Phase 1 · The money
FDA clearance is routine. The question is who pays.
Of 1,614 FDA-cleared AI devices, only 78 (4.8%) have ever had a way to be paid per use. This report shows who pays for the rest, how Medicare pays, how four countries compare and who captures the money.
October 2026 · FDA device list through June 29th 2026
The report at a glance
Five key findings
High confidence. Clearance is not reimbursement. Only 78 of 1,614 FDA AI devices (4.8%) have a separate billing code, lab code or add-on…
More, and why this confidence?
…and Medicare's physician payments on AI codes were $10.3M in 2024, 96% of it for coronary CT.
Counted from FDA and CMS files. A floor: a device counts only when a rule or the company names it for a code.
Medium confidence. Payers pay per case for AI that produces a result of its own (a measurement, a risk score, an autonomous exam).
More, and why this confidence?
AI that helps a clinician read, or works inside a procedure that is already paid, is a hospital cost in every country we checked.
Consistent across 36 payment cases in four countries and with CMS claims data, but the cases are our selection.
Medium confidence. The hospital is the main buyer, and the value tends to flow to platforms (scanner makers, EHR vendors, provincial systems) rather than to standalone AI companies.
Why this confidence?
Consistent across US, Chinese and German sources; vendors' revenue mix is not measured yet.
Medium confidence. AI labs and big tech sell to budgets and stay outside US device rules: none of OpenAI, Anthropic, Amazon, NVIDIA or Epic has an FDA AI clearance…
More, and why this confidence?
…and their policy asks are about FDA's boundary and data access, not payment.
Clearance counts are exact; products and policy come from a 176-row source table, spot-checked at source.
High confidence. Our device labels get the type of payer right but the exact mechanism only about 60% of the time, so device-level shares are a first cut.
Why this confidence?
28-device case review checked against makers, FDA and CMS; the fixes it called for were applied on 2026-10-13.
Who pays
High confidence. Only 78 devices (4.8%) are named for a billing code, a Medicare lab code or an inpatient add-on; 73 still have it in force.
More, and why this confidence?
Eight are paid like lab tests, $200-$4,500 per test, and two prosthetic-arm controllers are on Medicare's equipment fee schedule (L6700, from 2025).
From CMS rules, Medicare lab fee schedule and coverage documents, and company coding guides.
Most AI is paid from a buyer's budget, not per use
All 1,614 FDA AI-enabled devices by who pays (colour) and segment (size). Select a block to list its devices
Paid per use: 78 devices named for a billing code, lab code or add-on, plus 2 maker-run services paid per use without a code.
Source: FDA AI-enabled device list; CMS; Medicare contractors; makers' documentsCleared to Paid
Look up any of the 1,614 devices
Search by name, company or FDA number, and filter by who pays, specialty and year
Two devices show "Name withheld".
Source: FDA AI-enabled device list; Cleared to Paid who-pays labelsCleared to Paid
More charts (2)
Only 5% of devices are paid per use
FDA AI-enabled devices by who pays, % of all 1,614
Paid per use includes 2 maker-run services that some insurers pay per use without a code; 78 devices (4.8%) are named for a billing code, a Medicare lab code or an inpatient add-on. The hardware share is an upper bound: scanner features whose packaging could not be checked stay counted as hardware.
Show the numbers
| Who pays | Devices | Share |
|---|---|---|
| Paid per use | 80 | 5.0% |
| No extra payment | 65 | 4.0% |
| Inside a hardware price | 242 | 15.0% |
| From a buyer's budget | 1,227 | 76.0% |
| All devices | 1,614 | 100% |
Source: FDA AI-enabled device list; CMS; Medicare contractors; makers' documentsCleared to Paid
Standalone software without a code is the largest group
FDA AI-enabled devices by who-pays label; green is paid per use
Labels built from payment rules and makers' documents are graded A or B; labels inferred from who filed the device are grade C. In a 28-device case review the type of payer was right for every device, the exact mechanism for 17.
