White House Approves Secret $9 Billion AI Chip Request for Spy Agencies, Authorizes NSA to Keep Using Anthropic

The White House has approved a secret $9 billion emergency funding request to expand AI computing infrastructure for US intelligence agencies, according to The New York Times. The request, which requires congressional approval, would fund specialized federal data centers capable of running Nvidia’s Grace Blackwell superchip infrastructure. In the interim, the administration is reallocating $800 million from other government budgets to accelerate near-term computing purchases.

Separately, White House Chief of Staff Susie Wiles has authorized the National Security Agency to continue using an advanced AI model built by Anthropic, as reported by The New York Times and confirmed by Crypto Briefing. The authorization comes despite the Pentagon having designated Anthropic as a national security supply chain threat.

The Chip Shortage Driving the Decision

US intelligence agencies are being outpaced by the computing demands of frontier AI models. The CIA and NSA rely on AI platforms to process millions of intercepted communications, satellite images, and data points, flagging anomalies that human analysts might miss. But today’s frontier models require processing power that, according to the Times of India, “far exceeds what defense experts and congressional committees anticipated just a year or two ago.”

The government cannot secure enough physical chips to install and test the latest AI tools within its classified networks. The $9 billion request targets custom data center builds that provide the massive electrical power and specialized liquid cooling required for Grace Blackwell systems. These cannot run on standard government computing infrastructure.

The Anthropic Paradox

The Anthropic authorization creates a direct contradiction in US AI policy. The Pentagon has formally designated the company as a supply chain threat, yet the White House is simultaneously approving its use by the NSA for classified intelligence work.

According to Crypto Briefing, the government and Anthropic are finalizing a classified agreement that would allow intelligence agencies to use Anthropic’s models “while preventing their deployment on Americans’ data.” The arrangement functions as a workaround: the chip shortage prevents agencies from running alternative models on their own infrastructure, so they must rely on a vendor they have formally flagged as a risk.

This is not the first time Anthropic’s relationship with the federal government has generated friction. The Pentagon blacklisting, which restricts Anthropic from certain defense supply chains, sits alongside Anthropic’s growing commercial traction with federal agencies. The company’s upcoming $30 billion funding round at a reported $900 billion-plus valuation underscores its strategic importance to the broader AI infrastructure landscape.

Urgency and Competition

The spending urgency stems from the intelligence community’s assessment that a chip shortage stalling AI deployment risks letting adversaries, particularly China, seize the computational advantage in global espionage. The New York Times reported that the White House views access to advanced AI chips and infrastructure as a major national security priority.

The $9 billion figure signals the scale of the gap between what intelligence agencies need and what they currently have. For context, the entire US intelligence community’s publicly acknowledged budget for fiscal year 2025 was approximately $73 billion. A $9 billion supplemental request dedicated to AI computing infrastructure represents a significant share of new spending directed at a single capability.

The Compute Access Question

For agent builders and AI infrastructure companies, the federal dynamic illustrates a broader tension: compute access is becoming geopolitical leverage. The question of whether a single vendor should control critical AI infrastructure for government use extends beyond Anthropic to every frontier lab. As federal agencies move from pilot programs to production AI deployments, the vendor relationships, exclusivity terms, and supply chain designations will determine which models actually run on classified networks.