Intelligence doesn’t come cheap as AI drives up costs for the NSA, hospitals, and insurers
Automation was sold as a way to cut labor bills. 2026 is teaching a different lesson. Powerful AI shifts costs into compute, audits, and a new kind of commercial arms race. The bills show up in surprising places, including classified testing programs, enterprise API invoices, and routine hospital claims.
Where the new costs come from
There are three distinct drivers to keep in mind.
- Compute and scale. Running and validating frontier models requires large pools of GPUs/TPUs and sustained usage. That hardware, plus the cloud or on-prem capacity to feed it, is the single largest line item for high-assurance work.
- People and process. Building secure evaluation pipelines, staffing expert reviewers, and maintaining continuous red-teaming or validation programs are labor-intensive and expensive.
- Commercial token economics. Many enterprise deals bill by token (units of input/output text). Token use is a useful proxy for inference compute on the same model, but actual cost also depends on model size, context window, latency SLAs, and throughput. Multiply tokens by model weight and SLA tightness, and a small consumer-style workflow can become a big enterprise bill.
NSA testing: reported “billions”, but proceed with caution
Trade reporting, specifically The Washington Sun as republished in other outlets, quoted two sources saying the National Security Agency told lawmakers it is spending “billions of dollars” this year testing advanced or frontier AI models. The NSA has not publicly confirmed that figure, and the reporting rests on anonymous sources. Treat the number as reported, not audited.
That report stands in stark contrast with an earlier civilian estimate from the Congressional Budget Office. The CBO put the cost of implementing a specific civilian AI-risk center in H.R. 9363 at roughly $20 million per year (CBO publication 62730). The two numbers are not directly comparable. The CBO figure covers a targeted civilian program, while the classified testing programs described in the trade reporting imply larger hardware and security requirements. The gap shows how quickly costs scale once you move beyond a small pilot.
Practical point: high-assurance, classified evaluation requires compute, secure facilities, and specialist staff. Those factors make agency-level budgets look very different from an office in a civilian agency doing rulemaking or limited testing.
Enterprise token economics: consumer plans are subsidized, enterprise is not
Consumer subscriptions feel cheap for a reason: they are often cross-subsidized or designed for light, casual usage. Industry analysts at SemiAnalysis produced headline-grabbing comparisons that show how misleading consumer pricing can be when applied to heavy enterprise use. Their work, which compares typical consumer-subscription usage to public API list prices, estimates (as an analyst scenario) that a $200-per-month consumer plan can correspond to thousands of dollars of list-priced API consumption (SemiAnalysis estimated up to $14, 000 for a $200 ChatGPT Pro equivalent and roughly $8, 000 for a $200 Claude Max equivalent). Those are analyst scenarios based on assumed usage and list pricing; enterprise contracts commonly include substantial discounts and different billing constructs.
An agent that loops through steps, calls tools, keeps long context windows, or performs retrieval-augmented operations can consume far more tokens than a single-turn chat. SemiAnalysis and other commentators have warned of worst-case scenarios where agentic patterns use orders of magnitude more tokens, with figures like “up to a thousand times more tokens” circulating as illustrative extremes. To make that concrete: a single chat might be 2, 000 tokens. A long-running agent that runs 1, 000 internal steps at about 2, 000 tokens per step could reach roughly 2 million tokens, a 1, 000× increase. That math shows how quickly token-based billing compounds.
Vendors are experimenting with alternative commercial models, such as task-based pricing, capped charges for agent runs, and lower-cost model tiers. OpenAI and others have previewed cheaper variants of some models (reported in trade press under names like Sol and Luna); industry reporting also says some providers are willing to offer outcome- or task-based billing to select enterprise clients. These moves lower risk for buyers, but frontier-class models, the largest, most capable systems, remain expensive to operate and are often restricted to higher-priced arrangements.
Health care: AI-assisted coding, incentives, and an iterate-and-appeal loop
The incentive structure in U.S. health care makes the cost effects of AI especially visible. The Blue Cross Blue Shield Association (BCBSA) released an analysis finding that adoption of AI-assisted coding and documentation tools correlated with about $942 million in additional insurer plan costs over two years. The report says hospitals coded cases as more severe and added secondary diagnoses and that, on average, hospitals collected roughly $12, 000 more per case after those coding changes. The BCBSA figure is a correlation reported by the association; it does not by itself prove intent or inappropriate care decisions.
Hospitals and vendors have financially benefited in some cases. McLaren Health Care’s CFO Dave Mazurkiewicz told reporters his system brings in an extra $1 million a month using software from SmarterDx. Vendors sometimes enter revenue-share deals with hospital customers. Insurers, for their part, are using analytics and AI to flag suspicious claims, which creates a workflow of denial, appeal, resubmission, and re-review. As Caroline Pearson of the Peterson Health Technology Institute put it:
“Both sides now go several rounds per dispute ‘because it’s cheap.'”
The result: what could have been a one-time documentation efficiency becomes ongoing administrative churn and higher net payments. The BCBSA analysis quantifies the scale of the correlation. Sorting how much of the added revenue reflects improved coding accuracy, versus revenue-driving upcoding enabled by automation, requires case-level audits and vendor-by-vendor review.
