A shortage of NVIDIA Grace Blackwell chips means many of the newest large models cannot be deployed on classified U.S. networks, making Anthropic’s Mythos — compatible models a near-term option to keep Claude accessible to the NSA.
A shortage of NVIDIA’s latest Grace Blackwell chips has left U.S. intelligence agencies unable to run the newest large models on classified networks, creating an operational gap that has positioned Anthropic’s offerings as a near‑term fix. Recent reporting indicates Anthropic’s newer Mythos model can run on older accelerators already present in secure environments, which matters because agencies must use models compatible with the hardware on their classified systems.
Intelligence customers lack sufficient quantities of Grace Blackwell hardware to deploy many current commercial models, including those from major vendors that require the newest accelerators. As a result, models that can operate on legacy chipsets are among the few immediately deployable options for sensitive networks, and agencies are prioritizing compatibility with existing infrastructure over adopting the latest architectures until supply improves.
Anthropic’s Mythos compatibility has made it a candidate to fill that short‑term need on NSA classified networks. According to the reporting, the arrangement to keep Anthropic models available to the agency was personally approved by White House Chief of Staff Susie Wiles. That operational decision has proceeded even as the Pentagon has labeled Anthropic a “supply chain risk,” highlighting a tension between immediate mission requirements and defense assessments of vendor risk.
The Pentagon’s designation reportedly stems from Anthropic’s earlier refusal to permit its technology to be used for “any lawful use.” Negotiators are finalizing a contract that removes that broad language and replaces it with a narrower clause explicitly barring the contracted model from processing Americans’ data. The White House intends to use the finalized agreement as a template for future contracts with commercial AI companies, signaling a bid to standardize enforcement mechanisms and carve‑outs for intelligence‑grade deployments of generative models.
To address the longer‑term hardware shortfall, the White House has approved a $9 billion fund to procure new AI chips, but that funding still requires congressional authorization. For builders and infrastructure teams, the immediate takeaway is concrete: classified networks remain constrained by accelerator availability, and short‑term operational choices may rely on narrow contractual safeguards to enable mission continuity even when suppliers have been flagged as supply‑chain risks.
Sources
Replies (0)
No replies in this topic yet.
