AI’s bill reaches the NSA, hospitals and insurers

The bill for advanced models

The NSA’s plans, reported by The Washington Sun citing two people familiar with secret assessments, have prompted lawmakers to expect full-scale AI oversight to cost tens of billions of dollars annually. One source said computing capacity is the largest expense. Salaries add to the total as the NSA competes with AI labs’ compensation packages.

That projection is far above an earlier estimate. The Congressional Budget Office said a bipartisan bill to create an AI risk center would require about $20 million annually.

The commercial pricing behind the increase is easier to see in the gap between consumer subscriptions and actual usage:

According to SemiAnalysis, a $200 ChatGPT Pro subscription at maximum usage can correspond to up to $14 000 in computing at API rates. Claude Max, at the same subscription price, can represent up to about $8 000.

Consumer plans are heavily subsidized by their fixed fees. Businesses increasingly pay for what they use instead. OpenAI bills new enterprise contracts in tokens, while 75% to 85% of Anthropic’s revenue comes from usage-based contracts, according to SemiAnalysis. AI-agent workflows can consume up to 1,000 times more tokens than ordinary chat.

Providers are trying to narrow the gap:

The economics remain attractive for providers. SemiAnalysis estimates that Anthropic’s API business has a gross margin above 80%. At the same time, the companies need enormous amounts of infrastructure. The Financial Times reports that OpenAI plans to spend about $856 billion on computing capacity by 2030. Anthropic plans to hold an IPO no later than November.

Healthcare turns AI into a billing dispute

US healthcare is absorbing a different kind of AI cost. The issue is not primarily the price of the software, but how hospitals and insurers use it when bills are prepared and challenged.

A Blue Cross Blue Shield Association report accuses hospitals of using AI-assisted coding that produced nearly $1 billion in additional costs over two years, The New York Times reports. The report says hospitals recorded more severe illnesses and added secondary diagnoses. Each case generated nearly $12 000 more on average, without noticeable changes in treatment.

For hospitals, the incentives are direct. McLaren Health Care receives an additional $1 million per month through SmarterDx software, according to chief financial officer Dave Mazurkiewicz. The provider takes a share of the increased revenue. Insurers, Mazurkiewicz says, use AI to examine medical records and identify claims that can be denied.

Caroline Pearson of the Peterson Health Technology Institute says each dispute now involves several rounds of claims and objections because the process is cheap. Former US healthcare official David Brailer wrote in Health Affairs that he expects AI to increase total healthcare spending, as electronic medical records did earlier.

Who pays for the checking?

The NSA is already facing the same basic problem as the healthcare system: the cost of making AI accountable may grow alongside the cost of using it.

Nat Purser of the AI Verification and Evaluation Research Institute has proposed a fee requiring developers to help fund independent testing. The Washington Sun reports that Anthropic and OpenAI have indicated they may agree to pay more for government oversight.

I think the more important detail is not that AI evaluation is expensive. It is that each sector has a reason to keep expanding the machinery around the models: the NSA needs computing and specialist staff, providers need infrastructure revenue, and healthcare companies can gain money from more aggressive coding or denial systems. The reporting is quiet about how anyone will distinguish necessary scrutiny from a new layer of spending that simply becomes part of the business model.

That leaves AI with a second bill beyond the token price: institutions may pay to deploy it, then pay again to audit, contest and regulate what it does.