NSA와 기타 미국 정부 기관들이 고급 AI 모델 테스트를 위한 컴퓨팅 파워에 수십억 달러를 지출하고 있으며, 공공 및 민간 기관 전반에서 비용이 급격히 상승하고 있습니다.

The National Security Agency is reportedly spending billions of dollars this year to test advanced AI models, according to The Washington Sun, which cited two sources familiar with classified estimates. Computing power represents the largest expense, with staffing costs adding substantially as the NSA competes with private-sector compensation packages for talent.

These costs have prompted lawmakers to significantly raise expectations for federal AI oversight. Full-scale oversight could cost tens of billions of dollars annually—far higher than earlier Congressional Budget Office estimates of approximately $20 million per year for a bipartisan center on AI risks.

AI providers' pricing structures reveal why computing costs burden large customers. Consumer subscriptions operate at a loss: a maxed-out $200 ChatGPT Pro subscription is worth up to $14,000 at API list prices, while Claude Max, priced identically, is worth approximately $8,000. Enterprise customers face different economics. OpenAI bills new enterprise contracts by token consumption, and at Anthropic, 75 to 85 percent of revenue derives from usage-based contracts, according to SemiAnalysis. Agent workflows can consume up to a thousand times more tokens than regular chat interactions.

Providers are adjusting their offerings in response. OpenAI now permits some large customers to pay only for completed tasks and has introduced GPT-6 Sol and Luna, priced at half the standard API rate. These discounts do not extend to frontier models, which remain expensive; providers increasingly distribute cutting-edge capabilities through closed cybersecurity programs.

Enterprise AI remains highly profitable. SemiAnalysis estimates Anthropic's API business maintains a gross margin exceeding 80 percent. This profitability supports aggressive expansion: OpenAI plans to spend approximately $856 billion on computing power through 2030, and Anthropic is targeting an IPO by November at the latest.

Health care faces a different cost dynamic. Rather than AI system pricing, the primary driver is how hospitals and insurers use AI in billing disputes. The Blue Cross Blue Shield Association documented that hospitals using AI-assisted coding inflated revenues by nearly $1 billion over two years. According to the report, hospitals coded cases as more severe and added secondary diagnoses, collecting nearly $12,000 more per case on average without any apparent change in treatment.

Hospitals benefit substantially from this approach. McLaren Health Care generates an additional $1 million per month using software from SmarterDx, according to CFO Dave Mazurkiewicz. The vendor receives a share of the incremental revenue. Insurers, meanwhile, use AI to comb through records for claims they can deny, Mazurkiewicz reports.

Both sides now cycle through multiple rounds of dispute because, as Caroline Pearson of the Peterson Health Technology Institute observes, it has become "cheap" to do so. Writing in Health Affairs, former US health official David Brailer predicts AI will increase overall health care spending, similar to the cost growth that followed electronic health records adoption.

For federal AI policy, the question of who bears the cost has already emerged. Nat Purser of the AI Verification and Evaluation Research Institute proposes a developer levy to fund independent testing. According to The Washington Sun, both Anthropic and OpenAI have indicated they might be willing to pay for expanded government oversight.