The US government has run into an awkward little problem: it banned Anthropic from Pentagon supply chains, then discovered it still needs the company’s AI to do the job. The NSA keeps using Anthropic models because modern chips and secure infrastructure are not yet available in enough quantity to replace them. In other words, national security is bumping into a silicon shortage.
The core issue is simple. Intelligence work now depends on systems that can sort through intercepted communications, satellite images, anomalies, and threat signals at a scale no human team can match. But the computing gear needed to run those models inside government-controlled environments is still scarce, which leaves agencies choosing between policy purity and operational reality. They are choosing operational reality.
Why Anthropic is still in the room
Anthropic has become one of the few companies whose models are considered advanced enough for serious threat analysis, even while officials treat it as a supply-chain risk because of ownership structure and foreign investment concerns. That contradiction is doing all the work here: the Pentagon can blacklist a vendor on paper, but it cannot conjure a replacement if the alternative is ”wait for the data center to be built.”
According to the New York Times, the White House has approved an emergency $9 billion allocation for specialized data centers for intelligence and military agencies. Those facilities will be fitted with Nvidia Grace Blackwell superchips, which need specialized power delivery and liquid cooling, so this is not a weekend retrofit. Until that infrastructure exists, the NSA keeps using Anthropic anyway.
The chip shortage is now a policy problem
The shortage is spilling far beyond the intelligence community. The US government is also redirecting $800 million from other budgets for immediate computing purchases, while memory makers such as Samsung, SK Hynix, and Micron shift resources away from consumer electronics toward AI demand. That means the shortage is not just slowing procurement; it is bending the entire supply chain around high-end AI hardware.
- NSA is still using Anthropic models despite the Pentagon blacklist.
- The White House has approved $9 billion for dedicated intelligence and military data centers.
- Another $800 million is being redirected for near-term compute buys.
- The new facilities are meant to run Nvidia Grace Blackwell superchips with liquid cooling.
A company that is banned and indispensable
The twist is that Anthropic is not just surviving this contradiction; it is benefiting from it. The company’s Claude model is seen as strong enough for sensitive analysis, and its Glasswing project reportedly identified more than 10,000 critical vulnerabilities in a month for Western intelligence services. At the same time, its revenue growth and lofty IPO chatter show just how central it has become to the AI arms race.
The real question is whether the US can build enough sovereign compute fast enough to stop relying on vendors it does not fully trust. If it cannot, this ban will keep looking a lot like theater, and the next emergency exception may not be an exception at all.