Show the numbers
| Label | Group | Devices |
|---|---|---|
| Separately reimbursed | Payer pays separately | 78 |
| Maker-run service | Payer pays separately | 2 |
| Folded into the exam payment | No extra payment | 36 |
| Billed under an existing service code | No extra payment | 29 |
| Sold with the scanner | Inside a hardware price | 185 |
| Sold with a medtech system | Inside a hardware price | 35 |
| Maker-run service | Inside a hardware price | 14 |
| Consumer buys the device | Inside a hardware price | 8 |
| Standalone software, no code | A buyer's budget | 937 |
| Scanner option or upgrade | A buyer's budget | 154 |
| Scanner maker's standalone software | A buyer's budget | 127 |
| Maker-run service | A buyer's budget | 9 |
Source: FDA AI-enabled device list; CMS rules; makers' filings, coding guides and product documentsCleared to Paid
More findings (4)
Medium confidence. Most cleared AI has no code, and most of it is funded from a buyer's budget (76%); 15% is paid inside a hardware price.
More, and why this confidence?
Even scanner AI is mostly sold separately: where makers' documents say, 154 of 195 scanner features are options, licences or upgrades, and only 41 are included in the machine's price. The other 144 are unverified and stay counted as hardware, so 15% is an upper bound for true bundling.
Packaging checked in makers' documents for 70 feature families (grade A or B); the rest are structural labels (grade C).
High confidence. The type of payer is reliable even where the mechanism is not: in the case review it was right for all 28 devices, but the mechanism only for 17.
More, and why this confidence?
Labels built from payment rules were all correct (6 of 6); labels inferred from the filer were weakest (1 of 4 for medtech systems, 0 of 3 for scanner makers' standalone software).
Every sampled device checked at source; per-label samples are small.
High confidence. Separate reimbursement clusters in a few places (coronary CT, eye screening, MRI measurement): only 30 of 181 device types have any reimbursed device.
Why this confidence?
Counted from the device table and code links.
Medium confidence. Two quieter routes earn nothing extra for the AI: 29 devices ride on services insurers already pay for (heart-rhythm monitoring, echo, a prostate procedure)…
More, and why this confidence?
…and 25 are run by their makers as per-case services, most paid through patient-specific implants or guides.
Linked only where a filing, coding guide or maker page says so; for 20 of the 25 maker-run services, the payment route is inferred from the offer.
Read the article: Paid for, but not by insurers (coming soon)
Medicare payments
High confidence. Medicare physician payments on AI codes grew from $0.14M (2019) to $10.3M (2024).
More, and why this confidence?
Measurement codes took $10.08M; detection and diagnosis codes $21.7K.
Medicare physician data, national rows.
Medicare paid $10.3M on AI codes in 2024, mostly heart CT
Medicare physician payments on AI billing codes, by year, $
Estimated as services times average Medicare payment. Coronary CT analysis (FFR-CT and plaque) took 96% in 2024. Codes for retinal photos read by people are excluded.
Show the numbers
| Year | Payments | Coronary CT share |
|---|---|---|
| 2019 | $141K | 100% |
| 2020 | $257K | 97% |
| 2021 | $752K | 96% |
| 2022 | $2.1M | 97% |
| 2023 | $4.3M | 97% |
| 2024 | $10.3M | 96% |
Source: CMS Medicare physician data, national rows; Cleared to Paid analysisCleared to Paid
Two codes take most outpatient AI claims
US hospital outpatient claims for AI billing codes, January to September 2025, by product type
24 AI billing codes had claims, grouped into 15 product types.
Show the numbers
| Product type | Codes | Claims |
|---|---|---|
| FFR-CT (coronary flow from CT) | 75580 | 19,352 |
| Lung nodule CT | 0722T, 0721T | 6,832 |
| Brain MRI volumes | 0866T, 0865T | 2,669 |
| Liver and organ MRI | 0649T, 0648T, 0698T, 0724T, 0697T | 1,950 |
| Coronary plaque from CT | 0625T | 1,191 |
| Autonomous diabetic eye screening | 92229 | 1,142 |
| 3D heart model for cardiac procedure planning | C9793 | 430 |
| Echo AI for heart failure (HFpEF) | 0932T | 428 |
| CT bone strength test | 0556T, 0555T | 341 |
| 3D anatomy model for surgical planning | C8001 | 246 |
| ECG simulation to locate arrhythmia | 0897T | 101 |
| ECG AI | 0764T, 0765T | 60 |
| Liver ablation planning | 0944T | 47 |
| Lung disease subtype from CT | 0877T, 0878T | 20 |
| Heart chamber volumes from chest CT | G0183 | 1 |
Source: CMS preliminary outpatient rate-setting file (2025 claims)Cleared to Paid
More charts (2)
Add-on payments last only while an AI counts as new
Inpatient new-technology add-on payments for AI and software devices, maximum per case, $
Add-on payments last two to three years while a technology is new, then end. Several other AI devices were denied, withdrawn or found ineligible.
