The National Security Agency is reportedly spending billions of dollars this year to test advanced AI models, according to The Washington Sun, citing two sources familiar with the classified estimates. Computing power is the biggest expense, with staffing adding to the bill because the NSA has to compete with the pay packages AI labs offer.
Lawmakers expect that full-scale AI oversight could cost tens of billions of dollars a year, according to the sources. Earlier estimates were far lower, with the Congressional Budget Office putting the cost of a bipartisan bill to create a center for AI risks at about $20 million a year.
Paying by the token means paying full price for compute. AI providers' pricing models show why computing power costs large customers so much. Consumer flat-rate subscriptions are heavily subsidized, with a maxed-out $200 ChatGPT Pro subscription worth up to $14,000 at API list prices, while Claude Max, which costs the same, comes to up to roughly $8,000.
Businesses, on the other hand, increasingly pay based on usage. OpenAI bills new enterprise contracts in tokens, and at Anthropic, 75 to 85 percent of revenue comes from usage-based contracts, according to SemiAnalysis. Agent workflows can use up to a thousand times more tokens than a regular chat.
Providers are responding. OpenAI now lets some large customers pay only for completed tasks, and the company introduced GPT-6 Sol and Luna, models priced at half the usual API rate. That doesn't apply to the most powerful systems, which remain expensive, and providers increasingly sell their capabilities through closed cybersecurity programs.
Business customers appear to be lucrative for providers. SemiAnalysis estimates the gross margin on Anthropic's API business at more than 80 percent. The companies need that money, though, with OpenAI planning to spend about $856 billion on computing power through 2030, and Anthropic planning its IPO for November at the latest.
Source: thedecoder
