The National Security Agency is spending billions of taxpayer dollars this year to evaluate and test advanced artificial-intelligence models, The Washington Sun reports, citing two sources familiar with classified intelligence estimates.
The exact amount spent so far remains unclear. The Pentagon declined to discuss the NSA’s AI spending, technical architecture or allocation of resources. Because the estimate is classified and comes through anonymous-source reporting, it should be understood as an attributed account rather than a publicly audited total.
Compute changes the price of oversight
One source told the Sun that additional computing power has been the NSA’s largest AI-related cost. Evaluating a frontier model can require extensive runs across powerful chips, specialized test environments and repeated attempts to expose failures or vulnerabilities. Those expenses rise with the size and capability of the systems being tested.
Specialized personnel are another cost. The government needs engineers and evaluators who understand advanced models, cybersecurity and national-security systems, but it competes with private companies offering compensation that federal agencies cannot easily match.
The reported total is dramatically larger than earlier public estimates for narrower oversight proposals. The Congressional Budget Office estimated that one proposed federal center focused on AI risks would cost roughly $20 million annually. Another House proposal for an AI reporting and tracking system carried an estimated cost of $36 million over five years.
The NSA figure covers a different and partly classified mission, so those numbers are not direct comparisons. Still, the Sun reports that lawmakers view the spending as a warning that a comprehensive federal evaluation regime could cost tens of billions of dollars each year.
Who should pay for independent testing?
The NSA says its Artificial Intelligence Security Center works to protect training data, model frameworks, capabilities and the broader machine-learning lifecycle. The center collaborates with industry, academia, the intelligence community and other government partners. Its public materials describe that mission but do not disclose the spending reported by the Sun.
The cost has added urgency to a policy dispute over who should fund model testing. One approach would leave much of the work to AI companies. Another would build independent government capacity with public money. The Sun also describes proposals to assess frontier-model developers and use that revenue to pay for the computing resources and expertise needed for outside audits.
No such funding model has been established in the reporting. For now, the NSA’s work is funded through classified portions of the national-security budget, leaving lawmakers to decide how much independent testing the government should perform and who should bear its rapidly increasing cost.
Technology