US intelligence agencies are racing to secure the computing power needed for next‑generation surveillance and analysis, even as their dependence on a blacklisted AI provider exposes deep tensions in US technology and security policy.
Early warnings and a growing AI crunch
For years, the CIA and NSA have struggled to run the latest AI models on their classified networks, lacking the advanced chips and data center capacity modern systems require.1 Frontier models now demand processing power far beyond what legacy government infrastructure was built to deliver, slowing efforts to fully install or test cutting‑edge tools.2
Blacklisting Anthropic, then relying on it anyway
The contradiction came into sharp focus when the Pentagon formally designated Anthropic a “national security supply chain threat,” effectively blacklisting the AI company.2 Yet, because there was “no alternative,” the NSA continued using an advanced Anthropic model to support operations, under authorization from White House chief of staff Susie Wiles.2
This stopgap was driven by a critical shortage of high‑end chips needed to host other frontier models on secure, air‑gapped systems. Intelligence officials framed the resulting reliance on Anthropic as a temporary but necessary compromise to avoid weakening national security.
A $9 billion push for sovereign spy compute
In response, the White House backed a secret $9 billion emergency proposal “to buy AI chips for spies” and build specialized federal data centers tuned for Nvidia’s Grace Blackwell superchip infrastructure.12 These facilities would require massive electrical power and liquid cooling and cannot run on standard government computing grids.2
While Congress must still approve the package, the administration has already begun redirecting $800 million from other budgets to start purchasing capacity, motivated by fears that China could seize a decisive computational edge in intelligence operations.2
Supporters argue the funding is a national security imperative to analyze vast troves of intercepted communications and satellite imagery. Critics warn the rush risks entrenching opaque, secretive AI infrastructure with limited public oversight.