Classified Estimates Show the NSA Is Paying Billions to Test AI Models

Classified intelligence estimates show the National Security Agency is spending billions of dollars this year to evaluate and test frontier artificial intelligence models, a sum far larger than previously known and signaling that comprehensive federal AI oversight could cost tens of billions per year. The NSA’s Artificial Intelligence Security Center launched intensive testing after a series of high-profile hacking and security incidents; those efforts are funded through classified portions of the national security budget, with exact totals undisclosed. The development sharpens a policy debate: the White House has resisted broad regulation, some tech leaders propose industry self-review, and lawmakers and experts are weighing alternatives ranging from third-party audits to levying taxes or developer fees to shift the financial burden away from taxpayers.

The biggest drivers of the NSA’s expense are massive computing power to run and stress-test advanced models and the personnel costs to recruit top AI engineers, where private-sector pay packages dwarf government offers. Prior Congressional Budget Office estimates for proposed oversight centers were orders of magnitude lower (roughly $20 million per year or $36 million over five years), underscoring how resource-intensive rigorous evaluation is. Industry discussions about a voluntary standards body between leading labs coexist with calls for independent, government-capable testing; several experts argue that such assessments are public goods but insist developers should help fund the necessary computing resources and expertise.