Bunkerhill Health

The deployment layer is a feature, not a company. Epic absorbs the governance and orchestration surface for its 56.9% of US beds, Aidoc's CARE plus aiOS takes the imaging-AI substrate above Bunkerhill's weight class on ten times the capital, and the G0680 code Bunkerhill created becomes the standard every cleared CAC vendor bills — HeartLung.AI, Coreline Soft and Riverain are already there. CMS does to G0680 what it did to the first AI NTAP: transitional payment, then retirement (REG-NTAP-002 — $1,040 max per case in FY2021, one extension, ZERO appearances in FY2023). A 2.5%-margin buyer declines to expand the IT line, agent counts plateau at pilot depth outside the flagship accounts, and the randomised evidence proves narrower than hoped once it leaves a single Stanford cohort and a primary-prevention population. The 20x decays hard as the base grows. With $55M cumulative against competitors at ten to twenty times that, the outcome is an acqui-hire or a modest trade sale — and because no post-money was ever disclosed, there is not even a down-round to observe, only a private mark that never becomes a print.

A real, well-executed company in a category whose ceiling is set by someone else. Growth continues off a small base — the 20x directionally true, the absolute number modest — and the customer list deepens without the step-change a platform thesis requires. G0680 works as a permit rather than pricing power: volume grows, competitors bill the same code, and the New Technology APC assignment holds without becoming material to anyone's P&L. Integration stays semi-bespoke, so gross margin lands between software and services and the sales cycle stays measured in quarters. Epic ships more native agents and takes the commodity use cases; Aidoc's foundation model takes the broad imaging coverage; Bunkerhill retains the specialised, clearance-backed, evidence-backed ones. It raises again privately at a mark that is negotiated rather than discovered, and the value question stays unresolved because nothing forces a price.

Deployment turns out to be the real bottleneck and the substrate is the thing worth owning. Agent counts across the other fourteen systems converge on the UTMB pattern, so revenue compounds inside accounts where integration cost is already sunk, and gross margin climbs as integration templatises. The FDA portfolio widens across opportunistic findings, each clearance becoming a new agent for an installed base that already bought the substrate, and the NOTIFY-PICTURE result pulls guideline and payer attention toward AI-detected incidental CAC. G0680 volume ramps through the lung-cancer-screening LDCT population, giving every health-system CFO a billable line to point at. Optum converts into a payer-side channel — the durable-payer mechanism the board's thesis says actually funds software margin. In this branch the shared code stops mattering because the moat was never the code; it was the twenty agents in production that a competitor would have to displace one at a time.

• Aimed at the correct bottleneck. The board's own measurement is that clinical AI fails on the money and deployment paths, not accuracy — the three flagship AI cardiac services generated 81 CY2024 Medicare OPPS claims worth $76,990.50 (canon REG-OPPS-010). Carebricks targets that gap, and 15 systems with 20+ agents at the deepest account is the evidence that counts
• Randomised outcome evidence, rare in this sector — NOTIFY-PICTURE found statin prescriptions of 44 (44%) in the notification arm vs 10 (9.8%) in usual care at six months, p<0.001 (Circulation; NCT05588895)
• Regulatory substance most agent-platform startups lack — its own FDA-cleared algorithms, with first-in-class positions in contrast-enhanced non-gated chest CT calcium quantification and mitral annular calcification. These are 510(k) CLEARANCES establishing substantial equivalence, not PMA approvals and not De Novo grants
• It authored its own payment identity — CMS established HCPCS G0680 effective 1 April 2026 (verified at CMS primary). Very few clinical-AI companies of any size have driven a national billing code to establishment
• Land-and-expand works. Once a system runs a fleet of agents on one governance substrate, the substrate is an operating dependency rather than a subscription — the one switching cost in clinical AI that is not a shared code someone else can bill

• Nothing about the revenue line is knowable — no dollar figure at any vintage, no ARR, no pricing, no gross margin, no NRR — only a '20x' ratio with no base. A 20x on $250k and a 20x on $5M are different companies
• The code it won is non-excludable by construction. G0680 describes the SERVICE: HeartLung.AI publicly welcomed the same coverage; Coreline Soft and Riverain Tech market ROI material against it. Bunkerhill bore the cost, the category collects the benefit
• APC 1492 is a New Technology APC — the discretionary bucket, one band above APC 1491's $0–$10 Level 1A floor. A low-band transitional assignment, and CMS has pre-committed to routing new software-as-a-medical-service analyses into these APCs (canon REG-CMS-010)
• The randomised evidence is thinner than the headline — SINGLE site (Stanford), open-label, 202 randomised from 6,188 screened, a prescribing process outcome rather than a cardiovascular one, and no effect in patients who already had ASCVD (29% vs 31%). Both founders are listed as Bunkerhill employees in the disclosures and a senior co-author holds equity
• Its buyer is the least able buyer in enterprise software — median hospital calendar-YTD operating margin 2.5% incl. corporate allocations in April 2026 across >1,300 hospitals, 26–99-bed cohort at −11.5% (canon CM-KH-01 / CM-KH-05)

• Agent-count expansion inside the existing 15 systems is the highest-return growth available: integration cost sunk, substrate live, incremental agent close to pure margin. Twenty-plus at UTMB is the template for the other fourteen
• G0680 is a wedge into every system running high-volume chest CT — most obviously lung-cancer-screening LDCT, where one acquisition now supports both the lung read and a billable cardiovascular assessment. Category creation with no new scan, order or visit
• The opportunistic-findings franchise generalises — bone density on abdominal CT, mitral annular calcification, any other actionable finding recoverable from imaging a system already owns. Each clearance is a new agent for an installed base that already bought the substrate
• Optum Ventures opens a channel into the largest US payer-provider organisation. Payer-side and value-based-care demand for opportunistic risk stratification is a different budget from a hospital IT line — potentially the correct-payer mechanism the board calls the only durable one
• Consolidation is plausible and Bunkerhill is a coherent target — a governance layer with real FDA clearances, randomised evidence and an owned CMS code is what an imaging incumbent, an EHR vendor, a payer or a diversified health-IT platform would rather buy than build

• Epic. 43.7% of US acute-care hospitals in 2025, 56.9% of beds, AI Charting GA from 4 Feb 2026, Penny at 200+ organisations, Art at 16M uses a month (board canon). Epic is not a tollbooth taking rent from AI vendors — it is a distribution owner that has started selling the AI itself, at exactly the integration point Bunkerhill must pass through
• Aidoc runs the identical strategy with an order of magnitude more resources: 34 FDA authorisations on canon's basis, its CARE foundation model cleared Jan 2026 as what it calls healthcare's first comprehensive foundation-model triage solution (11 new plus 3 existing indications in one workflow), delivered through the aiOS operating system, $150M raised Apr 2026, 150+ health systems
• The Viz.ai precedent for payment-pathway decay, verified at primary in canon: ContaCT was the first AI software granted a Medicare NTAP, at $1,040 max per case in the FY2021 IPPS final rule, extended once for FY2022, and appearing ZERO times in the FY2023 final rule (REG-NTAP-001/002)
• The integration layer may not templatise. If each system's Epic build, PACS estate and governance committee makes deployment bespoke, gross margin stays services-shaped, the sales cycle stays long, and 'platform' is a story the unit economics do not support. No disclosed metric distinguishes the two futures
• Agent-governance regulation is unsettled. FDA's posture on adaptive, LLM-mediated clinical agents — as opposed to the locked 510(k)-cleared algorithms Bunkerhill holds — is still forming, and a platform selling 'deploy many agents fast' is exposed to any tightening that makes each agent a separately regulated device