Show the numbers
| Technology | Maker | Per case | Latest decision |
|---|---|---|---|
| InVision Precision Cardiac Amyloid (InVision PCA) | InVision Medical Technology Corporation | $2,275.00 | approved (FY2027) |
| Ceribell Delirium Monitor System | Ceribell, Inc. | $2,171.00 | approved (FY2027) |
| Nelli Seizure Monitoring System | Neuro Event Labs, Inc. | $975.00 | approved (FY2027) |
| TriVerity Test | Inflammatix, Inc. | $243.75 | continued (FY2027) |
| Annalise Enterprise CTB Triage-OH (obstructive hydrocephalus) | Annalise-AI Pty Ltd | $241.39 | continued (FY2026-FY2027) |
| BriefCase-Triage: CARE Multi-Triage CT Body | Aidoc Medical Ltd. | $137.53 | approved (FY2027) |
| Bayesian Health Sepsis Flagging Device | Bayesian Health, Inc. | $61.84 | approved (FY2027) |
| Caption Guidance | Caption Health, Inc. | $1,868.10 | ended (FY2024) |
| ContaCT | Viz.ai, Inc. | $1,040.00 | ended (FY2023) |
| EchoGo Heart Failure 1.0 | Ultromics Limited | $1,023.75 | ended (FY2026) |
| Ceribell Status Epilepticus Monitor | Ceribell, Inc. | $913.90 | ended (FY2027) |
Source: Final inpatient payment rules, Federal Register, FY2021-FY2027Cleared to Paid
Only three AI products have a permanent code
AI product types with a permanent billing code: years from first FDA authorization (line start) to the permanent code (dot)
Only three AI product types have a permanent (Category I) code. The two cardiac ones spent years on temporary codes first.
Show the numbers
| Product type | First FDA | Temporary code | Permanent code |
|---|---|---|---|
| Autonomous diabetic eye screening | 2018 | none | 92229 (2021) |
| FFR-CT (coronary flow from CT) | 2014 | 2018 | 75580 (2024) |
| Coronary plaque from CT | 2019 | 2021 | 75577 (2026) |
Source: FDA; CMS final physician payment rulesCleared to Paid
More findings (3)
High confidence. A price is not use. Of 24 AI codes with hospital outpatient claims in Jan-Sep 2025, FFR-CT had 19,352 and the lung-nodule risk code 6,829; most others had a few hundred or fewer.
Why this confidence?
CMS preliminary rate-setting file (2025 claims).
High confidence. Only three AI product types have reached permanent codes: eye screening (2021), FFR-CT (2024) and coronary plaque (2026).
More, and why this confidence?
The two cardiac ones spent 5-6 years on temporary codes first.
CMS final rules.
High confidence. Triage tools' only separate payment is the time-limited inpatient add-on (NTAP): three have held one, at $137-$1,040 per case.
Why this confidence?
Final inpatient rules in the Federal Register.
Across countries
Medium confidence. Two questions predict per-case payment: does the AI produce a result of its own, and is it attached to a service that is already paid?
More, and why this confidence?
Per-case codes exist for a lung-nodule risk score (6,829 outpatient claims, Jan-Sep 2025), echo AI for heart failure (428) and ECG AI for low ejection fraction (60). AI that helps read an image or works inside a colonoscopy or bronchoscopy is funded from budgets.
Codes, prices and claims are CMS data; the rule is our reading of 36 cases.
Per-use payment mostly goes to results and treatment
How each type of AI is paid for, by country: 36 payment cases in America, China, Germany and Britain
Bold: the main route in that country. Own code: an insurer pays per use through the AI's own price item or code. Patients or others: per-use payment by patients or third parties. Budget: a hospital or government buys it, including central purchase.
