NSA Spending Billions on Frontier AI Model Testing

NSA AI Testing Costs Reach Billions

The National Security Agency (NSA) is spending billions of dollars in taxpayer funds this year to evaluate and test advanced artificial intelligence models. According to sources familiar with classified intelligence estimates, these expenditures are significantly higher than previous public estimates for AI regulatory oversight, suggesting that a comprehensive federal AI regulatory system could eventually cost the U.S. government tens of billions of dollars annually.

These efforts are being led by the NSA’s Artificial Intelligence Security Center, which began testing frontier models to identify potential national security vulnerabilities following a series of high-profile hacking and security incidents. Currently, these operations are funded through classified portions of the federal national security budget.

Primary Cost Drivers: Compute and Talent

The high price tag of the NSA's AI initiatives is driven by two primary factors: the extreme cost of hardware and the competitive market for specialized personnel.

Compute Infrastructure

The most significant expense is the acquisition of computing power. The surge in global demand for AI chips has led to insufficient production, driving prices up. For context, industry reports indicate that companies like Anthropic have pursued computing deals that could cost as much as $517 billion.

Specialized Personnel

Hiring top-tier AI engineers is a major financial challenge for the government. Frontier AI labs have offered salaries in the hundreds of millions of dollars to lure top talent, far outpacing the government's ability to match such pay packages, despite the NSA being regarded as having the deepest technical roster in the federal government.

The Debate Over Regulatory Funding Models

The scale of the NSA's spending has intensified the debate over who should bear the financial burden of AI safety and oversight.

Government vs. Industry Funding

Previous bipartisan proposals, such as the AI Security and Innovation Act, were estimated by the Congressional Budget Office (CBO) to cost only $20 million per year. The current reality of the NSA's spending reveals a massive gap between those early estimates and the actual cost of rigorous testing. This has led to calls for frontier labs to shoulder the cost of reviewing their own models.

Proposed Funding Alternatives

• Industry-Led Standards: Some leaders, including Elon Musk, have suggested that frontier labs review each other's models without government oversight. Google, OpenAI, and Anthropic are reportedly working on a joint plan to create an AI safety-focused "standards body."
• AI Taxes: Some experts propose levying a tax on AI companies to fund independent government safety operations, ensuring that the financial burden does not fall on taxpayers.
• Developer-Funded Audits: Experts from the AI Verification and Evaluation Research Institute suggest that assessments on frontier developers should be required to fund independent audits and the necessary computing resources.

Community Perspectives and Critical Analysis

Discussion among technical observers suggests skepticism regarding the official "testing" narrative and concerns over the application of these models in intelligence operations.

Skepticism of "Testing"

Many observers believe the billions spent are not merely for safety testing but for active offensive and defensive operations.

"They would have to be incompetent if they were not using them both for security penetration as well as social engineering / election meddling as everyone else is."

Intelligence Applications

Technical commentators highlight the utility of Large Language Models (LLMs) for mass surveillance and data processing. Specifically, the ability of LLMs to perform zero-shot classification makes them highly effective for monitoring communication streams, particularly in foreign languages, and extracting "dangerous" targets from massive troves of intercepted data.

Accountability Concerns

There are concerns that the NSA's AI activities will operate with little to no oversight, mirroring historical precedents of mass surveillance programs that were legalized ex-post-facto under the guise of national security.