State of AI for Healthcare
Where margin actually accrues in clinical AI once you follow the money path rather than the technology — US-first, August 2026
Own only the businesses whose payer is pharma's commercial budget or the provider's own collected revenue; the technology is real, the money path is not, and the entities that solved it are private — so deploy small, name the trigger, and hold the cash.
Healthcare is the largest AI application market with a structurally broken money path: the layer that generates demand does not pay, the layer that buys clinical AI runs a 2.5% median operating margin (Kaufman Hall, April 2026 CYTD, 1,300+ hospitals), and the layer that theoretically pays for validated AI has now been measured and is a moat around a puddle — CMS designates 36 HCPCS codes as Software as a Medical Service and moves the aggregate rate across the 30 already-rated O1 codes by +0.53%, while the three flagship AI services it names generated 81 Medicare OPPS claims nationally in CY2024 across their standalone and add-on codes (0625T 22, 0698T 55, 0724T 4), which at the $950.50 APC 1511 rate is $76,990.50 against the $110.9B of total CY2027 OPPS payments CMS projects — one dollar in every 1.44 million. Regulatory assets are barriers to entry, not margin pools: 1,524 FDA authorisations stand against three Category I CPT codes, a 508:1 ratio, and GE HealthCare holds the largest AI-clearance stock in the world at 130 authorisations on a 13.9% Advanced Imaging Solutions segment EBIT while Doximity, holding zero clearances and no CPT code, earns an 89.1% GAAP gross margin and $317.5M of free cash flow. Eight of seventeen registered shift-points survived verification and only three have a clean, liquid, US-listed instrument, so the book is deliberately small — Doximity 7% as the control case, Waystar 4% as the collected-revenue expression, HeartFlow 2% short into the CY2027 OPPS final rule — with 89% in cash held against a named trigger, the 2026-27 digital-health IPO slate. Six private marks (~$33.05B) roughly equal the entire 17-name US-listed pure-play cohort (~$33.90B), and the one US-listed ETF marketed as healthcare AI holds nine therapeutics or life-science-tools names in its top ten with zero clinical-workflow, EHR, RCM or ambient exposure, so the index route does not exist either. Not investment advice.
Margin sits where the buyer is pharma's commercial budget or the provider's own collected revenue — never where the buyer is the provider's IT opex line and never where the payer is a CMS fee schedule. Doximity, holding zero FDA clearances and no CPT code, earned an 89.1% GAAP gross margin, $196.1M of net income and $317.5M of free cash flow on $644.9M of FY2026 revenue. GE HealthCare, holding the single largest FDA AI-clearance stock in the world at 130 authorisations — 107 filed under a GE name plus 23 filed by companies it has since bought — earned a 13.9% Advanced Imaging Solutions segment EBIT on $3,771M of segment revenue in Q2 2026, and paid $2,293M for a workflow platform holding zero AI authorisations against the roughly $435M of consideration disclosed in its current 10-K for the 2024-2026 acquisitions that carried 15 of those 23 authorisations, with the other ten unpriced. Those are two different margin lines — a gross margin against a segment EBIT — and are not directly comparable; what is comparable is the direction of the two businesses. Clearance stock is a barrier to entry at a $26,067 filing fee, not a margin pool. What sets price is distribution, and distribution belongs to Epic — 43.7% of US acute-care hospitals in 2025, a fifth consecutive year of gains — which collects its rent not through a toll (Vendor Services runs roughly $1,700-1,900 per vendor per year) but by substituting its own AI into a contract it already holds.
Full-stack value chain
Each layer scored on whether anyone actually pays for the AI it contains — a code, a contract, or a claim line — rather than on how good the technology is. August 2026
The stack has no chokepoint; it has a broken money path. Demand sits with patients, who do not pay. Purchasing sits with providers, running a 2.5% median operating margin. The payer of last resort has now been measured and is empty — 81 Medicare OPPS claims in CY2024 across the three flagship AI services CMS names, $76,990.50 against $110.9B of projected CY2027 OPPS payments. So margin does not follow clinical capability or clearance stock; it follows payer identity. The two layers that earn it — the clinical system of record, and clinician attention and knowledge distribution — are paid by someone other than the provider's IT budget: distribution rent inside a contract already held, and pharma's commercial budget. Every durable business in the evidence base solved the same problem the same way, by finding a fourth payer.
Upgrade to Pro to see the full value chain.
Opportunity board
Ranked by whether a listed instrument actually expresses the thesis — six of the ten are avoids, which is itself the finding.
Durable compounder
1Undervalued / high-potential
3Short / avoid
6Upgrade to Pro to see the full opportunity board.
Full board analytics
The prediction matrix, scenario bands, regime timeline, book construction, premise tests and the shift-point register — the deep analytical layer behind this board.
- Prediction matrix & scenario bands
- Book construction & sizing
- Premise tests & falsifiers
- Shift-point register