Source: Cleared to Paid payment cases from pricing rules, payment data and filingsCleared to Paid
More charts (1)
Who should pay, in theory
Our analysis, not evidence: who captures the value of each type of AI, and so who would have to pay for the arrangement to last
| Type of AI | Lasting payer (our analysis) |
|---|---|
| Read-assist imaging | Buyer's budget. The provider's budget (per-case vendor pricing is fine); insurers pay per case only temporarily while evidence builds, or permanently once the AI gives a distinct, validated output (it then counts as a measurement or risk score); government for population screening |
| Measurement or risk score | Payer pays per use. The insurer, per use, tied to evidence |
| Autonomous screening | Government buys centrally. A long-horizon budget holder: government, single payer, Medicare, capitated plans, employers |
| Pathology AI | Buyer's budget. The lab or hospital budget (per-case vendor pricing is fine); drug makers for companion diagnostics; insurers only for a distinct, validated output |
| AI during a procedure | Buyer's budget. The facility's budget, because the procedure is already paid; an add-on payment only when the technology lets the procedure do something new; possibly quality-linked payment; no per-case fee for doing the same procedure better |
| Digital therapeutics and apps | Payer pays per use. The insurer, per use, with part of the price tied to outcomes |
| Surgical robotics | Inside a device price. The hospital's capital budget; a per-procedure price only where the robot does part of the work |
| Reconstruction and acquisition AI | Inside a device price. The device owner, through the maker |
| Documentation and decision support | Buyer's budget. The provider's own budget, above all providers that bear financial risk |
| Operations and claims | Government buys centrally. The payer itself, an insurer or government |
This is our reading of the cases, not a finding: it says where payment should settle if whoever gains is the one who pays.
Source: Cleared to Paid analysis of the payment cases (value capture, wrong-pocket problem, lab-test logic, payment-system design)Cleared to Paid
More findings (4)
Medium confidence. Hospital IT and workflow AI is a hospital cost, or a government or payer purchase, in all four countries; image-acquisition AI comes with the device.
Why this confidence?
Clear in every country we checked; Germany and the UK have few cases.
Medium confidence. Same technology, opposite payment: CT-FFR is Medicare's largest AI payment line but has no national price in China…
More, and why this confidence?
…while FFR computed from angiogram images has an insured ¥500-600 price in China and no separate US payment.
CMS data and Chinese provincial price lists; national coverage differs by province.
High confidence. Where per-case payment exists, prices differ about 11-fold: AI eye screening about $2.4 per test in China against about $27 per Medicare exam.
Why this confidence?
Company annual results and Medicare data.
Low confidence. In theory, per-case payment for read-assist AI lasts only while evidence builds: mammography CAD was paid per case, then folded into the exam codes.
More, and why this confidence?
Lasting payment sits with whoever captures the value.
Our analysis (theory overlay), not evidence.
Read the article: Pay for answers, not for help (coming soon)
China
High confidence. China's price system never pays read-assist AI on its own.
More, and why this confidence?
Radiology, pathology, clinical scales and digestive endoscopy charge AI at the main item's price, and local AI price items were ordered stopped in February 2026.
NHSA pricing guides and notices, read at source.
The US approved first in 11 of 11 like-for-like areas
First AI approval by frontier: US (line start) and China (dot), 11 like-for-like frontiers
China's lead is in scale and vendor count rather than timing.
Show the numbers
| Frontier | US first | US products | China first | China products |
|---|---|---|---|---|
| GI endoscopy AI | 2021 | 24 | 2022 | 6 |
| Fracture and musculoskeletal X-ray/CT AI | 2018 | 19 | 2020 | 10 |
| Ophthalmology screening (diabetic retinopathy) | 2018 | 7 | 2020 | 9 |
| Radiotherapy planning and contouring AI | 2017 | 64 | 2019 | 17 |
| ECG and cardiac signal AI | 2017 | 60 | 2020 | 11 |
| Lung nodule and chest CT detection | 2016 | 16 | 2020 | 15 |
| Stroke and neuro imaging AI | 2018 | 65 | 2022 | 20 |
| Coronary CTA analysis incl. CT-derived FFR | 2014 | 8 | 2020 | 11 |
| Ultrasound AI | 2016 | 92 | 2023 | 5 |
| Breast imaging AI | 1998 | 87 | 2019 | 4 |
| Digital pathology AI | 1995 | 8 | 2023 | 4 |
Source: FDA AI-enabled device list; NMPA Class III approvals (UDI database) and Beijing's official AI listCleared to Paid
More findings (3)
Medium confidence. Hospitals and governments pay instead: 73.9% of top-tier hospitals had imaging AI in 2022 but only 4.4% charged for it, and Guangdong gives one AI reading system free to 2,224 institutions.
