US blacklists Anthropic as security threat but NSA keeps using Claude

At a glance:

• Pentagon officially blacklists Anthropic as a national‑security supply‑chain threat
• NSA continues to run Anthropic’s Claude model because of a critical AI‑chip shortage
• White House approves a secret $9 billion emergency package to buy Nvidia Grace Blackwell superchips for classified data centres

What happened

The Department of Defense announced this week that Anthropic, the San Francisco‑based creator of the Claude family of large language models, is now on a formal blacklist for national‑security supply‑chain concerns. The designation, issued by the Pentagon, cites worries over Anthropic’s corporate structure and foreign‑investment ties, marking the company as a potential vector for espionage or sabotage.

Despite the blacklist, the National Security Agency has been authorized to keep using an advanced Claude model for classified workloads. According to reporting by the New York Times, White House chief of staff Susie Wiles signed off on the exception after senior intelligence officials warned that no viable alternative existed on the agency’s current hardware.

Why the blacklist matters

The blacklist is not merely a symbolic gesture; it restricts Anthropic’s ability to sell its models to other federal contractors and blocks the company from participating in certain government procurement programs. The move reflects a broader trend of the U.S. government trying to tighten control over AI vendors that could expose sensitive data or supply‑chain vulnerabilities.

However, the decision also highlights a paradox. While the Pentagon is trying to limit Anthropic’s reach, the NSA’s operational need for a high‑performing reasoning model—Claude being among the most capable—forces it to sidestep its own policy. This tension underscores how limited the government’s AI‑hardware ecosystem has become.

Hardware shortage and the $9 billion emergency plan

The root cause of the NSA’s reliance on Claude is a severe shortage of advanced AI chips. Frontier models demand far more processing power than the classified networks were originally built to support. Memory wafers that would normally feed consumer devices from Samsung, SK Hynix and Micron are being redirected to AI workloads, leaving the intelligence community scrambling for capacity.

To address the gap, the White House has approved a secret $9 billion emergency funding request. The money will be used to construct specialised federal data centres built around Nvidia’s Grace Blackwell superchip infrastructure. These facilities will require custom power delivery, massive electrical capacity and liquid‑cooling systems that cannot be accommodated by existing government computing grids.

An interim $800 million has already been re‑allocated from other budgets to begin immediate procurement of computing capacity, while Congress prepares to formally vote on the full package. The urgency is driven by fears that China could capture a decisive computational advantage in global intelligence operations.

Impact on intelligence work

AI tools are now integral to the daily workflow of the CIA, NSA and other spy agencies. They sift through millions of intercepted communications, analyse satellite imagery and flag anomalous patterns that human analysts might miss. A prolonged chip shortage would blunt these capabilities, which the government has labeled a national‑security emergency.

Anthropic’s own Project Glasswing, which grants 50 select partners access to Claude Mythos for vulnerability discovery, reported finding more than 10,000 critical flaws in a single month. This makes Anthropic simultaneously a valuable security asset and a perceived supply‑chain risk, deepening the policy dilemma.

Future outlook

Anthropic’s revenue surged from $9 billion to $30 billion annualised between late‑2025 and early‑April 2026, and the company is eyeing an IPO later this year with a potential valuation of up to $800 billion. The juxtaposition of a massive market boom with a Pentagon blacklist illustrates the structural tension in U.S. AI policy: the government wants to dictate which private AI firms can touch sensitive data, yet frontier AI capability remains concentrated in a handful of companies.

The $9 billion hardware program is intended to give the intelligence community the flexibility to run any model it chooses, reducing dependence on a single vendor. Until those classified data centres become operational, the NSA will likely continue to run Claude despite the official ban, effectively letting the chip shortage dictate policy.

The situation may force a reassessment of how the United States secures its AI supply chain, potentially prompting new legislation, increased domestic chip production incentives, or a broader push to diversify AI‑model providers within classified environments.