawful use” language that derailed earlier talks is reportedly not part of the deal. The White House wants to use this contract as a template for future agreements with other AI providers.
Separately, the White House has approved $9 billion for new AI chips to address the chip-shortage longer term. Congress still needs to sign off. The Grace Blackwell platform — Nvidia’s combined CPU-GPU architecture, with the B200 GPU and Grace CPU — is the current state-of-the-art for AI inference at the scale required for classified-network deployments.
Who’s Affected
The NSA gains operational access to current-tier AI capability on its classified networks. Anthropic gains a public-record validation that its responsible-deployment framework can coexist with national-security contracts. Pentagon procurement officials gain a template for future AI-vendor arrangements that work around the standard contractual constraints. OpenAI, Google DeepMind, xAI, and other frontier labs face a precedent that their own NSA-supply terms can be negotiated against. The Bloomberg-reported Mistral CEO Arthur Mensch warning to France against allowing Anthropic Mythos to scan military code bases gains a counter-narrative.
What’s Next
The contract is reportedly being finalised; specific signing date and contract value have not been disclosed. The $9 billion chip-purchase package needs Congressional approval. The broader pattern — White House personal approval of AI-supply arrangements with companies that have responsible-deployment constraints — is expected to continue as similar arrangements are negotiated with OpenAI, Google, and others.