Why this confidence?
A survey reported by the media, plus an official provincial notice.
High confidence. Big tech is paid as a builder of government platforms, not per use: Zhejiang awarded about ¥206M to Alipay (Hangzhou) for a provincial medical AI project…
More, and why this confidence?
…and the national insurer paid Alibaba Cloud ¥745K for fraud screening.
Government procurement award notices.
Medium confidence. China is not first on approvals (the US came first in 11 of 11 like-for-like frontiers).
More, and why this confidence?
Its lead is in scale and vendor count, such as 11 CT-FFR products from 11 firms.
Approval data are official; deployment figures are mostly company and media counts.
Value chain
Medium confidence. Four kinds of organisation build AI for care delivery, and in all four the hospital pays unless a billing code exists…
More, and why this confidence?
…then an insurer pays per use, whether or not the startup came out of research. Of the 34 companies with a separately reimbursed device, 15 are research spinouts and 10 independent startups. Drug makers, payers and governments, on the other side of healthcare, buy AI for their own goals.
Mapped from the health-system scan, the data-supply scan, payment cases and company filings; each model's cases are graded A or B. Company origins were checked at source for all 34 (8 unknown).
Drug makers and insurers make the largest disclosed AI and data payments
14 named examples we could source, not a market census: what each buyer pays for and how much, $ on a log scale. Colour: what is bought. Shape: the kind of figure
- Per-use tests and results
- Data and models
- Software and platforms
- A whole company
- Annual revenue
- One-off deal
- Contract ceiling
- Annual spend
- IQVIA technology and analytics $6.6B (annual revenue)
- Roche buying PathAI (up to) $1.1B (one-off deal)
- Tempus data and software $316.4M (annual revenue)
- Tempus–AstraZeneca–Pathos data and model fees $200.0M (one-off deal)
- Regeneron stake in Truveta's data network $119.5M (one-off deal)
- Tempus diagnostics $955.4M (annual revenue)
- iRhythm (AI-read heart monitoring) $747.1M (annual revenue)
- HeartFlow (FFR-CT) $176.0M (annual revenue)
- Medicare payments on all AI codes $10.3M (annual spend)
- VA ambient-AI contract $775.7M (contract ceiling)
- GE HealthCare bought Intelerad $2.3B (one-off deal)
- Tempus bought Paige $81.0M (one-off deal)
- Health systems' AI software spend $1.0B (annual spend)
- RadNet digital health (imaging AI) $92.7M (annual revenue)
Low confidence. Named examples, mostly from public companies; private deals and spend are missing. Revenue and spend recur every year; deals are one-off and a contract ceiling is a maximum over five years, so compare like with like. Hospitals' income from selling data is not disclosed.
Show the numbers
| Buyer | What | Item | $M | Kind of figure | Year | Grade |
|---|---|---|---|---|---|---|
| Drug makers | Data and models | IQVIA technology and analytics | 6,626.0 | annual revenue | 2025 | A |
| Drug makers | Data and models | Tempus data and software | 316.4 | annual revenue | 2025 | A |
| Drug makers | Data and models | Tempus–AstraZeneca–Pathos data and model fees | 200.0 | one-off deal | – | B |
| Drug makers | A whole company | Roche buying PathAI (up to) | 1,050.0 | one-off deal | 2026 | B |
| Drug makers | Data and models | Regeneron stake in Truveta's data network | 119.5 | one-off deal | 2025 | B |
| Insurers and Medicare | Per-use tests and results | Tempus diagnostics | 955.4 | annual revenue | 2025 | A |
| Insurers and Medicare | Per-use tests and results | iRhythm (AI-read heart monitoring) | 747.1 | annual revenue | 2025 | A |
| Insurers and Medicare | Per-use tests and results | HeartFlow (FFR-CT) | 176.0 | annual revenue | 2025 | A |
| Insurers and Medicare | Per-use tests and results | Medicare payments on all AI codes | 10.3 | annual spend | 2024 | A |
| Governments | Software and platforms | VA ambient-AI contract | 775.7 | contract ceiling | 2026 | B |
| Platform owners and acquirers | A whole company | GE HealthCare bought Intelerad | 2,300.0 | one-off deal | 2026 | A |
| Platform owners and acquirers | A whole company | Tempus bought Paige | 81.0 | one-off deal | 2025 | A |
| Hospitals | Software and platforms | Health systems' AI software spend | 1,000.0 | annual spend | 2025 | A |
| Hospitals | Software and platforms | RadNet digital health (imaging AI) | 92.7 | annual revenue | 2025 | A |
Source: Company filings and results; CMS; Menlo Ventures 2025; GE HealthCare; trade press for Roche–PathAI, Regeneron–Truveta, the VA contract and Tempus–AstraZeneca–Pathos (grade B)Cleared to Paid
The largest known sums sit with data companies, not data sources
Known annual revenue or spend of named firms by position in the chain, 2025, $
Low confidence. Data companies: IQVIA and Tempus data only. Paid per use: Tempus diagnostics, iRhythm and HeartFlow. Software for providers: health systems' total AI spend. Named public firms only, so each bar is a floor.
