For now, the biggest money in medical AI rides inside the test

Article 2 found that insurers pay mainly for a result, not the AI. This article asks who earns the money.

What we can count

Worldwide, we counted 58 listed (publicly traded) companies that call themselves AI companies and earn mostly from clinical work, or whose main product is AI authorized by a medical-device regulator. Forty-two are listed outside America, but the 16 listed there earn 83.1% of the money. Of their $4.1B of revenue in the latest fiscal year, mostly 2025, tests with AI inside took 61.1%; AI software sold on its own, 14.1%. Tests took more than half of this revenue in every year from 2023 to 2025; the jump to 61.1% in 2025 comes mostly from Tempus buying Ambry Genetics and from Caris's growth.

Three firms dominate: Tempus and Caris for genomic tests and iRhythm for heart monitoring, 68.6% of the total. Insurers paid about half, by our estimate, nearly all to those three, for the test rather than the AI. Among the 58, the largest firm paid for the AI itself is HeartFlow: $176M in 2025 for heart-scan analyses that providers bill to insurers. Revenue is not profit: Tempus, Caris and iRhythm all lost money in 2025.

Outside America, insurers rarely pay for the AI: China's national insurance agency folds AI help in reading a scan into the scan's fee, so Chinese firms earn from governments, hospitals, patients and drug makers. Korea's Seers is closer to iRhythm: hospitals bill national insurance for its heart monitoring.

We count listed firms because they publish audited revenue, and use this rule because anyone can check it against filings. Most of the 58 qualify by calling themselves AI companies, so the count shows what AI-branded firms earn, not everything AI earns: it misses private firms (providers spent about $600M on note-writing AI in 2025, by one venture firm's estimate), equipment makers that sell AI with their machines, and diagnostics firms such as Veracyte, Guardant and Castle that do not call themselves AI companies.

The whole American market

For the AI itself, we turn to America and count money only where the AI is priced or the billing code pays for the algorithm's output: HeartFlow's analysis is in; Tempus's genomic test, billed as sequencing, sits beside it. Our analysis, not a finding: on that rule, Americans spent about $2.9B on medical AI in 2025, plus about $0.6B that drug makers paid for data: $3.5B in all (outer bounds $2.4–5.5B). Hospitals and practices paid almost half, mostly for AI subscriptions, from note-writing to patient calls, and scanner options; insurers paid over a quarter, for results that come from AI, such as Decipher's score; drug makers paid most of the rest. Beside it sit pools where AI is only one part, led by about $2.8B of insurer-paid tests and heart monitoring, about 1.7 times what hospitals and practices paid for AI in patient care and patient access.

Who buys data and companies

Drug makers and other life-science firms pay whoever holds the data: about $470M to the 58 firms, two-thirds of it to Tempus; in our small sample, about half of hospitals' data deals with known terms paid no cash. Companies go to whoever owns the point of use, the scanner, lab or software where the AI runs: of 25 AI firms bought for a disclosed price since 2023, 17 went to device makers, imaging chains or labs. The point of use can be worth more than the AI: GE HealthCare paid $2.3B for Intelerad, an imaging-software firm that hosts other firms' AI but has no FDA-authorized AI of its own; GE expects about $270M from it in its first full year, more than HeartFlow earned.

Devices are not dollars

AI software firms hold nearly two-thirds of FDA's recent AI authorizations but earn about a third of our $3.5B US estimate. Nearly two-thirds of it goes to products FDA does not authorize: most AI lab tests, note-writing and other assistants, and data.

Who gets paid, and why

Our analysis, not a finding: the money goes to whoever owns something already paid for, such as a test, a scanner or a drug maker's data or marketing budget. In patient care, the hospital usually pays, unless the AI is paid per use in its own right, which almost only younger AI firms and test labs achieve, or someone outside the hospital pays for their own reasons: drug-company advertisers (OpenEvidence, free to clinicians), drug makers looking for patients, or patients themselves.

AI sold on its own faces hospital budgets and the verification tax, the clinician's time to check each result. Note-writing AI pays less of it, because clinicians edit a draft rather than check each finding; Microsoft already sells it widely, and Epic, which owns the record, is moving in.

So what?

• Investors: judge revenue by what pays for it and who can reprice it: in 2021 Novitas, which then priced iRhythm's Medicare claims, set rates that would have cut its 2020 revenue by 15.6%; since 2023 the risk sits in Medicare's yearly national fee rules.
• Founders and product managers: attach the AI to something already paid for: price a result per use where a billing code is covered and priced, sell time saved by subscription, and sell scanner features through the maker, or to imaging centers across makes before the makers build their own.
• Health systems: buy from the budget that gains and keep data rights if a vendor is sold; to earn, license to firms with a payer, join a data pool drug makers pay for, or spin out. Of the 40 companies behind the 79 devices with a separate payment of their own (own code, add-on or per-test price), 17 are research spinouts (firms founded on university or hospital research), against 3 of 39 sampled others (a fragile gap).

Phase 1 in three lines

1. Most FDA-authorized AI devices are bought from budgets, not billed.
2. Insurers pay for answers, not for help.
3. For now, the biggest money rides inside the test, not the AI.

Phase 2: The buyer asks which budget pays when no insurer does.

Update, October 27, 2026: the US estimate now also counts AI assistants and agents beyond note-writing, such as AI that answers patients' calls, schedules visits, handles intake and makes care-team calls. The total moves from $3.2B to $3.5B and hospitals' share from 45% to 48%; no conclusion changes.

If you've bought AI, was it its own line item, or inside something you already pay for, like a scanner, a test or the health record?

Join the discussion on LinkedIn: the post