The US has condemned Anthropic as a security threat. Its spy agencies use Claude anyway.
The TL;DR
Chip shortages forced the NSA to continue using Anthropic’s AI despite the Pentagon’s blacklist. White House approves $9B for decentralized data centers.
The US government has a problem it cannot solve publicly. The Pentagon has officially listed Anthropic as a national security threat. The NSA uses Anthropic AI anyway because there is no other way.
White House Chief of Staff Susie Wiles has authorized the National Security Agency to continue using the improved Anthropic model, the New York Times reports. The decision was forced by a severe shortage of advanced chips needed to run frontier AI models on decentralized networks.
The agreement came alongside a secret $9 billion emergency funding request approved by the White House. The money is designed to help America’s top intelligence agencies, including the CIA and NSA, secure the advanced semiconductors needed to run productive AI in secret infrastructure.
The computing requirements of modern AI have exceeded the expectations of defense experts and congressional committees. Frontier models use more processing power than distributed networks are designed to deliver. Since the government cannot protect enough physical chips, spy agencies have not been able to fully install or test the latest AI tools.
$9 billion will fund specialized data centers built for Nvidia’s Grace Blackwell superchip infrastructure. These systems need to be custom built with high power and special liquid cooling. They cannot use standard government computer grids.
Congress will formally vote to approve the package. In the meantime, the White House is redirecting $800 million from other federal budgets to begin buying computing power immediately. The urgency is driven by fears that China will take advantage of the computer in intelligence operations around the world.
Military and intelligence services rely on AI to sift through millions of intercepted communications, satellite images, and data points. AI flags anomalies and uncovers threats that human analysts might miss. The shortage of chips that block these devices is, in the government’s analysis, a national security emergency.
The contradiction is stark. The same government that designated Anthropic as a supply chain threat is now relying on Anthropic models because it lacks the hardware to implement other things in its infrastructure. The blacklisting was prompted by concerns about Anthropic’s corporate structure and foreign investment relationships. Dependence on performance is driven by the fact that Claude is among the most capable thinking models available.
Anthropic’s Project Glasswing, which gives 50 select partners access to the Claude Mythos for vulnerability detection, found more than 10,000 critical bugs in one month. The company is simultaneously the most powerful AI security tool available to Western defense and the Pentagon’s supply chain risk management company.
The chip shortage extends beyond the intelligence community. The same reallocation of memory that is killing the cheap smartphone, as Samsung, SK Hynix, and Micron redirect wafers from consumer electronics to AI, is also hindering the government’s ability to build a decentralized AI infrastructure. The $9 billion request is the government’s version of a similar problem affecting consumer electronics around the world.
Anthropic’s revenue has increased from $9 billion to $30 billion annually between the end of 2025 and the beginning of April 2026. The company is preparing for a possible IPO later this year for as much as $800 billion. At the same time it is banned by the Pentagon and important to the NSA.
The situation reflects a structural tension in US AI policy. The government wants to control which AI companies have access to sensitive jobs. But frontier AI capabilities are concentrated in a few private companies. When the hardware to implement alternatives is not available, the government’s control over its supply chain disappears.
$9 billion is earmarked to address the hardware gap. Decentralized data centers, once built, will provide the intelligence community with the infrastructure to use any models it chooses without depending on any single provider. Until those agencies are up and running, the NSA will continue to use an AI company the Pentagon says it shouldn’t trust. The lack of chips makes that choice for them.