Source: IQVIA, Tempus, iRhythm and HeartFlow 2025 results and filings; Menlo Ventures 2025Cleared to Paid
In all four care models, the hospital pays unless a code exists
Six business models, four in care delivery and two on the other side of healthcare: who supplies the data, builds the AI, files with FDA and pays. Open a row for its evidence and our read
| Model | Used for | Data from | Builds | Files with FDA | Pays |
|---|---|---|---|---|---|
| Care delivery | |||||
| Hospital builds for itself (no company) | Clinical alerts · Documentation and admin | Hospital (its own records) | Hospital (own team) | nobody (run as decision support) | Hospital (budget, grants, donors) |
Evidence (grade A) and our read
Our read: Cheap to start and close to the problem, but nobody outside pays. It lasts where the hospital itself gains, as with shorter stays, and rarely leaves the building. | |||||
| AI startup (research spinout or independent) | Diagnosis and treatment · Clinical alerts | Parent (archive, for equity) or Broker (bought, revenue share) | Startup | Startup | Hospital or insurer (budget; insurers per use where a code exists) |
Evidence (grade A/B) and our read
Our read: Two ways to start, one way to get paid. A spinout starts with its parent's archive and evidence and pays the parent in equity; an independent startup buys its data. Either reaches insurers only with a result of its own and a code. | |||||
| Device maker bundles AI into hardware | Diagnosis and treatment | Device maker (installed base and clinical partners) | Device maker (own or acquired) | Device maker | Hospital (capital budget, often a paid option) |
Evidence (grade A) and our read
Our read: Owns the machine where images are made, so its AI ships with the hardware. Hospitals pay in the capital budget, often as a paid option, and scanner AI rarely gets a code of its own. | |||||
| EHR and imaging-IT platform bundles AI into software | Clinical alerts · Documentation and admin · Diagnosis and treatment | Platform (records of customer hospitals that opt in, de-identified) | Platform (own models, plus partners' AI) | nobody (mostly not filed: decision support and admin tools) | Hospital (software subscription) |
Evidence (grade A/B) and our read
Our read: Owns the screen where clinicians work, so it decides which AI they see. Its models arrive inside a subscription the hospital already pays, which standalone vendors have to beat. | |||||
| Other side of healthcare | |||||
| Life sciences: drug makers pay for data and AI | Drug research | Data and diagnostics firms (license hospital and network data) | Data and diagnostics firms (build models and tests) | Data and diagnostics firms (companion diagnostics when needed) | Pharma (licence and model fees) |
Evidence (grade A/B) and our read
Our read: The biggest cash payer for clinical data and models, but for drug development, not bedside care. Its money reaches hospitals mainly through data companies. | |||||
| Payers and governments buy for themselves | Documentation and admin · Diagnosis and treatment | Payer or government (claims and public patient records) | Vendors | Vendors (file only for clinical tools) | Payer or government (a share of savings, or central purchase) |
Evidence (grade A/B) and our read
Our read: Pays only where it keeps the saving itself, as in claims review or screening at scale. Central purchase skips hospital budgets but leaves vendors with one big customer. | |||||
"Our read" is our analysis, not evidence; the bullets are the evidence.
