The US government just admitted it has a serious problem. Anthropic, an AI company officially blacklisted by the Pentagon as a national security risk, remains a vital supplier to the NSA. The catch? The government lacks the hardware to run alternatives.

Behind the scenes, the White House approved a secret $9 billion emergency fund to buy advanced AI chips. These Nvidia Grace Blackwell superchips are critical to running the latest language models on classified networks. But the supply shortage means spy agencies can’t deploy frontier AI at scale yet.

Until these new data centers are built, the NSA is stuck using Anthropic’s Claude AI models. Despite the Pentagon’s blacklist—rooted in concerns over Anthropic’s refusal to remove ethical restrictions on use—Claude remains among the best reasoning models available. The company’s Mythos model recently found over 10,000 cybersecurity flaws in one month, proving its power.

This contradiction exposes a stark reality. The government wants control over which AI tools can access sensitive systems. But only a handful of companies deliver frontier AI. Without enough chips to run alternative models in secure environments, the government’s leverage evaporates.

To address this, Congress faces a $9 billion vote to fund specialized data centers. These facilities require massive power and liquid cooling just to operate Nvidia’s latest chips. Even with fast-tracked funding, the buildout will take time. Meanwhile, $800 million has already been reallocated for quick-start chip purchases.

The stakes couldn’t be higher. US intelligence depends on AI to sift through mountains of intercepted data and spot threats humans would miss. Falling behind China’s computational capabilities risks national security. The chip shortage isn’t just about consumer gadgets; it’s a strategic bottleneck for government AI.

Anthropic’s Ethical Standoff and Legal Fight

Anthropic’s troubles began when the Pentagon demanded unrestricted military use of its AI models. The company refused to remove clauses barring mass domestic surveillance and fully autonomous weapons without human oversight. The Pentagon responded by labeling Anthropic a “supply chain risk,” a label typically reserved for adversarial foreign companies.

Anthropic fought back, winning a preliminary court injunction against the blacklist, citing First Amendment protections. The dispute highlights a new divide in AI: companies drawing ethical lines versus those surrendering to government demands. Anthropic’s stance cost it Pentagon trust but gained significant public and market support. Its user base surged, and the company’s valuation soared toward an $800 billion IPO prospect.

Meanwhile, the NSA’s contract with Anthropic includes a clause banning the use of AI on Americans’ data. This compromise aims to ease privacy concerns while satisfying Pentagon requirements. The White House hopes this deal will serve as a template for future government contracts with AI firms.

Hardware Limits Shape AI Policy

The real bottleneck isn’t just corporate ethics or policy—it’s physical hardware. Intelligence agencies run classified AI in air-gapped environments that demand local compute capacity. These can’t tap commercial cloud resources freely. Amazon’s government cloud services, despite a $50 billion upgrade plan, won’t deliver the required chip density or security soon enough.

Until the new $9 billion data centers come online, the NSA relies on Anthropic’s older-model-compatible AI. OpenAI’s models, while powerful, require newer chips not yet available for classified use. The hardware gap forces agencies to keep using a blacklisted vendor.

It’s a textbook case of technology outpacing policy. The US wants to control its AI supply chain tightly. Yet, the frontier AI market is concentrated in a few private firms. Without the infrastructure to run alternatives, the government’s options are limited. This dynamic will likely define AI governance and national security strategy for years.

The chip shortage also reflects a global race for semiconductor supremacy. Export bans on advanced chips to China underscore the geopolitical tensions fueling this scramble. Whoever masters AI compute power holds a key advantage in intelligence and military operations.

For now, the US spy community balances risk and necessity. It blacklists Anthropic but keeps using its AI. It funds billion-dollar chip buys while scrambling for short-term solutions. The AI arms race is here, and the hardware shortage just made it messier than anyone expected.