Policy options and industry signaling
Who pays for independent testing, secure infrastructure, and continuous validation remains an open policy question. Nat Purser of the AI Verification and Evaluation Research Institute has proposed a levy on developers to fund independent evaluation and oversight. That suggestion is meant to make developers contribute to the public good of independent testing. Trade reporting also notes that some developers, Anthropic and OpenAI among them, according to those reports, have signaled willingness to discuss funding oversight. Those signals are not formal commitments and would require governance guardrails to avoid conflicts of interest.
One more data point circulating in trade press: the Financial Times reported a large multi-year compute spending expectation for OpenAI, phrased as about $856 billion on computing power through 2030. That figure is extraordinary and has prompted follow-up questions about scope and context. Treat it as a reported projection that requires close scrutiny of the FT reporting and any clarifying statements from the company.
What leaders should do now, concrete steps and KPIs
For C-suite and IT leaders the short answer is: treat AI as infrastructure with measurable operating costs, not as a free add-on. Practical, implementable steps:
- Measure tokens per transaction. Put a KPI on tokens-per-customer-interaction (or tokens per automated workflow) and track it in your usage dashboard. Use vendor dashboards (for example, your provider’s usage console) or open-source tokenizers to compute tokens before production runs.
- Simulate agent loads before scale. Run synthetic tests. Model a representative multi-step agent (for example, 10-100 steps with 1, 000-5, 000 tokens per step) and extrapolate monthly spend at projected concurrency. That stress test will reveal whether a pilot could explode into an unexpected bill.
- Negotiate outcome- or cap-based pricing for agents. Ask vendors for task-based pricing, credits per completed automation, committed-volume discounts, or hard monthly caps for exploratory agent workloads.
- Govern AI-assisted billing tightly. Health systems should treat AI coding tools as compliance risks: random audits, clinical validation samples, vendor revenue-share transparency, and contractual penalties for demonstrable upcoding should be standard.
- Control rate and scope. Use rate limits, model-size routing (send high-volume routine tasks to smaller cheaper models), caching, and local inference where appropriate to reduce token spend.
- Budget for validation. If you build or rely on high-assurance validation (security, fairness, safety), include a recurring budget line for compute and expert staffing, those are real and recurring costs, not one-off investments.
Example calculation (illustrative): a 10-step agent averaging 2, 000 tokens per step equals 20, 000 tokens per interaction. If your projected volume is 10, 000 interactions per month, that’s 200 million tokens total. Convert that figure using your vendor’s list or negotiated per-token rate to estimate monthly spend before you sign up.
Key takeaways, questions you should be asking
-
How big is the government bill for testing frontier AI?
The Washington Sun reported that the NSA told lawmakers it is spending “billions of dollars” this year testing advanced models (the claim is based on two anonymous sources and has not been publicly confirmed). By contrast, the Congressional Budget Office estimated about $20 million per year for a targeted civilian AI-risk center (CBO publication 62730); the two numbers cover different programs and are not directly comparable.
-
Are consumer AI subscriptions hiding the real cost of enterprise use?
Analyst work (SemiAnalysis) warns that comparing consumer subscriptions to API list pricing can be misleading: a $200 consumer plan could equate to thousands of dollars at list API rates under some usage assumptions. These are analyst scenarios based on list prices and presumed usage patterns; enterprise contracts commonly include discounts, but heavy agentic usage can still drive large bills if not managed.
-
Can AI-driven billing actually increase health-care spending?
BCBSA’s analysis found about $942 million in additional insurer costs over two years correlated with use of AI-assisted coding and about $12, 000 more revenue per hospital case on average after coding changes. The report documents a strong correlation and highlights incentive effects; determining causation and clinical impact requires deeper audits and responses from vendors and hospitals.
-
Is there momentum to make developers pay for oversight?
Proposals exist (for example from Nat Purser of the AI Verification and Evaluation Research Institute) to levy developers to fund independent testing; some companies have signaled openness to discuss funding oversight, but no binding, industry-wide scheme has been agreed.
AI can cut labor and time in many workflows, but the ledger lines matter. Compute, staffing for validation, and economic incentives around monetization can easily outweigh headline savings if leaders mistake consumer-priced access for enterprise economics. Budget for it, measure it, and contract to contain it.
Sources
Key reporting and analyses that informed this piece include The Washington Sun and downstream coverage of that reporting on NSA AI spending; the Congressional Budget Office estimate for H.R. 9363 (CBO publication 62730); a SemiAnalysis analyst note on consumer subscriptions versus API list pricing and on agentic token multipliers; the Blue Cross Blue Shield Association analysis on AI-assisted coding (reported as roughly $942 million in additional insurer costs); trade reporting on vendor pricing and lower-cost model variants (e.g., The Information); and press coverage summarizing vendor and hospital statements (including comments attributed to McLaren Health Care’s CFO Dave Mazurkiewicz and Caroline Pearson of the Peterson Health Technology Institute). Many of these are reported figures and analyst estimates. Where possible, treat extraordinary numbers as reported or estimated rather than definitive audited totals.