Source: Cleared to Paid scans of health systems, data supply, payment cases and leaders (2026-10); grades A primary or filing, B company or trade pressCleared to Paid
More charts (3)
Drug makers and insurers pay the biggest sums
Revenue or payments around medical AI, 2025 (Medicare: 2024), $
Bars share one scale. IQVIA's figure is its technology and analytics business; RadNet's is its digital-health arm.
Show the numbers
| Revenue line | $M | Who pays |
|---|---|---|
| IQVIA technology and analytics | 6,626.0 | Drug makers (data and software) |
| Tempus diagnostics | 955.4 | Insurers (tests) |
| iRhythm (AI-read heart monitoring) | 747.1 | Insurers (tests) |
| Tempus data and software | 316.4 | Drug makers (data and software) |
| HeartFlow (FFR-CT) | 176.0 | Insurers (per-use code) |
| RadNet digital health (imaging AI) | 92.7 | Hospitals, imaging centres, patients |
| Medicare, all AI codes (2024) | 10.3 | Medicare (per-use codes) |
Source: IQVIA, Tempus, iRhythm, HeartFlow and RadNet 2025 results and filings; CMS physician dataCleared to Paid
Data is paid for in the middle, not at the source
Follow the data: where money changes hands. Each step hands data on; the receiver pays as shown
- Paid in cash
Hospital archivedata →Broker or spinout
Receiver pays: in kind, in equity or with a revenue share
- Gradient Health: 'typical' 20% revenue share
- MSK → Paige: slide archive for equity
- Of 24 hospital-sourced arrangements we found, 4 paid cash and 10 did not disclose terms
Broker or spinoutdata →AI developer
Receiver pays: a licence fee
- OneMedNet → OncoSoft: purchased CT data (FDA K242994)
- Segmed → NeuroCare.AI: FDA test set (K252954)
- Brokers mostly supply FDA test sets (8 of 11 mentions)
Data companydata →Drug maker
Receiver pays: cash, the largest sums
- IQVIA technology and analytics: $6.6B revenue (2025)
- Tempus–AstraZeneca–Pathos: $200M in data and model fees
Named cases show where money changes hands; the counts come from 73 arrangements we could source, not a census.
Source: Cleared to Paid data-supply scan (2026-10): company filings, FDA summaries and broker pages; grades A primary or filing, B company or trade pressCleared to Paid
Hospitals supply most traceable data but are seldom paid cash
Arrangements that supply clinical data for AI, by supplier and by how the supplier is paid: 73 examples we could source across specialties
Examples, not a census: bars show patterns, not market shares. FDA summaries rarely say where training data came from: 79 of 1,614 name a source (13 a public dataset, 19 a hospital or research site, 64 a country).
Show the numbers
| Supplier | Paid in cash | Equity | Free, grant or in kind | Not disclosed | Total |
|---|---|---|---|---|---|
| Hospitals and health systems | 4 | 3 | 7 | 10 | 24 |
| Public datasets and challenges | 0 | 0 | 12 | 0 | 12 |
| Brokers and data exchanges | 11 | 0 | 0 | 0 | 11 |
| Health-system data networks | 1 | 2 | 2 | 1 | 6 |
| Device makers' own customers | 0 | 0 | 5 | 0 | 5 |
| Clinical trials and studies | 0 | 0 | 1 | 3 | 4 |
| Real-world data companies | 3 | 0 | 0 | 0 | 3 |
| Government data programmes | 1 | 0 | 1 | 0 | 2 |
| Other | 1 | 1 | 2 | 2 | 6 |
Source: Cleared to Paid data-supply scan (2026-10): company pages, filings, FDA summaries and papers; grades A to CCleared to Paid
More findings (6)
Medium confidence. Hospitals supply most of the training data we could trace but are seldom paid cash for it: of 24 hospital-sourced arrangements, 4 were paid in cash, 3 in equity…
More, and why this confidence?
…7 free or in kind and 10 undisclosed. Brokers, data exchanges and data companies were paid in cash in all 14 we found.
73 sourced arrangements across specialties, graded A to C; examples, not a census.
Medium confidence. Per-use payment is what turns clinical AI into revenue: HeartFlow made $176M in 2025, 98% from a per-use coded test…
More, and why this confidence?
…while products deployed in thousands of hospitals without a code earn little (RadNet's digital health arm about $93M).
Company filings; deployment counts are company-stated and not comparable.
Medium confidence. Pharma pays the biggest proven sums in the chain, for data and tests: IQVIA's technology and analytics arm earned $6.6B in 2025 and Tempus $316M from data.
Why this confidence?
Company filings; private vendors' revenue is unknown.
Medium confidence. Owning the point of read or the buyer brings power: GE paid $2.3B for the PACS vendor Intelerad, Radiology Partners builds its own AI…
More, and why this confidence?
…and General Catalyst now owns a health system (Summa Health) that buys from its own portfolio.
GE's release checked; the rest from company sources.
Medium confidence. Health systems are a new force as builders and data owners, not as sellers.
More, and why this confidence?
Only 2 of 1,614 FDA AI devices were filed by a health system, about 11 trace to the 21 large US systems we scanned, 35 of their 64 AI initiatives stay internal, and none discloses AI revenue.
Device counts are exact; initiatives come from a 21-system scan, 40 of 64 rows from primary sources.
Medium confidence. New payment routes are central or tied to outcomes, not new per-use AI codes: Medicare's ACCESS model (outcome-aligned payments since 2026-07-05), WISeR (vendors paid a share of avoided spending)…
More, and why this confidence?
…the VA's ambient-AI contract ($775.72M ceiling) and central purchase by Chinese provinces.
ACCESS checked at CMS; the VA ceiling is from a secondary source.
Read the article: Follow the money, not the model (coming soon)
Evidence and error rates
Medium confidence. The classifier is reliable for company type (94%), radiology function (98%) and model type (96%), and for whether training data is mentioned (92%); weaker for clinical area (85%), modality (83%) and the training-data source (79%).
More, and why this confidence?
Deployment (50%) is used nowhere.
Measured on 100 random devices against reviewed labels: the owner's review where the evidence was unsure, otherwise answers checked against the full FDA summary with a quote; an adversarial re-check found no wrong answers in 42. 95% intervals are about ±7 points.
Labels match reviewers 79–98% on the fields we use
Classifier labels that match reviewed labels on 100 random devices, with 95% intervals
Deployment (grey) is used in no finding. Where a fix followed the check, the rate shown is after it; the notes give the rate before.
Show the numbers
| Label | Match | 95% interval | Note |
|---|---|---|---|
| Company type | 94% | 88–97% | Measured before the round-2 fix that this check triggered (99% after) |
| Radiology function | 98% | 93–99% | |
| Clinical area | 85% | 77–91% | |
| Modality | 83% | 74–89% | Refined 2026-10-12 from full summary text (75% before) |
| Model type, 4 values | 96% | 90–98% | 97% before the 2026-10-13 fix |
| Names a model (yes/no) | 97% | 92–99% | |
| Training data, 4 values | 79% | 70–86% | Fixed 2026-10-13 for line breaks and wording (77% before) |
| Mentions training data (yes/no) | 92% | 85–96% | Fixed 2026-10-13 (88% before) |
| Deployment | 50% | 40–60% | Not used in any finding; OS-only summaries count as on-premise (ruling 1) |
Source: Cleared to Paid label check: blind labels, an online check of every open label, an adversarial re-check and the author's reviewCleared to Paid
More findings (3)
Medium confidence. FDA summaries disclose more since 2022: naming a model type rose from 35% to 51%, mentioning training data from 17% to 49%…
More, and why this confidence?
…and naming where training data came from (a dataset, institution or country) from 0.2% to 6.9%.
Keyword rules, fixed on 2026-10-13; they match the reviewed labels on 92% of devices for whether training data is mentioned. The source share was checked sentence by sentence; the keyword tags had overstated it (17%).
High confidence. Who-pays labels: the type of payer was right for 28 of 28 devices, the exact mechanism for 17 (61%, 95% CI 42-76%).
Why this confidence?
Case review at source.
High confidence. A widely repeated claim that 12 AI services entered China's national insurance list in April 2026 is false; it traces to a vendor sales-staff post that an AI assistant then 'confirmed'.
Why this confidence?
Traced to the original post and NHSA's catalogue plan.
How every number was built, and its limits: the method.